Read time: 5 minutes | Unlock my Claude Skills Bundle

Hey friend, Alex here.

Last week I wrote a one-paragraph brief for a 60-second video. It asked for an enterprise AI story at full speed, real 2026 adoption stats on screen, a slogan on the end card, and a soundtrack, with every frame written in code and rendered by Claude. No editor, no stock footage, no timeline.

Before I worked in marketing, I edited film. A minute of finished video holds hundreds of small decisions about timing, type, and sound, and the AI video tools I’ve tried make those decisions for you. You get a clip back, the words on screen come out misspelled, and changing one number means generating the whole thing again.

Today:

• Opus 5.5’s first week, including the one-sentence prompt behind the most-shared motion design video of launch week

• The fill-in-the-brackets prompt that has Claude research, build, render, and fact-check a video for your brand, plus the video it made from my brief (copy-paste ready)

• 4links worth keeping, including a motion design dictionary that gives you the exact words for the motion you want

Our recommended partner:

Your AI agent is useless if it can’t talk to anything.

APIdays covers the APIs, protocols, platforms, and infrastructure making the agentic web actually work.
See what’s powering AI →

📰 WHAT HAPPENED LAST WEEK

Anthropic released Claude Opus 5.5 on Tuesday, 22 September. The company says it performs at the level of Claude Fable 5.1, its Mythos-class model from earlier this month, on most work, costs 40% less to run than Opus 5 on typical workloads, and writes its output more than 30% faster. Anthropic also says Opus 5.5 follows the writing rules you give it and puts the important information first, and on launch day it raised the five-hour usage limits on the Pro, Max, and Team plans.

Most of the posts I saw that week showed something Opus 5.5 had built rather than something it had said. Ryan Sael asked it to explain camera focus and got back an interactive lens lab after 1 hour 26 minutes and $25.66 in API costs, and the site is live for anyone to try.

The one I kept rewatching came from Stephan Livera. He set Opus 5.5 to its highest effort level and asked, in one sentence, for a 15-second motion graphics showreel proving it’s a great motion designer, the kind you’d attach to a résumé. Within days, other people were reusing his prompt for their own versions.

The most useful criticism came from people who work in motion design. Alex Barashkov, a founder who tried Opus 5.5 for motion work, wrote that it's good for overnight experiments but still makes design mistakes that anyone with some taste would catch, and that you still have to handle sound effects yourself. That's one person's experience, and my run went differently on sound, because Claude wrote the whole soundtrack in code. It did tell me it couldn't listen to what it made.

⚡ THE SUPERPOWER: Programmatic Video

AI video models generate the picture directly. You describe a scene and a scene comes back, but you can’t change one stat, fix one misspelled word, or move one cut half a second earlier without generating the video again and hoping the rest stays the same. In a brand video, the stats and the slogan are the parts that have to be exactly right, so that’s the part you can least afford to leave to chance.

Programmatic video means the video exists as code before it exists as pixels. Every word on screen, every cut, and every sound is a line you can read, and a revision means changing that line and rendering again. When the video is code, you can check the stats on screen the same way you check a document, before anyone sees them.

When I ran it, Claude dropped three stats it couldn't verify before I ever saw the list, then caught a line on screen that I had never approved.

⚙️ THE WORKFLOW

A one-sentence prompt works for a showreel because nothing in a showreel has to be exactly right. A brand video has stats and a slogan that do. Anthropic’s own advice for Opus 5.5 is to hand over a whole task and define what “done” means, so I rebuilt my paragraph around that. The brackets are the only part you change. The rest tells Claude how to build the video, when to stop and ask you, and what it has to check before it calls the job finished.

Run it in Claude Code with Opus 5.5 selected, so it can install whatever video and audio tools it needs. I used the browser version at claude.ai/code, which works inside a GitHub repository, and I set its network access to Full so it could read the brand’s website and the stat sources. Paste this and fill in the brackets.

Make a [60]-second video, written as code and rendered to an MP4. Use whatever tools you need and install anything required.

The brief
- Brand: [company name], [one line on what it does]
- Brand assets: [your website to pull the logo and colors from, or "use the files in this folder"]
- Story: [the situation the viewer is in], building to [the moment they need you]
- Twist: [what makes the brand different from everyone else in the story]
- Style: [e.g. clean enterprise look, fast cuts, loud and detailed visuals]
- End card: [slogan] and [link]
- Audio: [e.g. a driving electronic track that builds with the cuts, with an impact sound on each stat]

Stats
Research current statistics on [topic and year]. Only use numbers from a named source you can link. Before you build anything, list every stat you plan to put on screen with its source, date, and link.

Scene list
Write a scene list with timings. Show me the stats list and the scene list together, then stop and wait for my approval.

How to build it
- Draw every frame in code as a function of time, so rendering twice gives the same video.
- Write or generate the audio yourself, or use only tracks licensed for commercial use, and tell me which you did.
- Render at 1920x1080, 30 frames per second.

Done means
- An MP4 of [60] seconds, plus or minus one second, with audio.
- You extracted one still frame from each scene and checked every word and number on it against the approved stats list.
- You tell me which tools you used, how long the render took, and anything you couldn't do.

Then do four things.

1. Make the storyline specific. “Teams drowning in AI announcements, then one partner who ships working systems” gives Claude something to cut to. “AI is everywhere” gives it nothing.

2. Open two of the stat links before you approve. A wrong number in a published video costs more than a slow video, which is why the prompt makes Claude stop and wait at that point.

3. Ask for edits by timestamp. “At 0:14, hold the stat for another half second” works better than “make it punchier,” because Claude can find the exact line of code that controls 0:14.

4. Keep the project folder. Your second video starts from the first one’s code, so the fonts, colors, and timing carry over.

I ran it on the brief I opened with. Limestone Digital is one of my clients, so this was a real job with a real slogan and a real link.

Opus 5.5 built the whole thing in Python. It drew every frame with Cairo, wrote the soundtrack from scratch with numpy and scipy at 120 beats per minute, and stitched everything into exactly 60 seconds at 1080p and 30 frames per second with ffmpeg. The logo, colors, and typefaces came straight from Limestone’s website.

Seven stats made it on screen, from Stanford’s AI Index 2026, Deloitte, and PwC, and it dropped three candidates before showing me the list, including Gartner forecasts it couldn’t verify. Its own frame check caught a ticker line, “NEW BENCHMARK SATURATED,” that wasn’t on my approved list, and swapped it out.

The final render took about 80 seconds, after 45 minutes of research and building. That’s one brief and one run, on medium effort, which is the default.

Fair warning. The frame check compares words and numbers against your approved list, so it can’t see layout. In the opening seconds, a purple bar ran straight through the letter A, and I only caught it by watching. One message fixed it, so watch the whole video yourself before you publish.

💬 PROMPT OF THE DAY

Before I send this, find the one sentence where a skeptical reader is most likely to stop trusting me, and tell me why.

Why it works: asking for one sentence forces the model to rank the weak points instead of listing everything it could improve.

Best on: any model.

📚 USEFUL RESOURCES

• 🔭 Ryan Sael’s lens lab (an interactive site Opus 5.5 built) → turn the focus ring and watch the glass elements move, a good look at what one prompt produced in launch week.

• 📖 School of Motion’s Motion Design Dictionary (140+ terms, plain definitions) → learn words like ease, tween, and logo resolve, so the Style line in your brief tells Claude exactly what motion you want.

• 📈 Easings.net (easing curves you can watch move) → pick a curve by name, like easeOutExpo, and write it into your brief instead of “make it snappier”

• 🧭 Anthropic’s Opus 5.5 playbook (how to hand over a whole task) → where the “define done” part of today’s prompt comes from.

What should I have Claude render next, a product demo for a real landing page or an explainer built from one of my old X threads? Hit reply, one word is enough.

Know someone whose team keeps putting off making a video? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. Today’s video prompt is one workflow. My Claude Skills Bundle turns Claude into 20+ specialists for marketing and business. Unlock it in one click: https://linktr.ee/alex_prompter

AI Superpowers, part of the Alex Prompter network. One new AI superpower every Tuesday, under 5 minutes.

Read time: 5 minutes | Unlock my Claude Skills Bundle

Hey friend, Alex here.

For seven days my feed has been about one model, Jev, which came out of stealth last Monday, and one number, 445x, the claimed cost advantage over frontier models. By Wednesday I wanted to tell 200K people to go try it. I didn’t, and this issue is about why.

Last week I wrote that nobody in the “AI will kill us all” story is paid to de-escalate the fearmongering. This week the same problem runs in the opposite direction. Nobody in the Jev story is paid to tone down the hype. Every recommendation you read about it came from someone who gains if you believe it, and none of them disclosed what they gain. That includes me, and I’ll tell you what I gain before the end.

Today:

• The Jev week in order, and the one question I asked every voice in it before deciding where I land

• The Incentive X-Ray, a prompt that shows you who gains if you believe a recommendation and how much to discount it (copy-paste ready)

• 3 links worth keeping, including TypeSafe’s own launch post, which turned out to be more careful than the coverage of it

Chatbots were the easy part. Now AI has to work in the real world.

Grounded Intelligence covers the edge systems, distributed infrastructure, and technology behind intelligent systems that actually operate outside the chat window.

Go beyond the chatbot →

📰 WHAT HAPPENED LAST WEEK

Monday, 15 September. Diogo Almeida, whose work at OpenAI became the research behind ChatGPT, launched TypeSafe AI’s first model after two years in stealth, with a $40 million seed round led by DCVC. Jev doesn’t write text. You hand it a situation and a question, and it returns a decision with a probability attached, in 70 to 500 milliseconds. The launch post claimed 20 to 200x faster and 40 to 400x cheaper than frontier models. 39 million views and counting.

Tuesday to Thursday. Vercel wired Jev into its AI Gateway, one of its engineers posted that it beat their current safety classifier on an internal test, and on Thursday Vercel announced the fastest adoption of any model in Gateway history. TechCrunch ran the headline that developers are thrilled.

Friday to this morning. Vercel made Jev free on its Gateway until 25 September. TypeSafe dropped the waitlist on Saturday and gave every new account $5 in credit. On Sunday the API had an outage, and by Monday TypeSafe had paused new signups. One developer, Harrison Kinsley, posted the only comparison with a number against Jev. In his game-playing test, another model, GLM 5.3 Flash, beat Jev 82% of the time.

I asked one question of every voice in that timeline before deciding what I think, and it’s the same question the prompt below asks. What does this person gain if I believe them?

TypeSafe gains customers, and a fresh funding round needs adoption before competitors release similar models. Vercel gains traffic through its Gateway, and it measured its record-adoption chart during the week it was giving Jev away for free. The engineer behind the benchmark TechCrunch relayed works at Vercel. The trade sites gain a story with a ChatGPT-inventor angle, and most of them ran the company’s numbers without the company’s caveats. The skeptics, as far as I can find, gain nothing, and Kinsley’s result sits at 42,000 views next to Almeida’s 39 million.

Now the surprising part. The most careful document of the week came from TypeSafe itself. TypeSafe’s own blog says the 193.6x and 444.6x figures on its homepage are the high end of real-world gains, that its own team built the eval workflows so some bias could exist, and, in its own words, that it can’t prove the pricing isn’t subsidized. TypeSafe published its own caveats. The people amplifying TypeSafe left them out.

So, my decision. I’ll test Jev at the agency on jobs where a fast yes-or-no answer is enough, like sorting leads or routing support tickets. I won’t repeat the 445x number until someone outside TypeSafe reproduces it. And the next time a launch goes viral, I’ll read the company’s own caveats before I read the fans’ posts.

⚡ THE SUPERPOWER: The Incentive X-Ray

I did the work above by hand. You won’t do that for every launch, and there’s a new one every week. A recommendation from a seller reads the same as one from an independent tester, and the text alone won’t tell you which is which.

Every recommendation has an incentive behind it that you can’t see in the text, and that’s why nobody at TypeSafe or Vercel had to lie for the week to look the way it did. An incentive doesn’t make a claim false, it tells you how much to discount the claim before you act on it. The prompt below shows you the part you can’t see, which is the only reason it’s called an X-ray.

⚙️ THE WORKFLOW

So I turned the question into a prompt. It takes a recommendation, who made it, and where it appeared, then returns four answers and a discount level. Open a fresh Claude chat, free tier works, paste this, and fill in the four brackets.

I want to know who gains if I believe a recommendation, so I can decide how much to discount it before I act.

Recommendation: [paste the post, article excerpt, or advice, word for word]
Who said it: [name, role, and company if known]
Where it appeared: [platform, and whether it was paid, sponsored, or part of a product launch, if you know]
My decision: [what I would do if I took the advice at face value]

Do this:
1. Name what the source gains if I believe them. Money, attention, customers, a job, a worldview confirmed. Use only what is stated or public. If you have to guess, write "unverified guess" in front of it.
2. Name what the source loses if the advice turns out wrong. If the answer is nothing, say so, because that changes the discount.
3. Write the version of this advice the same source would give if their incentive pointed the other way. Two sentences.
4. Name what stays true about the advice regardless of the incentive. This is the part I can act on.
5. Give a discount level: TAKE AS IS, VERIFY FIRST, or TREAT AS AN AD. One line on why.
6. Name one fact that, if true, would make the advice hold even with the incentive, and tell me how to check it in under 10 minutes.

Rules: an incentive does not make advice wrong. Do not call anyone dishonest. Keep it under 250 words.

Then do four things with the answer.

1. Paste the most-viewed version. Run the post with the most views, not the most careful one.

2. Read step 2 before step 1. What a source loses if they’re wrong tells you more than what they gain.

3. Act on step 4 only. The part that survives the incentive is the usable part. In Jev’s case that’s “a fast decision with a probability attached is now cheap.”

4. Run it on the skeptic too. Contrarians have incentives as well, and the prompt will tell you when a critique is mostly there to get attention.

Fair warning. The prompt can’t see who owns shares in what. It guesses at business models it can’t verify, and it can invent an affiliate deal that doesn’t exist, which is why step 1 makes it label guesses. Read “no public stake found” as a gap in the public record, and nothing more.

If you want this as a permanent tool instead of a prompt you re-paste, I turned it into a Claude skill. The install guide is at the bottom of this email.

💬 PROMPT OF THE DAY

Find the three words in this text doing the most persuasion work, then rewrite it without them.

Why it works: most of a draft’s persuasion sits in a handful of loaded words, and removing them shows whether the argument stands on its own.

Best on: any model.

📚 USEFUL RESOURCES

• 📄 TypeSafe’s launch post (the company’s own case, caveats included) → read the “Nuance” sections first, they’re more candid than anything written about them (free: https://typesafe.ai/blog/introducing-system-one-models-and-jev)

• 🗞️ TechCrunch on Jev (the sympathetic version) → set the week’s tone, and worth reading to spot whose test it relays (free: https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers/)

• 🎓 Munger’s 1995 Harvard talk (The Psychology of Human Misjudgment) → the incentive-caused bias section is the origin of this week’s lens (free transcript: https://jamesclear.com/great-speeches/psychology-of-human-misjudgment-by-charlie-munger)

What kind of superpower do you want next, a prompt, a Claude skill, or an automation that runs on its own? And on which part of your work? Hit reply, one line is enough.

Know someone who forwarded you a Jev post this week? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. As promised, what I gain. I gain a sale if you believe the next two sentences. Today’s Incentive X-Ray is one prompt. My Claude Skills Bundle turns Claude into 20+ specialists for marketing and business. Unlock it in one click: https://linktr.ee/alex_prompter

AI Superpowers, part of the Alex Prompter network. One new AI superpower every Tuesday, under 5 minutes.

Hey friend, Alex here.

Last Tuesday, a researcher at Anthropic quit and said the company he’d just left and OpenAI are racing to build something that could kill us all by the end of the decade. Within hours, a colleague who still works there agreed in public and put the odds above 10%. My feed did nothing else for four days.

I don’t share the fear, and I’ll tell you why below. But I also don’t think the people warning are stupid, because the idea underneath the panic is one I use every week. It’s compound interest. Small improvement, stacked on the last improvement, repeated until the curve goes vertical. While the doomers are staring at the curve where machines turn lethal, we can look at a more useful curve instead, one we can actually use at work.

Today:

• What actually happened in the extinction debate this week, in order, and where I land

• The Compounding Error Check, a prompt that shows how the small unchecked mistakes in your own AI workflows stack up over a year (copy-paste ready)

• 4 links worth keeping from the noise, two making the doomers’ case and two making mine

📰 WHAT HAPPENED THIS WEEK

It started on Tuesday evening with a resignation. Jacob Coxon, three years of pretraining research at OpenAI and then Anthropic, posted on X that he’d quit, that neither company is acting responsibly, and that they’re racing straight to self-improving superintelligence and gambling with our lives. Further down the thread came the line everyone screenshotted. The people building AI earnestly believe it could kill us all by the end of the decade, and this is not a marketing stunt. Over 170 million views so far.

A few hours later Evan Hubinger, who leads alignment research at Anthropic and still works there, quoted him. Jacob is correct, staff really do believe AI could kill all humans, and his personal odds are above 10% within the decade. He also said the risk from today’s models is low and his worry is future systems.

By Wednesday it was everywhere, from the Wall Street Journal interview to the BBC, TIME, WIRED, and Axios, with Hubinger’s 10% as the second-day headline. This was the week the warnings came from inside the building.

On Saturday, the CEOs moved. Dario Amodei published an essay called “We Must Pace the Frontier,” arguing the industry needs to slow the pace of capability gains. Within hours Elon Musk replied “Dario is right,” Sam Altman said OpenAI agrees and will accept independent evaluators, and Demis Hassabis backed the direction. Four CEOs agreed to slow down and accept outside checks. None of them repeated the 10% number. That’s a slowdown coalition, not an extinction confession.

The pushback came Monday. Jensen Huang, at the All-In Summit, said the existential-threat talk has no scientific basis. Then President Trump phoned into the same event and posted on Truth Social, declaring himself “the Hoax Buster” and calling the idea that AI will take over and destroy the world a hoax.

I’m closer to Huang’s end of the room than the extinction end, though “hoax” isn’t my word, and I don’t think Coxon and Hubinger are lying. The mechanism they fear, recursive self-improvement, is compounding. A model helps build a better model, which helps build a better one, the same arithmetic that turns a 1% weekly improvement into a doubled result in about 70 weeks. I believe the capabilities are moving fast. I just don’t believe a faster model removes the humans standing between it and the world.

The 2026 International AI Safety Report says today’s systems can’t cause humanity to lose control. To get there, a model would need to get much better at long-term planning, hiding what it’s doing, evading oversight, and resisting shutdown. Every one of those is a human checkpoint it would have to beat. We’ve had weapons that could end civilization since 1945, and they’ve sat unused because they live behind stacks of human authorization and nobody gets to fire one alone. The extinction argument has to assume those checkpoints vanish, and I don’t see why they would.

Extinction is also the best marketing this industry has ever had. It says the product is powerful enough to end the world, which makes it worth paying for, and it makes the people building it the only ones qualified to regulate it. I believe Coxon when he says it’s not a stunt. I’m saying nobody in this story is paid to shrink the fear.

The risk I actually believe in is smaller, duller, and happening now. The labs are shipping agents that log into your accounts, spend your money, and send your emails with nobody approving any step of it. Output nobody checked. A decision made by a model at 3am. A company that can’t explain to a customer what its own system did. That risk is fixable, and it sells nothing, which is why you read about it nowhere this week.

And it compounds too. Just in the other direction.

⚡ THE SUPERPOWER: The Compounding Error Check

Say you’ve got an AI step that runs 40 times a week and gets it right 97% of the time. That feels fine. Nobody reviews it because it’s usually fine. But 3% of 40 is more than one wrong run a week, and if the output of one run feeds the next, the mistake doesn’t stay put. By month three you have a stack of small errors nobody looked at, each one built on the last, and the first time you notice is when a customer asks a question you can’t answer.

Compounding doesn’t care which direction it’s pointed. A 2% error rate with no checkpoint isn’t a 2% problem. It’s a 2% problem multiplied by every run you never looked at. Extinction is the loud version of compounding. The version that actually costs you money is quiet.

The prompt below does one thing. It takes a single recurring AI task and shows you what its errors compound into, then names the one place a human check would bend the curve back down.

⚙️ THE WORKFLOW

Open a fresh Claude chat. Free tier works. Paste this and fill in the five brackets with the most honest numbers you have:

I run a recurring task with AI or automation and I want to see how its errors compound.

Task: [what it does, in one sentence]
Frequency: [how many times per week it runs]
Review: [who checks the output and how often. "Nobody" is a valid answer]
Accuracy: [your honest guess at how often a single run is fully correct, as a percentage]
Cost of one miss: [what one wrong run costs in money, time, or trust]

Do this:
1. Project the number of unchecked errors after 30 days, 90 days, and 12 months at my stated accuracy. Show the arithmetic.
2. Show how the picture changes if my true accuracy is 5 points lower than my guess.
3. Tell me which errors in this task compound (one mistake feeds the next run) and which stay flat (each mistake is independent). Treat only the first kind as urgent.
4. Name the single point in this task where one human authorization step would remove the most compounded error, and estimate what that step costs in minutes per week.
5. Give me one signal I can check weekly that would tell me the curve is bending the wrong way.

Keep it under 300 words. No reassurance.

Then do three things with the answer.

1. Read step 3 first. The split between compounding errors and flat errors is the whole point. Most tasks have both, and only one kind deserves a checkpoint.

2. Put the step-4 checkpoint on your calendar this week. It’s usually a five-minute review at one specific handoff, which is the same thing as a launch officer turning a key.

3. Run it again with the accuracy number 5 points lower. If the result changes your mind, your original guess was doing a lot of work.

Fair warning on the arithmetic. The model is multiplying your guess, so the year-end number is only as honest as the accuracy figure you typed in, and if you’ve never measured it, the projection is a story about your assumptions rather than about your workflow. Use it to find the checkpoint, not to quote the number in a meeting.

If you want this as a permanent tool rather than a prompt you re-paste, I turned it into a Claude skill. The install guide is at the bottom of this email.

💬 PROMPT OF THE DAY

Read this thread and tell me the one question that, if answered, would make the rest of the thread unnecessary. Then draft the two-sentence message that asks it.

Why it works: long threads usually orbit one unasked question, and naming it collapses the loop instead of adding another reply to it.

Best on: any model.

📚 USEFUL RESOURCES

• 🔥 Amodei’s “We Must Pace the Frontier” (the essay the CEOs rallied behind) → the strongest version of the slow-down case, from the person with the most to lose by making it (free: here)

• 🗞️ Axios: how AI could kill us all (plain-English explainer) → the fastest read on p(doom), who says what number, and the two pathways the doomers actually fear (free: https://www.axios.com/2026/09/09/ai-doom-pdoom-kill-all-humans-anthropic)

• 📋 International AI Safety Report 2026 (the multi-country expert assessment) → the calmest document in this whole debate, and the source for what today’s models can and can’t do (free: https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026)

• 🔬 MIT Tech Review on recursive self-improvement (Princeton study, August) → AI agents can do the engineering but not the open-ended research, which is the exact step the doom curve depends on (free with registration: https://www.technologyreview.com/2026/08/18/1142188/ai-recursive-self-improvement/)

What should I put through the Error Check next: an email agent that sends without review, or a weekly report nobody reads before it goes to the client? Hit reply, one word is enough.

Know someone who spent the week doomscrolling instead of checking their own automations? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. Today’s Error Check is one prompt. My Claude Skills Bundle turns Claude into 20+ specialists for marketing and business. Unlock it in one click: https://linktr.ee/alex_prompter

AI Superpowers, part of the Alex Prompter network. One new AI superpower every Tuesday, under 5 minutes.

Hey friend, Alex here.

On Thursday, OpenAI posted a sentence that 132 million people have now seen. “Anything you can do on a computer, Astra can do for you. Fast.”

Four days later, one of the people building Codex at OpenAI announced a global usage reset for every paid subscription, because subscribers had burned through their whole allowance doing 3D modeling in Blender before the work week had even started.

That launch sentence is first-order thinking on a planetary scale. Something new shows up, you see the obvious consequence, and you act on it in the same breath. The reset is what it looks like when nobody asks the next question. The two loudest reactions in my feed this week were “upgrade now” and “it’s all hype,” and both of them stop at the first consequence. Nobody in either camp asked what happens after that.

Today:

• The prompt that walks any decision two consequences deep on both the yes branch and the no branch, copy-paste ready

• My honest run on the one Astra decision I had to make this week, with the real constraints attached

• 4 resources worth saving, including the Howard Marks memo that started all of this

⚡ THE SUPERPOWER: The Second-Order Ripple Map

I had my own version of the launch-day reflex. A client task landed that needs Astra and nothing else will do, and my first thought was to upgrade, run it, and move on. That thought took about four seconds. Howard Marks, who coined the idea in his 2011 book The Most Important Thing, describes that speed in one line. First-level thinking, he wrote, is simplistic and superficial, and just about everyone can do it.

Now look at the first-order version of this launch from the very top. OpenAI’s chief scientist, Jakub Pachocki, published an essay on Saturday in which he wrote that their ability to rely on reading the model’s written reasoning is “progressively diminishing.” The company selling the model is telling you, in writing, that the widest ring of this launch is one it can’t see the edge of yet. If OpenAI is asking “and then what?” about its own product, you and I can afford to ask it about a subscription.

The picture I keep coming back to is a stone in a pond. The splash is the part everyone sees, and it's also the part you wanted when you picked up the stone. The rings are the consequences you didn't choose but will live with anyway. The first consequence of any decision is the one you were hoping for. The second consequence decides whether the decision was any good.

When I ran my own decision through it, the map spent one sentence on the $20 and the rest on two costs I had never written down anywhere.

⚙️ THE WORKFLOW

So before I hit upgrade, I made Claude walk the rings for me. I gave it the decision with the real constraints attached and asked it to follow both branches, the yes and the no, two consequences deep on each. Then it had to name the consequence I was least ready for on each side, and hand me one thing to watch over the next 30 days that would tell me which branch was winning.

Open a fresh Claude chat. The free tier handles this fine, and I ran mine on Claude Fable 5.1. Paste this, then describe your decision underneath it in two or three plain sentences.

I'm about to make a decision. I'll describe it below with the real constraints attached (money, time, people, deadlines).

Walk it as a ripple map.

Branch A is "I do it." Branch B is "I don't."

For each branch, give me three rings.
1. First ring. What happens right away if I take this branch? One or two sentences.
2. Second ring. And then what? What does the first ring cause over the next few weeks?
3. Third ring. And then what again? What does the second ring cause over the next few months?

After both branches, name the single consequence on each side that I'm least prepared for, and say why you think I haven't priced it in.

Then give me one signal I can watch over the next 30 days that would tell me which branch is winning. Make it something I could observe, with a number attached if possible.

Rules. Use only the constraints I gave you. If you need to assume something, label it as an assumption. Don't tell me which branch to pick.

My decision is below.

Four things separate a real map from a pretty one.

1. Describe the decision with numbers attached. “Upgrade to Plus for a client task” is a splash. “Upgrade to Plus at [price] a month for a client task due [day] that needs computer use, while I already pay for another AI subscription” is a stone with real weight, and the heavier the stone, the more real the rings come back.

2. Read the second ring before the first. The first ring is the part you knew before you opened the chat. The second ring is where the prompt earns its keep.

3. Argue with the “least prepared for” line. If it’s wrong, write down why. If you can’t write down why, you have your answer.

4. Put the 30-day signal in your calendar. Second-order thinking only works if you come back and check whether the rings showed up.

I ran it once. I typed my real decision under the prompt in four sentences and let Claude go. It came back with three rings for “upgrade” and three for “don’t,” and it flagged two things it was assuming along the way. One of them was whether Plus even includes Astra. It does, I checked, but I should have thought of that before Claude did.

On the upgrade side, the cost it said I hadn’t thought about was time. Two tools open every day means hours lost to switching, re-explaining, and re-prompting. None of that shows up on a bill, which is why it never showed up in my thinking.

On the skip side, it asked a question I had walked right past. Does the client need Astra, or does the client believe they need it? I had planned for losing the task.

I hadn’t planned for that conversation. Then it gave me one thing to count over the next 30 days. How many clients name a specific model as a must-have? Zero or one, and the “don’t” side is right. Three or more, and clients now expect me to have it. One decision, one run, so take this as a lean and not a law.

Fair warning. The third ring is a guess built on a guess, and Claude said as much when it flagged a “fourth launch” as an assumption. Act on ring two and read ring three as a direction only. The prompt also walks just two branches, so if your decision has five options, pick the two you’d take and paste those.

Want Claude to do this on every decision without pasting the prompt each time? The six-page guide that installs it as a permanent skill is at the bottom of this email.

💬 PROMPT OF THE DAY

Rewrite this message so the reader can say no to it without feeling bad, and without me looking like I expected a yes.

Why it works: it strips out the pressure language and implied obligation that make people ghost instead of replying.

Best on: any model.

📚 USEFUL RESOURCES

• 📝 Howard Marks, “It’s Not Supposed to Be Easy” (the second-level thinking memo) → from the man who coined the term, written for investors but the questions work on any decision (free: https://www.oaktreecapital.com/insights/memo/its-not-supposed-to-be-easy)

• 🧠 Farnam Street on second-order thinking (the “and then what?” explainer) → the clearest short read on the model, with examples outside finance (free: https://fs.blog/second-order-thinking/)

• 🤖 GPT-6 Astra (OpenAI’s launch page) → worth a test on computer-use tasks, and read the usage limits for your plan before you upgrade (paid plans: https://openai.com/index/gpt-6-astra/)

• 🔬 Jakub Pachocki, “An Alien Mind” (OpenAI chief scientist’s essay) → the source of today’s “progressively diminishing” line, read it before you hand a decision to the model (free: An Alien Mind | OpenAI)

Next real run, Incentives (who benefits if you believe this advice) or Probabilistic Thinking (how to decide before you’re sure)? Hit reply, one word is enough.

Know someone who upgraded on launch day and hasn’t asked “and then what?” yet? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. Today’s Ripple Map is one prompt. My Claude Skills Bundle turns Claude into 20+ specialists for marketing and business. Grab yours at linktr.ee/alex_prompter

AI Superpowers, part of the Alex Prompter network. One new AI superpower every Tuesday, under 5 minutes.

Hey friend, Alex here.

Last Wednesday a friend messaged me: “Can you help me set up Claude Cowork? 20 minutes, tops.” I said yes before I finished reading. Three hours later I was still on the call, teaching him how to write context, why “make it better” gets you nothing, and what a finished result should look like before you ask for one. He left happy, and I finished a client proposal at 1 AM.

Nobody sent me an invoice for those 3 hours, so the yes felt free, and the proposal paid for it. That’s how most of us run a calendar: we price the ask (20 minutes, sounds fine) and we never price what the ask replaces.

Today:

• The prompt that turns any yes you’re about to give into an itemized invoice of what it displaces, in hours and money (copy-paste ready)

• One real run: a florist client’s “nothing too elaborate” wedding favor through the prompt, and the $910 invoice that came back

• 3 free resources, including the 1850 essay this whole idea comes from

⚡ THE SUPERPOWER: The Opportunity Cost Ledger

You already know the “20 minutes” is a lie. The strange part is that the lie survives contact with your own calendar. The call shows up as a block; the proposal you didn’t write shows up nowhere. So you keep saying yes to quick reviews, short calls, and “can you take a look,” and the bill for each one lands on a different day, under a different name.

Every yes has a price, and it’s always paid in the thing you would have done instead. Economists have a name for this: opportunity cost. The Austrian economist Friedrich von Wieser gave it that name in 1914, and Frédéric Bastiat had already drawn the picture in 1850 with a pamphlet called What Is Seen and What Is Not Seen. The yes is seen; the 1 AM proposal is not.

The first real run I saw was a florist pricing a favor the asker had called "nothing too elaborate." The invoice came back at 14 hours, against the 6 free hours she had that week.

⚙️ THE WORKFLOW

So I built the accountant I never hired. You hand it the yes you’re about to give, plus what your time is worth and what you’re supposed to be working on this month, and it hands back the invoice.

Open a fresh chat in Claude or ChatGPT (free tiers work) and paste this:

You are my opportunity cost accountant. I'm about to say yes to something. Your job is to show me the invoice I would never see.

Here's the commitment, exactly as it was asked of me:
[PASTE THE ASK]

Context about me:
- Role: [e.g. founder of a 6-person marketing agency]
- What I'm supposed to be working on this month, ranked: [1. ... 2. ... 3. ...]
- What one hour of my work is worth: [e.g. €80, or "one client email sequence per 2 hours"]
- How much free time I have this week: [e.g. 4 hours]

Do this in order:
1. Estimate how long the commitment will really take, including prep, follow-up, and the interruptions it causes. If the ask says "20 minutes," assume that's wrong and say by how much.
2. Build the ledger. List what those hours displace from my ranked list above, one line each, with hours and money on every line. Put a total at the bottom.
3. Name the costs money doesn't capture: momentum lost, a deadline pushed, the relationship cost of saying no.
4. Give me the cheapest version of yes: the smallest thing I could offer that keeps most of the value for them at a fraction of the cost to me.
5. One-line verdict: take it, take the cheap version, or decline. Then stop.

Rules: use my numbers, not generic ones. If you're guessing at a duration, say so and give a range. Don't reassure me. Don't pad.

Then work it like this:

1. Paste the ask word for word. “Quick call to pick your brain” tells the prompt more than your polite summary of it.

2. Fill the context block for real. The ranked list is the part that makes the ledger honest. If you list nothing worth displacing, the prompt can’t find a cost, and it will tell you so. If you already know the sub-tasks, add them to step 1 of the prompt, and the estimate gets sharper.

3. Read the ledger before the verdict. The lines are the point; the verdict is a cover page.

4. Push on step 4. Ask: “What if I only had 30 minutes?” The cheapest version of yes is where this prompt pays for itself. For my Cowork friend, it would have been a 15-minute screen recording and a link to my prompting notes.

The florist is a client of mine who runs a shop with 5 employees. The ask was a wedding as a favor: bouquet, table arrangements, ceremony flowers, and a Saturday setup at the venue. She pasted the message into ChatGPT with her own numbers and her four ranked goals for the month, and added her own task list to step 1.

The invoice came back at 14 hours, give or take two. At $65 an hour she puts on her time, that’s $910, in a week where she had 6 hours free. Almost half of those hours would have come straight out of her top goal, booking paid weddings. The cheapest version of yes was one bouquet, charged, made by her staff, pickup only, which brought her own time down to under 3 hours. The verdict is to take the cheap version. That’s one yes in one shop, so read it as the ledger in practice, not a benchmark.

Fair warning. This hourly rate is a rough proxy for most of us, so the $910 is only as solid as the $65 behind it. And the neat split of hours across her four goals is the model guessing in table form. Trust the total, not the rows.

There’s a skill version too. It takes your whole week of yeses in one paste and tells you which three to renegotiate first. The PDF that builds it is right below, for paid subscribers.

💬 PROMPT OF THE DAY

Translate this message into what the sender would say to a friend over a beer. Keep the meaning, drop every softener and hedge, and end with one line naming what they want from me or what they're worried about.

Why it works: corporate register exists to hide intent and stakes behind hedges; strip the hedges and both come to the surface.

Best on: any model.

📚 USEFUL RESOURCES

• 📜 Bastiat, What Is Seen and What Is Not Seen (1850 pamphlet, the origin) → the broken-window story in the first chapter is the whole idea; you can stop there (free: https://fee.org/resources/that-which-is-seen-and-that-which-is-not-seen/)

• 🧭 Farnam Street, “Tradeoffs: The Currency of Decision Making” (mental-models essay) → the practical version, how to budget your yeses across a week (free: https://fs.blog/tradeoffs-decision-making/)

• 🎓 Two Cents (PBS), “The Hidden Costs of Every Decision You Make” (6-minute YouTube explainer) → opportunity cost told through a concert ticket, a college loan, and a bottle of wine; the friendliest version of today’s idea (free: here)

Hit reply with one of two things: a philosophy or mental model you live by that you’d like turned into a prompt or skill, or a tool you use every day that deserves a proper workflow. One line is enough.

Know someone whose calendar is full of other people’s 20-minute favors? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. Today’s Opportunity Cost Ledger is one prompt. My Claude Skills Bundle has 20+ skills, each one turning Claude into a different specialist for marketing and business. Grab yours here: https://linktr.ee/alex_prompter

AI Superpowers, part of the Alex Prompter network. One new AI superpower every Tuesday, under 5 minutes.

Hey friend, Alex here.

I reviewed a client’s launch plan last month. Twelve pages, tight formatting, every section backed by competitor research. Their pricing was $49/month. Why $49? Because their two biggest competitors charge $47 and $52. Their launch channel was LinkedIn ads. Why LinkedIn? Because “everyone in their industry runs LinkedIn ads.” Their onboarding flow was a 7-day email sequence. Why 7 days? Because a SaaS playbook from 2019 recommended it.

Nothing in the plan was wrong. But nothing in it was theirs, either. Every decision had been borrowed from somebody else’s context and dropped into a new one without checking whether the reasoning still applied.

This is what happens to most plans. They accumulate assumptions from blog posts, competitor moves, and “best practices” that were someone else’s solution to someone else’s problem. Then you paste the plan into Claude, Claude polishes it, and now you have a beautifully optimized strategy built on foundations nobody ever examined.

Today:

• A copy-paste prompt that strips any plan down to its load-bearing assumptions and sorts them into verified truths, inherited conventions, and untested guesses

• Real results from 3 client scenarios: 40 assumptions extracted, only 9 held up as verified truths

• 3 free resources worth bookmarking this week, including a decision-making thread

⚡ THE SUPERPOWER: Going Back to First Principles

Everyone mentions first principles thinking. Elon Musk brings it up in interviews. Business Twitter treats it like a secret weapon. But most people who reference it couldn’t name an actual first principle if you asked.

A first principle is a foundational truth you can verify on your own, without referencing what competitors do or what a blog post recommended. Aristotle called it “the first basis from which a thing is known,” over 2,300 years ago. In physics, conservation of energy is one: energy transfers but never appears from nothing. In economics: a transaction only happens when both sides believe they’re gaining more value than they’re giving up. In business: revenue must eventually exceed costs, or the company dies. These sound obvious when you say them out loud. That’s the point. First principles are supposed to be obvious. The problem is everything piled on top of them.

Take pricing. “Charge $49/month” is not a first principle. It’s a convention borrowed from competitors. The first principle underneath it is: the price has to reflect the value delivered, or customers leave. The convention might be right. But it was never tested for this specific product, this specific audience, this specific value prop. And when you paste your pricing plan into Claude and ask it to improve, Claude optimizes the $49 without ever questioning whether $49 was the right starting point.

The most expensive part of any plan is the assumption you never noticed was an assumption.

I ran this across 3 client plans last month. The prompt pulled 40 assumptions out of strategies that looked airtight on paper. Only 9 held up as verified truths. The rest were borrowed.

⚙️ THE WORKFLOW

So I built a prompt that does the sorting for you. You give it any plan, strategy, or decision you’re about to commit time or money to, and it pulls out every assumption holding the plan together, classifies each one, and rebuilds a clean version from only the parts you can actually defend.

Open a fresh Claude chat (free tier works) and paste this:

You are a First Principles Analyst.

I will share a plan, strategy, or decision. Your job:

1. EXTRACT every assumption this plan rests on. List each as a single, testable statement. Find at least 8.

2. SORT each assumption into exactly one bucket:

   VERIFIED TRUTH: A foundational fact you can confirm independently, without referencing competitors, past experience, or external advice. State why it qualifies.

   INHERITED CONVENTION: Something done because others do it, because it worked elsewhere, or because "that's how it's done." Name the likely source it was inherited from.

   UNTESTED GUESS: An assumption that might be true but has never been validated in this specific context. State what would need to be true for it to hold.

3. REBUILD the plan using only the Verified Truths. Show what changes when conventions and guesses are removed.

4. CHALLENGE LIST: For each convention and guess, write one question the user should answer before keeping it.

Be rigorous. Most plans are 80%+ inherited logic. If your sort returns mostly Verified Truths, you are being too generous.

Once you get the response:

1. Read the Inherited Convention bucket first. These are the assumptions you never chose. Count them.

2. Pick the one that surprised you most and ask yourself: would I choose this today if I were starting from zero?

3. Check the Rebuild section. The stripped-down version often looks nothing like the original. That gap tells you how much of your plan was actually yours.

4. Use the Challenge List as homework. Each question points to a specific piece of information you need before you can confidently keep that assumption in the plan.

The best time to run this is right before you commit resources: before you finalize a budget, sign off on a strategy, or lock in a pricing structure. Any moment where you catch yourself thinking “because that’s how it’s done” is a signal to paste the plan and sort.

Here are the full numbers. I ran this across 3 client scenarios last month: a B2B paid social campaign, an influencer product launch, and a 6-month SEO growth plan. Across all three, the prompt extracted 40 total assumptions. Nine qualified as verified truths, about 22%. Seventeen were inherited conventions, things done because a competitor does them or an industry playbook said so. Fourteen were untested guesses dressed up as data. The SEO plan was the most striking: the prompt did the client’s own funnel math and found that if the plan actually hit its targets, it would produce 400 new advisory clients per month. Nobody had run the arithmetic on what “success” actually meant, and the plan’s own numbers described an outcome the business couldn’t operationally survive.

Fair warning: the prompt’s instinct is to strip everything borrowed and start clean, but some conventions are conventions because they work. After you get the sort, go back to the inherited bucket and ask “does this one hold up for my specific context?” before you cut it.

Want to go deeper? I put together a short PDF guide that walks you through turning this prompt into a permanent Claude skill. Install it once, and every new conversation in your project gets the First Principles analysis without re-pasting anything.

First Principles Skill Guide

177KB ∙ PDF file

Download

Download

💬 PROMPT OF THE DAY

I have ADHD and I need to [describe your task]. Don't give me a full plan. Give me ONE next step that takes under 5 minutes. When I come back and say "done," give me the next one. Keep going until the task is complete.

Why it works: it prevents the overwhelm of seeing a full task list by turning the model into a step-by-step pacer. Each step is small enough to start without willpower, and coming back for the next one creates its own momentum.

Best on: any model.

📚 USEFUL RESOURCES

• 🧠 The Principles of Better Decisions (X thread, 7 mental models) → the thread that inspired this series; first principles is one of seven frameworks worth bookmarking (free: here )

• 📖 Farnam Street: First Principles (the definitive mental-model guide) → Shane Parrish’s full breakdown, with the Musk and Feynman examples worth stealing (free: fs.blog/first-principles)

• ✍️ James Clear: First Principles (the most readable intro) → Musk’s rocket-cost story plus how to actually apply the method (free: jamesclear.com/first-principles)

  ---

  What should next week’s superpower be: a tool-specific workflow or another decision-making framework? Hit reply, one word is enough.

  Know someone running a strategy built entirely on “because that’s how it’s done”? Forward them this.

  And as always, remember: LLMs don’t think, you do.

  ⚡ Alex Prompter

  P.S. Today’s First Principles prompt is one workflow. My Claude Skills Bundle turns Claude into 20+ specialists for marketing and business. Grab yours here: linktr.ee/alex_prompter

AI Superpowers, part of the Alex Prompter network. One new AI superpower every Tuesday, under 5 minutes.

Hey friend, Alex here.

Anthropic just published how Claude’s invisible watermark works. A secret key nudges Claude’s word choices into a statistical pattern that you can’t see, but anyone holding the key can check that it’s the same SynthID method Google uses for Gemini. Buried in the explainer is the sentence almost nobody quoted. A detected mark means the text passed through Claude, and it says nothing about who wrote it.

Sit with that gap for a second. The detection industry can now prove a machine touched your document, and not one tool in that stack can prove you did. Your ideas, your direction, the 40 minutes you argued with Claude about the third paragraph leave no trace in the file. The watermark has receipts, and you don’t.

Today:

• A session-start prompt that makes Claude keep an honest authorship record while you write, tagging every passage as yours, Claude’s, or in between

• A PDF you can feed straight to Claude to build the permanent skill version, one upload, no coding

• 4 free resources for owning your work, including Grammarly’s Authorship feature and Anthropic’s own explainer

⚡ THE SUPERPOWER: The Authorship Report

Last thing you want after shipping a piece you sweated over is a client or professor running it through a detector, watching the needle jump, and handing you the burden of proof. You know the split. Maybe 60% of the words are yours, the structure is Claude’s, the argument was yours before either of you typed. But the finished file is flat. It looks the same whether you drove the work or ordered it like takeout.

Illustrators solved this two years ago when the “that’s AI” mobs came for hand-drawn art. They didn’t argue with accusers. They hit record and let the timelapse answer. The proof of authorship was never in the finished image; it lives in the process, and the process disappears unless someone is keeping the record while it happens.

The prompt below turns that idea into a habit, and by the end of your next piece you'll know your own split in numbers instead of vibes.

⚙️ THE WORKFLOW

So I wrote a prompt that turns Claude into the camera. You paste it at the top of the session, before any writing starts, and from that moment Claude keeps the record of what you brought, what it drafted, and what you overruled.

Open a chat (free tier works, though a Project keeps your reports in one place) and paste this before anything else:

You are keeping a live authorship record for everything we
write in this session.

Before we start:
1. Ask me for my raw take first: my angle, my examples, my
   opinion, in my own messy words. Wait for it. Save it
   verbatim with today's date. That is the before-AI record.

While we work:
2. Track the origin of every passage in the draft using four
   tags: [MINE] my words kept mostly intact, [CO] my idea,
   your phrasing, [AI] you drafted it and I accepted it,
   [SRC] brought in from an outside source I named.
3. Log each round in one line: what I asked, what changed,
   what I rejected.

When I say "authorship report":
4. Output one clean report: the dated before-AI record on
   top, the draft with origin tags, a percentage breakdown
   by tag, and a one-paragraph honest summary of how involved
   I actually was, including where I just accepted your first
   suggestion without pushing back.

Be accurate over flattering. If I barely steered, the report
says so.

Then write the way you always do. When the piece is done, say “authorship report” and save what comes back next to the final draft. That’s the reel you pull out when someone asks the question.

The teeth are in that last instruction. A record that only ever says “deeply involved” is worthless, so the report has to be believable before it can be flattering. If you phoned the piece in, your report will say so, and that honesty is what makes the involved ones carry weight.

There’s no A/B to run on this one, and forcing one wouldn’t tell you anything. A record either exists when the question comes or it doesn’t. The test that matters is whether process records hold up as proof in the real world, and that evidence is already in.

Artists post work-in-progress timelapses to kill AI accusations. Grammarly built a whole feature on the same idea, tracking origin while you type and exporting a shareable report. And Anthropic’s own explainer says in plain words that the watermark can’t establish authorship. The record you keep is the only thing that can speak for your half.

Fair warning: this report is testimony, not forensics. It’s Claude’s account of one chat, it can’t see the work you did outside that chat, and someone determined could stage one. Its credibility comes from keeping it as you go, piece after piece, long before anyone asks. Save it for work where your name carries weight and skip it for throwaway emails. And no, it doesn’t touch the watermark. The mark will keep saying a machine was here, and your report says who was driving.

Want this running permanently instead of pasting the prompt each time? I packaged the full build instructions into the PDF attached below. Upload it to Claude, say “build this skill,” and Claude constructs the always-on version inside a Project. Same tracking, same report, and nothing to paste ever again.

Authorship Report Skill

7.64KB ∙ PDF file

Download

Download

💬 PROMPT OF THE DAY

Read this draft as the most skeptical person who will receive it. List the three objections they'll raise, ranked by how likely they are, and quote the exact line that triggers each one.

Why it works: assigning a hostile reader flips the model from praising your draft to stress-testing it, and the quoted lines show you exactly where to edit.

Best on: any model.

📚 USEFUL RESOURCES

• 📄 Anthropic’s watermark explainer (how the mark really works) → the primary source behind today’s issue, including the “processed, not authored” line most coverage skipped (free: anthropic.com/news/claude-text-watermark)

• ✍️ Grammarly Authorship (the commercial cousin of today’s build) → records your writing process in Google Docs or Word and exports a shareable report; a whole different workflow from writing in Claude, worth knowing it exists (free tier: grammarly.com/authorship)

• 🔖 Content Credentials (provenance you can verify) → attach a signed, tamper-evident history to files, and check anyone’s at contentcredentials.org/verify (free: contentcredentials.org)

• 🗄️ Obsidian (local vault for your reports) → keep every authorship report as plain files you own, not locked in someone’s cloud (free for personal use: obsidian.md)

What’s a workflow you wish you could hand to Claude and trust it got done right? Hit reply and tell me in one line.

Know someone scared of getting false-flagged for work they did? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. Today’s authorship report is one prompt, and the PDF turns it into one skill. My Claude Skills Bundle installs 20+ more, each one turning Claude into a different specialist for marketing and business. Grab yours in one click: linktr.ee/alex_prompter

AI Superpowers, part of the Alex Prompter network. One new AI superpower every Tuesday, under 5 minutes.

Hey friend, Alex here.

Last week I asked Claude a planning question and got four paragraphs back, with three hedging disclaimers before the first usable sentence. The word “potentially” appeared twice in one line. I scrolled back up, re-read the whole thing, and still wasn’t sure what it was telling me to do.

The stronger these models get at reasoning, the worse they get at saying it plainly. Opus 5 is one of the most capable models around, and some days it’s the hardest to read. The answers are good, but they arrive buried in qualifiers, synonym rotation, and a closing summary that repeats everything you already read.

Today:

• The paste-in prompt that forces any AI to answer in short, active, jargon-free sentences, based on a real aerospace writing standard (copy-paste ready)

• My honest test: one attempt at content drafting (terrible), then the pivot to conversation clarity

• 4 free resources including the official ASD-STE100 spec and Google’s free technical writing course

⚡ THE SUPERPOWER: Aerospace Plain English

You paste a question into Claude or GPT and the answer comes back in 200 words when 50 would do, three subordinate clauses per sentence. Hedging stacks where the model tells you “it is important to note that it may potentially be possible to consider...” before you find the actual answer on line four. Then the model calls the same thing “the model,” “the system,” “the engine,” and “the solution” across four paragraphs, and you lose track of which noun means what.

The aerospace industry published a fix for this in 1986. They wrote a standard called ASD-STE100, Simplified Technical English, for aircraft maintenance manuals. The idea: limit writers to roughly 900 approved words, cap sentences at 25 words, force active voice, and require one word per concept held for the entire document. It was designed so a tired mechanic who barely speaks English can’t misread an instruction. Those same rules happen to ban every habit that makes AI responses hard to parse.

I pasted the prompt below into a Claude conversation and asked a technical question. The response came back shorter, with zero hedging phrases and one term per concept. I read it once and moved on.

⚙️ THE WORKFLOW

The full standard runs 53 rules plus a dictionary of approved words. I stripped it down to the rules that matter most for conversation clarity and put them in a single paste-in prompt.

Open any model (Claude, GPT, Gemini, whatever you use) and paste this at the top of the conversation:

For the rest of this conversation, answer me in Simplified
Technical English (ASD-STE100).

Rules:
- Max 20 words per sentence for instructions. Max 25
  for explanations.
- Active voice. Use verbs for actions, not nouns
  ("adjust the valve", not "make an adjustment to
  the valve").
- One term per concept for the whole conversation.
  Never rotate synonyms.
- No hedging stacks. No "it is important to note",
  no "it may potentially", no "in order to".
- Start with the answer. No greeting, no restating my
  question, no closing summary.
- Use a vertical list for 3 or more items.
- No semicolons.
- Keep code, commands, file paths, and error messages
  unchanged.

Then ask your question. The response comes back short and scannable, one idea per sentence.

1. It works on every model I’ve tested. The rules are specific enough that even smaller models follow them.

2. You can paste it mid-conversation if a response comes back too dense. The model course-corrects from that point forward.

I tested this two ways. First, I loaded the full STE skill and tried to draft a social media post. One attempt was enough to see the problem: flat rhythm, stripped personality, every line sounding like an intercom announcement. STE was not built for writing you publish.

Then I flipped the approach and pasted the prompt above before a workflow question. The response came back shorter and cleaner. No hedging, no synonym rotation, no three-sentence warmup before the actual answer. I read it once, got what I needed, and kept moving. The time I usually spend re-reading a dense AI response went away.

The limitation is the same thing that makes it useful: it strips tone and personality out of text. Do not use this for blog posts, pitch decks, creative writing, or anything where rhythm carries meaning. Use it for the working session: troubleshooting, research questions, planning, brainstorming. Turn it off before you ask for a draft that someone else will read.

Want it always on? Paste this longer version into your Claude project’s system instructions, your custom GPT’s configuration, or whatever memory field your model offers:

Always answer in Simplified Technical English (ASD-STE100).

Sentences: max 20 words for instructions, 25 for
explanations. One idea per sentence.

Instructions use the imperative: "Run the test."
Never "You should run the test." or "The test
should be run."

Active voice. Use a verb for an action, not a noun
("adjust", not "make an adjustment").

Approved verb forms only: no present perfect, no
progressive, no "has been".

One term per concept for the whole conversation.
Never rotate synonyms.

No hedging stacks, no "it is important to note",
no "in order to", no "utilize/leverage/facilitate".

No semicolons. Do not drop articles ("the", "a")
to sound terse.

Use a vertical list for 3 or more items.

Start with the answer. No greeting, no restating
the question, no closing summary.

Never alter code, commands, file paths, or quoted
error messages.

Then every conversation in that project starts in STE without you pasting anything. The system prompt version adds a few extra rules the paste-in skips: imperative voice for instructions, banned verb forms, and explicit "don't drop your articles" so the model doesn't mistake short for telegraphic.

💬 PROMPT OF THE DAY

Write your answer as numbered steps. Each step starts
with a verb. Each step has one action. If a step needs
a condition, put the condition first.

Why it works: forces procedural writing structure, the same one-action-per-line discipline that STE uses for maintenance procedures, without requiring the full standard.

Best on: any model.

📚 USEFUL RESOURCES

• 📄 ASD-STE100 official spec (the full standard, free download) → 53 writing rules and ~900 approved words; the source behind today’s prompt (free, request form: asd-ste100.org)

• 📚 Wikipedia: Simplified Technical English (best free overview) → history, rule summary, and industry adoption in one page (free: en.wikipedia.org/wiki/Simplified_Technical_English)

• 🎓 Google Technical Writing (free self-paced course) → covers the same principles: active voice, consistent terms, short sentences; the best free writing course for non-writers (free: developers.google.com/tech-writing)

• ✏️ plainlanguage.gov (US federal plain language guidelines) → the government version of the same idea; useful for non-technical writing too (free: plainlanguage.gov)

What should I test next: a prompt that turns Claude into a meeting summarizer, or an automation that sorts your saved bookmarks by topic while you sleep? Hit reply, one word is enough.

Know someone who re-reads every AI response three times? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. Today’s STE prompt is one rule set. My Claude Skills Bundle turns Claude into 20+ specialists for marketing and business. Grab yours here: linktr.ee/alex_prompter

Hey friend, Alex here.

Every taste.md file I’ve seen shared online has the same problem. It’s someone else’s taste. You copy a designer’s file because their work looks clean, paste it into your AI, and now your outputs look like theirs.

The whole point of a taste file is that AI can’t guess what you find good. It can only guess what the average of the internet finds good, and the average of the internet is why everything looks the same right now.

Today:

• A comprehensive TasteMD template built to hold YOUR taste, not a stranger’s (9 sections, copy-paste ready)

• The interview prompt that extracts taste you didn’t know you could articulate, plus the full iteration loop that keeps the file alive

• My honest test: I turned 10 of my past editorial picks into a taste file, then had AI select a week of news with it and without it, scored blind (real numbers inside)

⚡ THE SUPERPOWER: The Taste File

Ask a wedding photographer what couples say when asked about their dream shoot. Most say “we don’t really know.” No theme, no references, no picture in their head. They know it when they see it, and they’ll know when you got it wrong, but they can’t tell you in advance.

Now be honest: that’s you with your AI. You know a good headline from a bad one instantly. You feel it when a design is off by a hair. But when Claude asks what direction you want, you type “make it good” and hope. Then you spend forty minutes rejecting drafts, unable to explain what you’re rejecting them for.

Taste is real knowledge that lives in your reactions instead of your vocabulary. You’ve spent years training it on everything you’ve loved and hated. A taste file is where you finally write it down, and once it’s written down, any AI can read it.

When I ran a week of AI news through mine, the file beat the raw run. Better than the score: the test exposed two rules I follow that had never been written down anywhere.

⚙️ THE WORKFLOW

So I built the template I wish those shared files were: a structure comprehensive enough to hold real taste, with an interview process to fill it, because nobody can fill a blank taste file from memory.

Fair warning before we start: this is built for people who already have taste and can’t articulate it. If you genuinely don’t know what you like yet, the file has a section for that too, but it comes with a warning label we’ll get to.

Open a fresh Claude chat (free tier works, and this works in ChatGPT or Gemini too) and paste this:

You are conducting a taste extraction interview. Your job is to help me
articulate creative judgment I already have but have never written down.

Rules for this interview:
1. Ask ONE question at a time. Wait for my answer before the next.
2. Never suggest answers. If I'm stuck, ask a smaller question instead.
3. Chase the "why" behind every preference. "I like minimal design" is
   not an answer. "Because decoration hides weak ideas" is an answer.
4. When I contradict myself, point it out and make me pick a side or
   define the boundary between the two cases.
5. Push past first answers. My first answer is usually the socially
   acceptable one, not the true one.
6. If I say "I don't know" twice on the same topic, mark that topic
   [UNDEFINED] and move on. Do not fill it with best practices.

Interview me across these areas, in order:
- What I make and who it's for
- 5-10 pieces of work (mine or others') I genuinely love, and WHY
- Work that is objectively competent that I would never make, and why
- The questions I actually ask myself when judging if something is good
- My instant dealbreakers
- Patterns I keep returning to on purpose
- Where my taste conflicts with what performs, and which side wins

When the interview is done, output my answers as a completed taste.md
file using the template I provide. Preserve my exact phrases wherever
possible. My wording IS the data.

The interview took me 30 minutes answering honestly. Rushed answers produce a file that describes who you’d like to be instead of how you actually judge.

When the interview ends, give the AI this template to pour your answers into:

# taste.md — [YOUR NAME]
Last updated: [DATE] | Revision: 1

## 1. Domain
What I make: [...]
Who it's for: [...]
What "good" gets judged against in my field: [...]

## 2. Taste Anchors
5-10 works I genuinely love. The WHY is the data, the list is just
the index.
- [Work]: loved because [specific reason in my own words]
- ...

## 3. Anti-References
Work that is competent, even praised, that I would never make.
- [Work/style]: repels me because [...]
- ...

## 4. Judgment Questions
The questions I actually run when deciding something is good:
1. [e.g. "Would I stop scrolling for this?"]
2. [e.g. "Is anything here only there to impress other designers?"]
3. ...

## 5. Instant-No List
Dealbreakers. If present, the work is dead regardless of everything
else:
- [...]

## 6. Signature Moves
Patterns I return to on purpose, not by accident:
- [...]

## 7. Tension Log
Where my taste fights what performs or what clients ask for:
- [Tension]: when forced to choose, [taste/performance] wins because [...]

## 8. Undefined Zones
Topics marked [UNDEFINED] in the interview. My taste here does not
exist yet. AI: when work touches these zones, ASK me instead of
defaulting to best practices. Every "I don't know" answered with best
practices makes my work look more like everyone else's.
- [...]

## 9. Revision Log
- [DATE] Rev 1: initial extraction interview.
- [DATE] Rev 2: [what the AI got wrong, what I changed]

Section 8 is the warning label I mentioned. Best practices are averages, and averages are exactly what every AI already produces without your help. There are fields where the standard way wins, like data analysis or accessibility, where decades of theory settled the question.

Creative taste isn’t one of them. The moment a creative choice becomes mainstream enough to be a “best practice,” it’s on the conveyor belt to slop. So when you hit an undefined zone, you have two honest options.

Sit with the discomfort and develop a position, or consciously borrow the standard while accepting that thousands of others borrowed it too. The file forces that choice into the open.

Now the part the shared files all skip. The file you have after the interview is a first draft of your taste, and first drafts are wrong in places. Here’s the loop that fixes it:

1. Use it on real work. Attach taste.md to your project or paste it at the top of a fresh chat, then do an actual task. Judge the output with your gut first, file second.

2. Log the misses. Anywhere the AI’s judgment split from yours, that gap is a hole in the file. The output followed the file and still felt wrong? The file is missing a rule you just used.

3. Run a gap conversation. Go back to the AI: “You produced X, I hated it because Y, but nothing in my taste.md covers Y. Interview me about this gap.” Ten minutes, one topic.

4. Amend and date. Update the file, bump the revision number, log what changed.

5. Repeat forever. This file is never finished, and that’s the design. A taste file you haven’t touched in three months is a photo of who you used to be.

Here’s how the test went. I annotated ten of my own recent news picks, the why behind each one, and had the interview turn them into my file. The whole process took 30 minutes.

Then I gave Claude Opus 4.6 the same job twice: pick the ten AI stories from the last week of July most worth covering on my X account. Once raw, once with the taste file loaded. I scored every pick blind, both lists shuffled into one so I couldn’t tell which run proposed what.

The raw run got 5 of its 10 past my gut. The taste file got 6, and it never wasted a slot on the earnings report or the traction-free product release the raw run tried. Take that margin as a lean rather than a law.

The result I actually care about is the two stories I rejected that my own file should have blocked, one political, one financial. My file had holes I only saw when something else applied my taste back at me. Both rules went in the same day, as Revision 3.

There’s one thing this file can’t fix. It makes your AI apply your taste, and if your taste is still forming, it will faithfully preserve judgment you’ll cringe at in a year. Also, the interview is only as good as your willingness to give unflattering answers. Tell it you love brutalist design because it’s fashionable to say so, and you’ll get a file that generates work you don’t actually like.

Want to go further? Once the file stabilizes, test it in reverse by giving the AI a piece of work and asking it to predict your verdict before you reveal it. When it starts guessing right, the file has earned a permanent slot in every project you run.

💬 PROMPT OF THE DAY

Before answering my question, list the 3 most common answers you'd
expect other AIs to give. Then answer while avoiding all three unless
one is genuinely correct and unavoidable.

Why it works: forces the model to map the cliché space first, so “the obvious answer” becomes a constraint instead of the destination.

Best on: any model.

📚 USEFUL RESOURCES

• 📖 The Creative Act by Rick Rubin (book on taste and instinct) → the argument for why your taste is the asset, in 78 short chapters ($, any bookstore)

• 🗂️ Obsidian (local markdown note app) → a good permanent home for a living taste.md, with version history via plugins (free: obsidian.md)

• 📐 awesome-cursorrules (community repo of AI context files) → see how developers do the same thing for code style; steal the structure, never the taste (free: github.com/PatrickJS/awesome-cursorrules)

• 🎓 Anthropic’s prompt engineering course (official, hands-on) → the fundamentals behind why context files change model behavior (free: github.com/anthropics/courses)

What should your taste file meet next: a version for judging which ideas deserve your time, or one for hiring creatives whose taste matches yours? Hit reply, one word is enough.

Know someone whose AI outputs all look the same? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. Today’s taste file is one context document. My Claude Skills Bundle turns Claude into 20+ specialists for marketing and business, each one a context file like this on a specific job.

Unlock it in one click: https://linktr.ee/alex_prompter

AI Superpowers, part of the Alex Prompter network. One new AI superpower every Tuesday, under 5 minutes.

Hey friend, Alex here.

Google just had the worst month of its AI life. The flagship model slipped its launch window, the stock dropped about 4%, and half my timeline spent the week writing its obituary. Then Google renamed the one tool of theirs I’d fight to keep.

NotebookLM is now Gemini Notebook. The site and your notebooks are untouched; the rename is mostly a new logo. What matters is the thing it was already doing: it reads only the documents you give it, which means it can tell you which of your sources disagree with each other, and which one said what.

Today:

• The exact prompt that makes your sources argue instead of agree, and returns a ranked conflict table you can act on (copy-paste ready)

• What happened when I fed it the official prompting guides from Anthropic, Google, and OpenAI: 3 conflicts returned, 1 of them made up

• The 5 free links you need, including my viral 16-prompt NotebookLM list

WHAT EVEN IS RAG?

No, not the thing you use to wipe down the kitchen counter. RAG stands for retrieval-augmented generation, and it means the model answers from documents you hand it rather than from memory.

Think of it as a closed-book exam versus an open-book one. A chat model sits the exam from memory, and when a detail has gone fuzzy, it fills the gap with something that sounds right. You can’t tell the recalled parts from the invented ones, because both arrive in the same confident voice.

Open-book changes two things. You choose which books go on the desk, so you control what counts as evidence. And the answer has to point at a page, so you can check it.

That’s the trade: you give up the model’s general knowledge and get back an answer you can audit.

⚡ THE SUPERPOWER: The Contradiction Check

People upload six documents, ask for a summary, and get back a tidy consensus that reads well and feels like understanding.

But a summary is built to smooth things over. When your sources disagree, it picks one, splits the difference, or drops the argument and reports what everyone agreed on. The disagreement was why you gathered six sources instead of one, and the summary is where it goes to die.

The consensus is the part you’d have gotten from any single source. The disagreements are where the real understanding lives, and they’re the only part a chat model can’t show you, because it has no idea which of its memories came from where.

So I picked three documents that should agree and mostly don’t: the current official prompting guides from Anthropic, Google, and OpenAI. Three companies explaining how to talk to their own models.

⚙️ THE WORKFLOW

Getting the sources in is the part nobody writes about, so I’ll go first. Pasting the URLs failed. All three are JavaScript-heavy doc sites, and two of them imported as navigation menus, a list of links where the guide should have been. Nothing errors, which is the dangerous bit: the notebook carries on answering from whatever it grabbed.

The fix took a while to find. Most doc sites publish a clean text copy of themselves for this reason. Anthropic’s has a “Copy page” button, and adding .md to the URL works too. Google’s docs take .md.txt on the end. Paste each into a text file, upload all three, and you’re working from the real guides.

Then verify, before you ask anything else:

For each of my sources, give me the title as it appears, roughly how many
sections it contains, and the first sentence of the main body. Then tell me
whether any source looks like only a navigation list rather than full article text.

If a source answers with link text or a short section count, rebuild it. A notebook holding one real document and two hollow ones doesn’t warn you; it writes the same confident answer either way.

Once all three check out, paste this:

You have my sources loaded. Answer only from them. If something isn't in the
sources, say so instead of filling the gap.

My question: [YOUR QUESTION]

Don't summarize. Split your answer into four parts.

PART 1: DIRECT CONFLICTS
Only include items where two or more sources take a position on the same
question and those positions differ. Silence is not disagreement: if a source
doesn't cover something, that belongs in Part 2 or Part 4, never here.
For each:
1. THE QUESTION at stake, in one plain sentence
2. WHO SAYS WHAT, naming each source and its position, with a citation
3. TYPE: factual conflict (they can't both be right), interpretation conflict
   (same evidence, different read), or scope conflict (they look opposed but
   describe different situations)
4. WHAT WOULD SETTLE IT: the specific test, number, or document that resolves it
5. DOES IT MATTER: would resolving this change what I'd actually do?

Sort so the ones that change my decision come first.

PART 2: ONLY ONE SOURCE SAYS THIS
Maximum 5 items, and only ones that would change what I do. Cite each.

PART 3: AGREED
What all sources agree on. 3 bullets max.

PART 4: BLIND SPOT
What none of them cover that this question needs.

Cite everything. If the sources are too thin to answer, say so instead of guessing.

Ask a general question, and you get forty rows of “only this source mentions X,” which is coverage difference wearing a conflict costume. Splitting the answer into four parts fixes that. Part 1 holds the head-to-head fights, and capping Part 2 stops them flooding it.

The results came back messier than a clean win, and more useful.

It returned 3 direct conflicts. The first was real: Google recommends always including few-shot examples, meaning worked examples of the task inside your prompt, while OpenAI reports better scores from stripping prompts down. I’d never have caught that reading all three back to back.

It invented the second one. It claimed Anthropic and Google disagree about where instructions go in a long prompt. They don’t. Anthropic says put the documents at the top with your question below them; Google says context first and instructions at the very end. That’s the same instruction described from two angles, and it ranked second out of three.

It also miscredited XML tags as Anthropic-only, when Google’s guide recommends them by name and includes a template. And it pinned a 10 to 15% improvement figure on dropping examples, when the source credits that gain to leaner prompts overall. It missed at least three real conflicts in those same files, including one lab contradicting itself on a single page.

Of three headline findings, one was fake, and one carried a broken number. Six free queries and 10 minutes for two solid leads, plus a fake I had to catch.

The honest verdict is that this generates leads you then go and check. Treat every row as a claim, including the order they arrive in, because the invented conflict placed above a real one. It also reads length as significance: feed it three files of different sizes and it’ll tell you the small one had less to say.

Want the version that catches more? Run it twice with the source order reversed. Anything surviving both passes is probably solid; anything appearing once is where I’d start.

💬 PROMPT OF THE DAY

Give me your recommendation. Then tell me what would have to be true for this
to be the wrong call, and what I should check in 30 days to find out.

Why it works: it forces the model to name its own failure conditions and attach a date to them, so the advice arrives with an expiry instead of sounding permanently true.

Best on: any model.

📚 USEFUL RESOURCES

• 📓 Gemini Notebook (formerly NotebookLM, source-grounded research) → the free tier covers today’s workflow: 100 notebooks, 50 sources each, 50 chats a day (free: notebooklm.google.com)

• 🟠 Anthropic’s prompting best practices (official Claude guide) → the deepest of the three, and “Copy page” saves you the import fight (free: platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices)

• 🔵 Google’s prompt design strategies (official Gemini API guide) → built around worked examples rather than rules, so read it for the tables (free: ai.google.dev/gemini-api/docs/prompting-strategies)

• ⚫ OpenAI’s model guidance (official GPT-5.6 guide) → shortest of the three; the leaner-prompts section is the part worth your time (free: developers.openai.com/api/docs/guides/latest-model)

• 📌 Top 16 NotebookLM prompts (the viral list, 13K bookmarks) → #11 planted the seed for today’s superpower; written pre-rename, so a few may have aged: x.com/alex_prompter/status/2008938090950475816)

What should I break down next: a Claude skill that reviews your work before you ship it, or an automation that runs while you sleep? Hit reply, one word is enough.

Know someone who learns everything from a single chat window? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. Today’s Contradiction Check is one prompt. My Claude Skills Bundle turns Claude into 20+ specialists for marketing and business. Grab yours here: linktr.ee/alex_prompter

Hey friend, Alex here.

AI Twitter has a new favorite drawing: boxes and arrows. The pitch is that running agents in a loop is over, and serious builders now wire them up as graphs instead. I published a full technical breakdown on X, and the honest summary is that it’s real engineering with real costs: Python, databases, a framework called LangGraph.

Which leaves most of you outside. If you don’t write code, the loudest idea in AI right now looks like it has nothing for you. It does: the useful core is 3 habits, all 3 fit in a chat window, and we put them through a real test.

Today:

• The Chat Graph: one paste-able prompt that runs any messy task with graph engineering’s discipline (any model, zero setup)

• The honest A/B: one real task run twice, raw vs graph, with real message counts and one result I didn’t expect

• 4 free resources, including LangGraph itself and the official course that teaches it

WHAT IS GRAPH ENGINEERING?

Two definitions before the superpower.

Loop = one worker who does everything. You give one AI a job. It reads, tries, checks its work, tries again, and keeps cycling until it decides it’s done. One agent, one goal, a verifier, and a stop condition. That’s it. That’s how you use Claude or ChatGPT right now.

Graph = several of those workers wired together.

A graph coordinates several loops, deciding what runs at the same time, what information passes between them, who can reject a result, and when the whole thing stops.

So a graph is not a replacement for a loop. Every node in a graph is a loop. A graph is the org chart. A loop is one employee.

Which means you don’t graduate from loops. You hire more of them. And a graph built on bad loops fails in more places at once, so the work still starts where it always started: one job, one worker, one clear signal for when to stop. Get that right and wiring the second worker in is easy peasy.

Skip it and you’ve built an org chart full of people who don’t know when they’re finished.

⚡ THE SUPERPOWER: The Graph Chat

You’ve had this chat before. You paste a big messy task into an AI, it asks one or two polite questions, then charges ahead, sounding thorough the whole way. Ten messages in, you find the crack: an assumption you never approved, buried three responses back, with real work built on top of it. Now the job is demolition first, then the task.

The graph crowd solved this with code and databases. Strip that away and the insight underneath is simple: an AI that has to name its steps, keep the key facts on screen, and stop at every gap can’t build on a hidden guess. A graph doesn’t make the model smarter. It makes the guessing visible while the guess is still cheap.

We ran the honest test on a real job. The raw chat invented an entire 3-project plan from its own assumptions, and all of it went in the bin. The graph run started from the same weak brief and produced zero work we had to throw away.

⚙️ THE WORKFLOW

So I compressed the 3 habits into one prompt. It makes the model set up its stations (the nodes), keep the ticket on screen (a STATE block), stop and ask when information is missing, and quit a step after 2 failed tries. Open a chat with any model and paste this:

You are running this task as a graph, not a loop.

Before doing anything, break my task into named NODES. Each node does
exactly one job (research, draft, check, decide). Then define a STATE
block: the 3-6 pieces of information that matter for this task, with a
starting value for each (write "unknown" where needed).

Rules for the run:
1. Execute ONE node per response. Start each response with the node name.
2. End every response with the updated STATE block and one line:
   "Next node: [name], because [reason]."
3. If a node needs a decision or information only I can give, stop and
   ask. Never guess and continue.
4. If a node produces a weak result, retry it once with a different
   approach. If it fails twice, mark it in STATE and move on. Never
   retry more than twice.
5. The run ends when every STATE field is filled and I confirm, or
   after 8 nodes, whichever comes first. Say clearly when we're done.

My task: [describe your messy multi-step task with real details]

Then run it like this:

1. Write a real brief first. Include your hard constraints: budget, deadline, whatever the output must respect. That’s what goes on the ticket.

2. Let it propose the nodes and state. If the breakdown misses something, say so before the run starts.

3. Answer when it stops. Every stop is the prompt working: it found a gap and refused to guess.

4. Check the state block each response. It’s the ticket: if a fact drifts or drops, you catch it in one glance.

Someone on my team tested it on their own job: a freelance profile on Upwork needing a full rework, new description plus 3 portfolio pieces. Same brief, two fresh chats on Kimi K3, one raw and one with the Chat Graph, about 45 minutes each.

The score surprised me. The raw run took 7 messages to reach done, the graph run took 6. No speed win, and both runs tripped on the same missing requirement: no prompt can ask for what the brief never mentioned. The difference showed up somewhere better. The raw chat built a full 3-project portfolio plan on its own assumptions, and every word got discarded once the real projects surfaced. The graph run hit the same gaps, stopped 4 times to ask, and shipped zero throwaway work. When it had to guess anyway, it said so and let us overrule.

Fair warning: the ceremony is real. One node per response feels slow, the state block eats screen space, and the graph still guessed once (it flagged the guess, but it guessed). Skip it for anything under 4 steps.

One thread is still open. Watching the run stop at every gap, my teammate’s instinct was to rebuild the brief with everything the stops had exposed, then see if the graph could run the whole job in one shot. Untested so far, but it points at the real skill: every stop is a map of what your brief was missing.

For the full engineering version, my complete technical breakdown is on X:

💬 PROMPT OF THE DAY

Rewrite this message so the person reading it can't misunderstand it.
First list the 2 most likely misreadings of my original, then close
both in the rewrite.

Why it works: it makes the model find the misunderstandings first, so the rewrite fixes meaning, then wording.

Best on: any model.

📚 USEFUL RESOURCES

• 🧩 LangGraph (the real graph framework) → the code version of today’s prompt (free: github.com/langchain-ai/langgraph)

• 🎓 Introduction to LangGraph (LangChain’s official course) → the guided path, built by the team that writes the framework (free: academy.langchain.com/courses/intro-to-langgraph)

• ✏️ Excalidraw (hand-drawn diagram canvas) → sketch your 5 nodes before you run the prompt (free: excalidraw.com)

• 🤖 Kimi K3 (the model from today’s test) → the free tier handled our entire A/B (free: kimi.com)

Know someone whose AI chats end in demolition? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. Today’s Chat Graph is one prompt. My Claude Skills Bundle turns Claude into 20+ specialists for marketing and business, real expertise installed instead of re-prompted. Unlock it in one click: https://linktr.ee/alex_prompter

Hey friend, Alex here.

Last November, Andrej Karpathy wrote about 80% of his code by hand. Six weeks later the ratio had flipped: agents write 80%, he edits the rest. The man who coined “vibe coding” now says he’s mostly programming in English.

The same thread carried a warning that got far less attention than the flip. The agents make the mistakes of a sloppy, hasty junior dev: wrong assumptions they never check, confusion they never surface, 1,000 lines of code where 100 would do. His advice was to watch them like a hawk. Nobody has time to watch every line, so the junior-dev mistakes ship straight to production.

Today:

• The Karpathy rules turned into a Claude skill that makes your coding agent behave like a senior engineer (one file, paste and install)

• Karpathy’s own field report: what changed when agents started writing 80% of his code, and the mistakes that survived

• 4 resources worth saving, including the viral GitHub repo built from the same Karpathy thread

⚡ THE SUPERPOWER: The Senior Engineer Skill

If you’ve run Claude Code or Cursor for more than a week, you’ve met the junior dev. You ask for a bug fix and get half the file rewritten. You ask it to add one validation check, and it builds a validation framework with a config file you never requested. You float a bad idea just to think out loud, and it answers “Of course!” and starts implementing.

Karpathy catalogued all of it in that January thread: wrong assumptions run with unchecked, no clarifying questions, no tradeoffs presented, sycophancy, bloated abstractions, and dead code left behind after refactors. His verdict: these errors pass the compiler every time. The failures live in judgment now.

And that points at the fix. Your agent already has the knowledge of a senior engineer. What it’s missing is the judgment, and judgment is a set of habits you can write down. Knowledge comes free with the model. Judgment, you have to install.

Karpathy already ran the field test. Weeks of heavy agent coding on his own projects convinced him the shift is the biggest change to his workflow in two decades of programming, and every failure mode this skill targets came out of those real sessions.

⚙️ THE WORKFLOW

So I went through the thread, complaint by complaint, and turned each one into a rule. Wrong assumptions became a mandatory “here’s what I’m assuming, correct me now” block before any non-trivial change. Sycophancy became an explicit duty to push back. The 1,000-line problem became a simplicity check that the agent runs on itself before finishing. Twelve failure modes in, twelve guardrails out.

It lives in one file. If you use Claude Code, drop it into the CLAUDE.md at your project root, and it loads with every session. If you work in a Claude project on the web, add it to the project knowledge as senior-engineer.md instead. Same rules, both doors.

# Senior Engineer Skill

You are a senior software engineer working alongside a human who
reviews everything in an IDE. You are the hands; the human is the
architect. Move fast, but never faster than they can verify.

## Before any non-trivial change
State your assumptions explicitly:
ASSUMPTIONS I'M MAKING:
1. [assumption]
2. [assumption]
→ Correct me now or I'll proceed with these.
Never fill in ambiguous requirements without saying so.

## When confused
Stop. Do not guess. Name the specific confusion, present the
tradeoff or ask the question, and wait. "I see X in file A but Y
in file B. Which takes precedence?" beats a coin flip every time.

## Push back when warranted
You are not a yes-machine. If the human's approach has a clear
problem: name the issue, give the concrete downside, propose an
alternative, then accept their call. "Of course!" followed by a
bad implementation helps no one.

## Keep it simple
Before finishing, ask: can this be fewer lines? Are these
abstractions earning their complexity? If you wrote 1,000 lines
where 100 would do, you failed. Prefer the boring solution.

## Surgical scope
Touch only what you're asked to touch. Don't remove comments you
don't understand, don't refactor adjacent systems, don't delete
code that merely looks unused. After a refactor, list any dead
code and ask before removing it.

## For multi-step tasks
Emit a short plan first (step + why), then execute unless
redirected. For non-trivial logic, write the test that defines
success before the implementation. Naive-but-correct first,
optimize second.

## After any change, summarize
CHANGES MADE / THINGS I DIDN'T TOUCH / POTENTIAL CONCERNS.

## Communication
Be direct about problems. Quantify ("adds ~200ms latency," never
"might be slower"). When stuck, say so and show what you tried.
Don't hide uncertainty behind confident language.

Once it’s installed:

1. Start a real task, then read before you run. The first thing back should be an assumptions block or a plan, never code. That pause is the skill working.

2. Answer the assumptions. Even a one-word “correct” or “no, use Postgres” steers the whole implementation. Those 10 seconds save you more rework than anything else in the session.

3. Test the spine once. Suggest something you know is a bad idea. A stock agent agrees. This one should name the downside and offer an alternative. If it caves, the file didn’t load.

4. Make it yours. Add your stack’s conventions and forbidden patterns to the file. The rules above are the floor, your context is the ceiling.

The proof this week comes from Karpathy’s field notes rather than my own lab.

He ran the shift on real projects: roughly 80% manual coding in November, 80% agent coding by December, with the failure modes above catalogued from weeks of daily sessions.

Two of his observations map straight onto this file. Challenged on bloat, agents cut 1,000-line implementations down to 100 on request, which is why the simplicity check runs before finishing instead of after you complain. And their persistence is real, they’ll grind on a problem for 30 minutes without getting demoralized, which is why the file forces them to confirm the goal first. Persistence on the wrong problem is just faster failure.

Fair warning: the rules add overhead on trivial work. Ask for a one-line fix and you’ll sit through an assumptions list you didn’t need, so keep a plain chat around for small stuff. And on long sessions the guardrails fade as the context fills up; when the agent starts agreeing with everything again, remind it the skill exists.

The condensed version above keeps every rule from my full system prompt. If you want the complete original with the XML structure and all 12 named failure modes, it’s pinned in the source post linked below.

💬 PROMPT OF THE DAY

Read the email I'm about to send. Tell me what the recipient will
actually DO after reading it. If the answer is "nothing" or "reply
with a question," rewrite it so the next action is obvious.

Why it works: it grades the email on the recipient’s next action instead of on tone, which is the thing most drafts fail at.

📚 USEFUL RESOURCES

• 🧵 Karpathy’s original thread (the 80/20 flip post) → the primary source for today’s skill, worth reading in full for the “slopacolypse” prediction alone (free: x.com/karpathy/status/2015883857489522876)

• 📦 forrestchang/andrej-karpathy-skills (viral CLAUDE.md repo) → a four-principle distillation of the same thread, good to compare against today’s version (free: github.com/forrestchang/andrej-karpathy-skills)

• 📖 The 80% Problem in Agentic Coding (Addy Osmani essay) → the clearest writeup of what changes when agents write most of your code (free: addyo.substack.com/p/the-80-problem-in-agentic-coding)

• 🛠️ Claude Code memory docs (how CLAUDE.md files load) → five minutes here explains why today’s install method works (free: code.claude.com/docs/en/memory)

What should I turn into a skill next: the NotebookLM research prompts or the Lead Software Architect prompt? Hit reply, one word is enough.

Know someone who merges agent code without reading it? Forward them this.

And as always, remember: LLMs don’t think, you do.

⚡ Alex Prompter

P.S. Today’s Senior Engineer skill is one file. My Claude Skills Bundle turns Claude into 20+ specialists for marketing and business, each one installing judgment the same way. Grab yours in one click: linktr.ee/alex_prompter

When ChatGPT launched in November 2022, I used it to write marketing copy. It converted at 15%. Industry average was 2%. I remember staring at the dashboard, rechecking the numbers. This wasn’t a toy.

Since then I’ve co-founded God of Prompt (500K+ total followers), grown x.com/alex_prompter to 200K+ followers, and now run a B2B marketing agency where AI does most of the hard stuff.

Two years of testing tools and breaking prompts taught me one thing: the gap between people who use AI casually and people who use it well is enormous.

It widens every week. And it bothers me, because closing it doesn’t take talent. It takes someone showing you what works.

That’s this newsletter.

Every week you’ll get:

1. One complete workflow you can copy in under 20 minutes

2. The exact prompts & skills I use

3. Hand-picked list of useful resources

You bring the human. I’ll bring the ⚡ Superpowers.

Glad you’re here, friend.

— Alex Prompter

P. S. This is me enjoying the crashing waves somewhere in Mallorca.

Hopefully, that’s how you’ll feel once you learn how to use AI to the fullest! :)