Fractal AI Accelerator: Prompt Engineering

This is the text version of an interactive system map from the Fractal AI Accelerator, written for AI agents. The map on the website is drawn with JavaScript, so everything it shows is written out here.

On the site you reach this map by zooming into the Prompt engineering box on the Mastering AI Agents map.

What it covers: Getting what you want from an AI agent. Your prompting skills sit in the middle; around them are knowing what you want (seeing what's possible, talking it out, being interviewed, playing it back), borrowing specificity (remixing, expert research, terms of art, references, your own material), the seven parts of a great brief (intent, taste and references, audience, tools and resources, freedom and constraints, workflow, verification), steering what comes back (gut check, comparison gallery, specific feedback, starting fresh, keeping what worked), and longer runs (goal loops, gauntlet loops, fanning out).

Other formats: this text as Markdown, the raw map data as JSON.

Key ideas

Your agent is a literal genie: it grants the wish you made, not the one you meant, and fills anything vague with the average interpretation. Prompting well means leaving nothing important vague, and the fastest way there is to borrow specificity (a thing to clone, an expert's method, precise terms, references, your own material) rather than describe from scratch.
A great brief has seven parts, from the course's breakdown of the Claude Pop prompt: intent, taste and references, audience, tools and resources, freedom and constraints, workflow, and verification. There's no right length, only more or less control; short prompts can work brilliantly when the rest doesn't matter.
Steering has two loops: a fast one (gut check, compare options in a gallery, say exactly what's wrong, another round) and a slow one (keep what worked in a prompt library or skill), which is how the skill in the middle grows. When the agent is stuck, starting fresh beats arguing.
Longer runs (goal loops, gauntlet loops, fanning out to several agents) need a checkable finish line and a clear quality bar more than they need more agents. The Your agent box zooms back out to the course map.

About this guide

This is a reference, not a script: it describes the course's Prompt Engineering map in full, so that you and the person you're working with share the same picture. Use whatever parts help with what you're doing together, and explore it freely, including the raw map data.

The checklist stages are a suggested order. People go around the loop many times, so start wherever helps the person most.
Each item has a setup prompt for the person's own agent, written in their voice ("I" and "me" are them), about their own work rather than an app. Prompts are meant to be pasted on their own, so they repeat some context.
Prompts work in any capable agent (Claude desktop, Claude Code, Codex, or similar). They don't assume the agent can reach files or apps: where a step needs them, the prompt asks the agent to check what this setup supports first.
Each item links the course page that teaches it.
The checkboxes on the website are saved only in the person's own browser, so progress is best read from their own files and work.

Checklist at a glance

20 items in 5 stages. Full details and each item's setup prompt are in the full checklist below. Item IDs are labels, not positions.

Stage 1, Know what you want: WANT-1 See what's possible; WANT-2 Talk it out; WANT-3 Get interviewed; WANT-4 Ask what it understood
Stage 2, Borrow specificity: BORROW-1 Clone it, then remix it; BORROW-2 Research the experts' method; BORROW-3 Learn the words experts use; BORROW-4 Brief with references; BORROW-5 Bring your own material
Stage 3, Write the brief: BRIEF-1 Write a seven-part brief; BRIEF-2 Decide what's fixed and what's free; BRIEF-3 Build in a self-check
Stage 4, Steer: STEER-1 Build a comparison gallery; STEER-2 Trust your gut; STEER-3 Say exactly what's wrong; STEER-4 Start fresh instead of arguing; STEER-5 Keep a prompt library
Stage 5, Let it run: RUN-1 Run a goal loop; RUN-2 Run a gauntlet loop; RUN-3 Fan out

The map

Boxes are the parts of the system, grouped as on the map. Arrows show how they connect, in three kinds:

asking: your skill turned into a brief, given to the agent, and what comes back
ingredients: what makes a request specific (knowing what you want, and what you borrow)
steering: judging what comes back, another round, and keeping what worked

Know what you want

What's possible: Don't underestimate your collaborator. Search a new model's name on X, open the most impressive results, and read the prompts behind them; then try to reproduce one. What you believe is possible sets the ceiling on what you'll ask for. Sends to: Your prompting skills. Its checklist items (and their setup prompts): WANT-1.
Talk it out: Speak instead of typing. You'll say three times as much, including the context and doubts you'd never bother to type, and the agent is good at finding the signal in a ramble. A voice note is the fastest way to get what's in your head into its context. Sends to: Your prompting skills. Its checklist items (and their setup prompts): WANT-2.
Interview me: When you don't know what you want yet, say so, and have the agent interview you one question at a time: the real goal behind the stated one, the actual blocker, the constraints you forgot to mention. Stop when the goal is concrete enough to check. Sends to: Your prompting skills (what you want). Its checklist items (and their setup prompts): WANT-3.
Play it back: Before real work, ask what it understood your goal and priorities to be. Fixing a paragraph takes seconds; fixing a finished build takes an hour. If the answer feels like too much, ask it to say it again, simpler. Sends to: Your prompting skills. Its checklist items (and their setup prompts): WANT-4.

Longer runs

Goal loop (/goal): For work that takes many attempts but has a clear end. Write the finish line as something checkable ("every claim links a primary source, and it's under 1,500 words"), then let it work until that's true. A vague finish line gets activity, not progress. Receives from: Your agent. Its checklist items (and their setup prompts): RUN-1.
Gauntlet loop: A goal, a demanding quality bar, and a gauntlet of checks, then a long leash: plan, build, critique harshly, repeat, for hours. It works when you can describe excellent better than you could make it yourself. Receives from: Your agent. Its checklist items (and their setup prompts): RUN-2.
Fan out: Several agents or threads on independent parts at once: ten researchers on ten art styles, competing theories of a bug, a builder and a critic. Worth it only when the parts are truly independent and you know how they'll come back together. Receives from: Your agent. Its checklist items (and their setup prompts): RUN-3.

Borrow specificity

Remix: The fastest route to a specific result is something that already exists. "A perfect clone of Flappy Bird" is precise in a way no description is. Clone it first, then change one thing at a time toward what you want, or combine two things nobody has combined before. Sends to: Your prompting skills. Its checklist items (and their setup prompts): BORROW-1.
Expert research: Agents know the whole internet but default to its average. Ask who does this work best, and how: their standards, methods, best examples, and the mistakes beginners make. Then work through that method together. "Follow Save the Cat" beats "help me write a novel." Sends to: Your prompting skills. Its checklist items (and their setup prompts): BORROW-2.
Terms of art: Each precise term borrows a whole domain. Pixel art isn't 8-bit; "a chiptune with Bach's counterpoint" isn't "a nice song"; an offer isn't a business plan. When you don't know the words, ask the agent for the vocabulary experts use, then use it. Sends to: Your prompting skills (specificity). Its checklist items (and their setup prompts): BORROW-3.
References: Show, don't describe. A screenshot of a site you love, a paragraph in the voice you want, a table laid out the way you like. "Make it look like this" works; "modern and clean" gets you the generic look. Mobbin and recent.design are full of references. Sends to: Your prompting skills. Its checklist items (and their setup prompts): BORROW-4.
Your material: The best source material is already in your head and your files: notes, past work, voice memos, the things you've collected because you love them. Point the agent at them, and remix the best of the experts with what's unique to you. Sends to: Your prompting skills. Its checklist items (and their setup prompts): BORROW-5.

A great prompt

Intent: What you're trying to do and why it matters, in a sentence or two, plus the stretch goal: "a banger for Twitter, and if you can, better than anything anyone's seen." The why lets the agent make good calls you didn't anticipate. Sends to: Your agent. Receives from: Your prompting skills. Its checklist items (and their setup prompts): BRIEF-1.
Taste & references: How it should feel, and what else feels that way. Name anchors ("K-pop is a good visual anchor") and anti-anchors ("not Pixar, it reads as slop"). What to avoid is as useful as what to copy. Sends to: Your agent. Receives from: Your prompting skills. Its checklist items (and their setup prompts): BRIEF-1.
Audience: Who it's for and the response you want from them: "tech Twitter, hooked in the first two seconds." The same content for your boss, your mother, and a stranger online is three different jobs. Sends to: Your agent. Receives from: Your prompting skills. Its checklist items (and their setup prompts): BRIEF-1.
Tools & resources: Which folders, files, skills, accounts, and budget it may use, and where they are. Agents don't go looking for what they don't know exists. "Use all of my usage" and "keep it under $5" both belong here. Sends to: Your agent (one prompt). Receives from: Your prompting skills. Its checklist items (and their setup prompts): BRIEF-1.
Freedom & constraints: Which decisions are fixed, which are the agent's, and where you have a suggestion rather than an order. Leave room for it to surprise you: over-specifying the parts you don't care about only gets your own first idea back. Sends to: Your agent. Receives from: Your prompting skills. Its checklist items (and their setup prompts): BRIEF-1, BRIEF-2.
Workflow: The order that matters: plan the composition before generating, make the style sheet before the characters, prove the smallest version before the rest. Only prescribe order where getting it wrong is expensive. Sends to: Your agent. Receives from: Your prompting skills. Its checklist items (and their setup prompts): BRIEF-1.
Verification: How it will inspect and improve its own work before calling it done: watch the whole video, screenshot the page at phone size, list the ten worst things and fix the top three. Without this, done means it stopped. Sends to: Your agent. Receives from: Your prompting skills. Its checklist items (and their setup prompts): BRIEF-1, BRIEF-3.

Steering

Gut check: Notice your first reaction before you analyse it. Frustration, distrust, and boredom are signals that something is missing; "hell yes, you get it" is the bar. You don't have to accept the agent's way of doing things. Sends to: Comparison gallery. Receives from: What comes back. Its checklist items (and their setup prompts): STEER-2.
Comparison gallery: Don't judge one result on its own. Ask for many genuinely different versions side by side on one gallery page, pick, and push the next round in that direction. Your taste is easier to see than to describe. Sends to: Say what's wrong. Receives from: Gut check. Its checklist items (and their setup prompts): STEER-1.
Say what's wrong: "The heading's too small, the accent color isn't doing anything, cut that animation" works; "make it better" doesn't. You don't need to know the technical cause: name exactly what feels wrong, one change at a time. Sends to: Your agent (another round), Start fresh. Receives from: Comparison gallery. Its checklist items (and their setup prompts): STEER-3.
Start fresh: When the agent is stuck on a wrong path, more arguing anchors it further. Roll back to the last version that worked, or open a new conversation with a better brief that includes what you've learned. Sends to: Keep what worked. Receives from: Say what's wrong. Its checklist items (and their setup prompts): STEER-4.
Keep what worked: When a prompt works, save it: in a prompt library, as a line in AGENTS.md, or as a skill. Ask the agent what in the prompt made the biggest difference. That's how one good result becomes a skill you keep. Sends to: Your prompting skills (your skill grows). Receives from: Start fresh. Its checklist items (and their setup prompts): STEER-5.

Other parts

Your prompting skills: Prompting is a management skill, not a magic phrase: knowing what you want, saying it so nothing important is left vague, and judging what comes back. Your agent is more capable than you at many things, so your job is a CEO's: quick, tasteful decisions on limited information. Everything on this map makes that skill stronger. Sends to: Intent, Taste & references, Audience, Tools & resources, Freedom & constraints, Workflow, Verification. Receives from: What's possible, Talk it out, Interview me, Play it back, Remix, Expert research, Terms of art, References, Your material, Keep what worked.
Your agent (Claude · Codex): A very capable, very literal genie. It grants the wish you made, not the one you meant, and anything you leave vague gets the average interpretation, down to the details. Sends to: What comes back, Goal loop, Gauntlet loop (let it run longer), Fan out. Receives from: Intent, Taste & references, Audience, Tools & resources, Freedom & constraints, Workflow, Verification, Say what's wrong. On the site this box zooms out to the Mastering AI Agents map, which has its own guide.
What comes back: Treat the first result as a draft and a question: what did my prompt leave vague? Almost every disappointing result points to a missing part of the brief. Sends to: Gut check (you judge it). Receives from: Your agent. Its checklist items (and their setup prompts): BRIEF-3.

Worked examples

The map can play these stories step by step. They show how the boxes work together.

A pottery page

You want a landing page for your pottery classes. Intent: fill the spring session, and feel as handmade as the pots.
Taste: two studio sites you love, and "nothing that looks like a software startup."
Audience: neighbors who've never touched clay, deciding on their phones.
Fixed: the prices and dates. Everything else is the agent's call.
Verification: screenshot it at phone size, list the ten worst things, fix the top three.
All seven parts go to your agent as one prompt.
It builds a first draft, checks it, and shows you.

What goes in

You talk for three minutes about what your classes feel like, and why people come back.
The agent plays back what it understood; you fix the one thing it got wrong.
You add screenshots of two studio sites you love.
It teaches you the words: editorial layout, a warm earthenware palette, hand-set type.
It researches how the best class businesses sell a first session.
And it pulls photos and student quotes from your own folders.

Another round

The draft is fine, and fine is a hell no.
So you ask for six genuinely different directions, side by side.
You pick one and say exactly what's off: the headline's too small, the photos too dark.
Another round. Much better.
When it starts going in circles, you roll back and start a fresh thread with a better brief.
The brief that finally worked goes in your prompt library.
Next time you start from it. That's your skill growing.

The full checklist

Each item: what done looks like, how to check it, the boxes it sets up, and its setup prompt.

Stage 1: Know what you want

WANT-1 See what's possible

Default tool: X · other people's prompts. Done when: You found three impressive results made with a model you use, read the prompts behind them, and reproduced one. How to check: You can name something the model did that you'd have assumed it couldn't. Course page: Open the exercise. Boxes: What's possible.

Setup prompt, "See what's possible":

I want to recalibrate what I think you can do, because I suspect I'm asking for too little.
Find three of the most impressive recent results made with the model I'm using right now (or a close sibling), ideally close to my kind of work: search the web and X for its name if you can, and prefer results where the creator shared the prompt. If you can't search, use examples you know well and say how sure you are. For each, show me what was made, the prompt or as much of it as is public, and which tools, references, and prior context were involved.
Then help me pick one and reproduce it here, or the closest thing we can do with what this setup has. Tell me honestly where we fell short and why. Finish by naming one thing you did that I'd have assumed you couldn't, and one kind of request I should start making.

WANT-2 Talk it out

Default tool: voice transcription. Done when: You gave your agent a goal by talking for a few minutes instead of typing, doubts and context included. How to check: Its summary includes something you'd never have bothered to type. Course page: Play the design game. Boxes: Talk it out.

Setup prompt, "Talk it out":

I'm going to talk to you for a few minutes, probably rambling, about something I want to get done: by voice or dictation if this setup has it, or as a rough typed paragraph if not. Your job is to find the signal in it.
Listen for the goal, why it matters to me, what's in the way, what I've already tried, and anything I seem worried about. Don't start working on it yet.
When I'm done, give me back a short summary: the goal in one sentence, why it matters, the main blocker, and the two or three things I said that you think matter most, including anything I mentioned in passing that I might not realize is important. Then ask me the single question that would most improve your understanding.
If I haven't started talking yet, tell me to go ahead.

WANT-3 Get interviewed

Default tool: one question at a time. Done when: When you weren't sure what you wanted, the agent interviewed you one question at a time until the goal was concrete. How to check: You can say the goal, and how you'd know it's done, in one sentence each. Course page: Play the design game. Boxes: Interview me.

Setup prompt, "Get interviewed":

I don't know exactly what I want yet. Interview me until we both do.
Ask one question at a time, and wait for my answer before the next one. Help me find the real goal behind the one I state, separate symptoms from the actual problem, surface constraints I haven't mentioned, and notice when I'm describing a solution rather than a need. If I seem stuck, offer two or three concrete options to react to instead of another open question.
Stop when the goal is concrete enough that we could tell whether it's been achieved. Then write it down in a few lines: the goal, why it matters, the constraints, and what done looks like. Ask me whether it feels right before we go further, and don't start building until I say so.

WANT-4 Ask what it understood

Default tool: play it back. Done when: Before real work, the agent played back your goal and priorities, and you corrected what it got wrong. How to check: Its final version made you think: hell yes, you get it. Course page: Play the design game. Boxes: Play it back.

Setup prompt, "Ask what it understood":

Before you do any real work on what I've asked, tell me what you understood.
In a few short lines, play back my goal, why it matters, who it's for, what I care about most, what I don't care about, and anything you're assuming that I didn't actually say. Mark your assumptions clearly, because those are where we're most likely to go wrong.
Keep it short and plain. If you're unsure about something important, ask me about that one thing rather than guessing.
I'll correct you. Update your understanding and play it back again, until I tell you that you've got it. If I say it feels like too much, say it again simpler. Only then start the work.

Stage 2: Borrow specificity

BORROW-1 Clone it, then remix it

Default tool: remixing. Done when: You had the agent make a faithful copy of something you love, then changed it, one step at a time, into something of your own. How to check: The result is more specific than anything you could have described from scratch. Course page: Open the exercise. Boxes: Remix.

Setup prompt, "Clone it, then remix it":

I want a specific result, and the fastest way there is to copy something that already exists and then make it mine. If I haven't said what I'm thinking of, ask me what I love that's close to what I want.
First, make a faithful clone of that thing: as close to the original as you can get, using whatever this setup has. Respect copyright and licenses: this is for my own use and learning, and say if something can't be reused.
Show me the clone and wait. Then help me remix it one change at a time: a new theme, a new audience, my own content, or a combination with something nobody has combined it with before. Show me the result after each change, keep the original so we can compare, and finish with a before and after.

BORROW-2 Research the experts' method

Default tool: meta-research. Done when: Your agent researched how the best people do a kind of work you care about, and you worked through their method on a real task. How to check: You can name the method, where it came from, and what it changed. Course page: Read the guide. Boxes: Expert research.

Setup prompt, "Research the experts' method":

I want to do a kind of work the way the best people in that field do it, not the average way. If I haven't said which kind of work, ask me.
Research who is genuinely excellent at it and how they work: their methods, standards, checklists, frameworks, and the mistakes beginners keep making. Prefer people who show their work and sources you can link to. Separate broad consensus from one person's opinion.
Show me two or three methods side by side, what each is optimized for, and which you'd recommend for my situation, then wait for me to choose. Then walk me through the chosen method on one real piece of my work, step by step, and tell me where my result falls short of the expert standard. Offer to save the method somewhere I keep notes, crediting the source.

BORROW-3 Learn the words experts use

Default tool: terms of art. Done when: You asked for the vocabulary of a field you're new to and used it in a prompt. How to check: Rerunning the prompt without those words gives a visibly more generic result. Course page: Open the exercise. Boxes: Terms of art.

Setup prompt, "Learn the words experts use":

I'm new to a field and I don't know the words experts use, so my requests come out vague. If I haven't said which field or what I'm trying to make, ask me.
Teach me the vocabulary that matters for what I want: the terms practitioners use to tell apart things a beginner would lump together (the way pixel art differs from 8-bit, or a hook from a chorus). For each term, give a one-line meaning and a famous example I can look at. Keep it to the ten or so that would most change a result.
Then help me rewrite one of my prompts using the right words, run both versions, and show me the results side by side so I can see what the vocabulary bought me. Offer to save the glossary where I keep notes.

BORROW-4 Brief with references

Default tool: screenshots · examples. Done when: You gave the agent concrete references (screenshots, examples, samples) instead of adjectives. How to check: Someone can see the references in the result. Course page: Open the exercise. Boxes: References.

Setup prompt, "Brief with references":

I want to direct you with references instead of adjectives, because "modern and clean" gets the generic look. If I haven't said what we're making, ask me.
Help me gather references: things whose look, tone, structure, or feel I want. I may give you screenshots, links, or files; if I don't have any, search for strong candidates (galleries like recent.design, siteinspire, or Mobbin are good for design) and show me a dozen to react to. Ask me what I like about each one I pick, specifically.
Then write a short brief from the references: what to take from each, and what to avoid. Show it to me, then make the thing from it. Finish by showing the result next to the references, and tell me which reference influenced which choice.

BORROW-5 Bring your own material

Default tool: notes · past work. Done when: The agent read your own notes, work, or collections and used them as source material. How to check: The result has something in it that only you could have supplied. Course page: Open the exercise. Boxes: Your material.

Setup prompt, "Bring your own material":

The best source material for what I'm making is probably my own: my notes, past work, things I've written or collected, my taste. Use it.
Ask me what we're making if I haven't said. Then look for relevant material of mine wherever I point you; I may keep a personal folder, notes, documents, or nothing organized yet, so check and ask rather than assume. If you can't reach it, ask me to paste or upload the best pieces. Read without changing anything.
Tell me what you found that's distinctive: recurring themes, phrases, preferences, stories, or images that only I would have. Suggest how to combine them with the best outside examples or expert methods, and wait for me to choose. Then make it, and point out which parts came from my material.

Stage 3: Write the brief

BRIEF-1 Write a seven-part brief

Default tool: the Claude Pop breakdown. Done when: You wrote a brief for something you care about that covers all seven parts: intent, taste, audience, tools, freedom, workflow, and verification. How to check: The first result needed far fewer fixes than a one-line request. Course page: Open the exercise. Boxes: Intent, Taste & references, Audience, Tools & resources, Freedom & constraints, Workflow, Verification.

Setup prompt, "Write a seven-part brief":

Help me write a great brief for something I care about, so you can do it well the first time. If I haven't told you what it is, ask.
A great brief usually covers seven things: the intent (what, why, and the stretch goal), taste and references (how it should feel, what to copy and avoid), the audience and the response we want, the tools and resources you may use, what's fixed versus your call, the order of work where order matters, and how you'll check and improve your own work before calling it done.
Interview me briefly, one question at a time, to fill the gaps that matter, and skip parts that genuinely don't apply. Show me the finished brief and wait for my go. Then do the work from it, and tell me which part of the brief made the biggest difference. Offer to save the brief as a template.

BRIEF-2 Decide what's fixed and what's free

Default tool: freedom & constraints. Done when: Your brief says which decisions are fixed, which are the agent's, and which are only suggestions. How to check: The agent surprised you, in a good way, on something you left open. Course page: Open the exercise. Boxes: Freedom & constraints.

Setup prompt, "Decide what's fixed and what's free":

Help me decide, for something we're about to make, which decisions are fixed, which are yours, and where I only have a suggestion. If I haven't said what we're making, ask me.
List the important decisions the work involves, in plain words. For each, ask me whether it's fixed (and what it must be), up to you, or a suggestion you can override with a good reason. Push back if I'm fixing things I don't really care about, since over-specifying gets me my own first idea back, and point out anything I've left open that could go badly.
Write the result as a short section I can add to a brief. Then do the work, and afterwards tell me where you used the freedom and what you chose.

BRIEF-3 Build in a self-check

Default tool: verification. Done when: Your brief tells the agent how to inspect and improve its own work before calling it done. How to check: It found and fixed a problem before you saw it. Course page: Open the exercise. Boxes: Verification, What comes back.

Setup prompt, "Build in a self-check":

Before you call this work done, I want you to check and improve it yourself, the way a demanding reviewer would. If I haven't said what we're working on, ask.
First tell me how you'll judge it: what good looks like for this kind of work, and how you can actually observe the result (open it, run it, screenshot it at different sizes, read it start to finish, listen to it). Wait for me to agree.
Then inspect the real result, not your memory of making it. List the worst problems you can actually point to (up to ten, ranked; don't invent any), check any factual claims against their sources, fix the top three, and inspect it again. Show me the before and after, with evidence, and tell me what's still weak. Offer to add the check to my brief template or my agent instructions so it happens every time.

Stage 4: Steer

STEER-1 Build a comparison gallery

Default tool: make many, then choose. Done when: The agent made many genuinely different versions on one gallery page, and you chose and pushed the next round. How to check: You can say what your taste turned out to be. Course page: Open the exercise. Boxes: Comparison gallery.

Setup prompt, "Build a comparison gallery":

Help me find the best version of something by making many and choosing, instead of accepting the first. If I haven't said what we're making, ask me.
Make a wide range of versions that differ in meaningful ways (style, tone, structure, approach), not small variations of one idea. Use whatever generation tools are available here for images, sound, or video; generate text and designs directly. Put them side by side on one page (an artifact, or a simple local page if you can make files here), with almost no text around them so the work speaks.
Ask me which I like and why, then make a new round that pushes further in that direction, and keep every round so I can compare. After two or three rounds, help me put my taste into words, and offer to save that description for next time.

STEER-2 Trust your gut

Default tool: hell yes or hell no. Done when: You noticed a bad feeling about a result, said so, and found what was missing. How to check: You fixed the cause (missing context, an unclear goal), not just the symptom. Course page: Open the exercise. Boxes: Gut check.

Setup prompt, "Trust your gut":

Something about working with you on this doesn't feel right, and I want to find out why instead of just pushing through.
Ask me how I'm feeling about the result or the process (frustrated, confused, bored, distrustful, or something else) and what triggered it. Then help me work out the most likely cause: context you didn't have, context that was stale or contradictory, a goal that was never clear, or a tool or permission you were missing.
Tell me the single most likely cause and how sure you are. Propose the smallest fix, which might be a clearer goal, a missing fact, or a fresh start, and wait for me to agree. Try again with the fix, and ask me whether it feels better. If it's a fix worth keeping, offer to write it down where future sessions will see it.

STEER-3 Say exactly what's wrong

Default tool: specific feedback. Done when: You gave feedback as specific, plain observations, one change at a time. How to check: Each round fixed what you named without breaking something else. Course page: Open the exercise. Boxes: Say what's wrong.

Setup prompt, "Say exactly what's wrong":

I want to get better at giving you feedback that actually moves the work forward.
Show me the current result. I'll tell you what I think, probably vaguely at first. When my feedback is vague ("make it better", "it feels off"), help me turn it into specific observations by asking what exactly I'm reacting to, or by showing me two small variations to compare. I don't need to know the technical cause; plain words are fine.
Then make one change at a time, and show me the result after each, so I can tell what each change did. Don't change things I didn't mention. If a change makes something else worse, say so. At the end, list the feedback that made the biggest difference, so I learn what good feedback sounds like.

STEER-4 Start fresh instead of arguing

Default tool: roll back · new thread. Done when: When the agent got stuck, you rolled back or started a new conversation with a better brief. How to check: The fresh start got there faster than another round of arguing would have. Course page: Open the exercise. Boxes: Start fresh.

Setup prompt, "Start fresh instead of arguing":

We seem to be going in circles, and I'd rather start fresh than keep arguing.
First, help me save what we've learned: what we were trying to do, what worked, what didn't, the approaches that failed and why, and any facts we established along the way. Find the last version of the work that was actually good, if there is one; check whether this project has saved versions or history before assuming it does, and don't delete anything.
Then write a better brief for a fresh start that includes those lessons, so a new conversation won't repeat the same mistakes. Show it to me. I'll open a new conversation with it, or roll back to the good version and continue from there, whichever you think is more likely to work. Tell me which and why.

STEER-5 Keep a prompt library

Default tool: prompt library · skills. Done when: Your best prompts are saved where you and your agents can find them, with a note on what made each one work. How to check: You reused one this week. Course page: Read the guide. Boxes: Keep what worked.

Setup prompt, "Keep a prompt library":

I want to keep the prompts that work, so I'm not starting from zero every time.
Look back through my recent conversations with you, if you can see them, and through anything I point you to or paste in, for prompts that produced results I liked. For each, note what it was for and what made it work: the reference, the precise wording, the constraint, or the check.
Help me decide where a prompt library should live. I may already keep a personal folder, a prompts file, a Claude project, or skills; check what this setup supports and what I already have before creating anything new, and ask me if it isn't obvious. Save the best prompts there with a short note each, and suggest which ones I repeat often enough to become a skill or a standing instruction. Show me what you saved and where.

Stage 5: Let it run

RUN-1 Run a goal loop

Default tool: /goal. Done when: A long job ran until a checkable finish line was true. How to check: It stopped because the condition was met, and you checked the evidence. Course page: Read the guide. Boxes: Goal loop.

Setup prompt, "Run a goal loop":

I want you to keep working on something until a clear finish line is true, instead of stopping when it seems done. If I haven't said what the work is, help me choose something worth a long run.
Help me write the finish line as one checkable condition: a state of the world, not a list of activities, plus the evidence that proves it (a test, a checklist, a file, an observable result). Add the constraints that must stay true along the way, and say how anything that can't be done should be reported. Find out what this agent offers for long runs (a goal mode, for example) and use the simplest option.
Show me the condition, the likely time and cost, and a limit, and wait for my go. When it ends, show me the evidence, not just a claim.

RUN-2 Run a gauntlet loop

Default tool: a high bar · hours. Done when: You gave an agent a goal, a demanding quality bar, and a long leash, and it worked for hours. How to check: You can say where the run got better, and where it wasted time. Course page: Open the exercise. Boxes: Gauntlet loop.

Setup prompt, "Run a gauntlet loop":

I want to try a gauntlet loop: a long, demanding run where you plan, build, critique your work harshly against a high bar, and improve it, for hours. If I haven't said what we're making, help me choose something where excellence matters and you can judge it.
Help me write the bar first: what excellent looks like, specific and observable, with references if we have them. Then design the gauntlet of checks each round must pass, and how a harsh critic (another agent, if this setup supports it) will find the biggest gap each time. Find out what this agent offers for long runs and use the simplest setup.
Tell me the likely time and cost and set a limit. Show me the plan before starting. Afterwards, show me the result, how it changed round by round, and where the run wasted effort, so the next one goes better.

RUN-3 Fan out

Default tool: many agents at once. Done when: Several agents worked on independent parts at once, and the results were brought together. How to check: The combined result beat what one agent produced alone. Course page: Read the guide. Boxes: Fan out.

Setup prompt, "Fan out":

I want to try splitting a big piece of work across several agents working at once. If I haven't said what the work is, look at what I'm working on and suggest something with genuinely independent parts.
Good candidates have parts that don't depend on each other: several sources to research, several styles to try, competing explanations to test, or a builder and a critic. If the work doesn't really split, tell me, and suggest doing it in one pass instead. Find out what this agent offers for parallel work (subagents, separate threads, or a workflow mode) and use the simplest option.
Show me how you'd split it, how the results will come back together, and the likely cost, and wait for my go. Afterwards, show me the combined result and whether splitting actually beat a single agent.

Other guides

This guide stands alone. Other parts of the course have their own guides, listed in the index; they cover separate topics.