Prompt Lab — Write Better, Get Better Results 
The difference between a useless AI response and a game-changing one? The prompt. Practice here — click each scenario to see real before/after differences, then hit Try Another for a fresh example.
Is Your Prompt Any Good?
Paste any prompt below. We'll scan it for the 8 things that separate amateur prompts from professional ones — and tell you exactly what's missing.
Before & After
The most common mistake: asking for something without saying what you actually want.
AI can't read your mind. The more relevant background you provide, the better the output.
Stop asking AI questions. Start giving it jobs with clear deliverables.
For complex tasks, tell AI to think step by step instead of jumping to the answer.
The biggest unlock: stop chatting and start delegating complete workflows.
Give AI a role and identity. It changes how it thinks, what it prioritizes, and how it communicates.
Build Your CLAUDE.md — Control Your Environment
Everything above makes one prompt better. A CLAUDE.md makes every prompt better. It's a plain text file that sits in your project folder, and the AI reads it at the start of every session — before you type a single word. Your context, your stack, your rules, loaded automatically. Most people have never thought about giving a project a brain — a memory that outlives the chat window. That's what this is. It's the difference between chatting with an assistant and directing one that already knows how you work.
Why I put a CLAUDE.md in every project I build. I run dozens of projects at once, and every single one has its own CLAUDE.md. It's the highest-leverage file I write — here's what it buys you:
- You stop repeating yourself. Say it once in the file instead of re-explaining your stack, your style, and your standards at the top of every session.
- You set the guardrails. “Never touch this folder.” “Always run the tests before saying done.” “Ask before deleting anything.” The AI inherits your boundaries instead of guessing them.
- You take control of the environment. You're not hoping the AI behaves — you're telling it exactly how work gets done here, and it follows those rules on every task.
- It compounds. Every lesson you learn becomes a new line in the file. Your project gets sharper over time because you're teaching it permanently, not one chat at a time.
One rule of thumb I follow: the CLAUDE.md holds the rules and conventions — the “how we work here.” Keep it lean and it stays powerful. (Where you are in the project — the running notes and state — goes in a separate notes file, so the rules never get buried.)
Fill in the four sections below, then drop the result into a file named CLAUDE.md at the root of your project. That's it — the AI does the rest, every time. And this is not just for coders — if you're building a dashboard or a webpage in the workshop, your project deserves a brain too.
What are you building, who is it for, and what stage is it in? The AI can't see the big picture unless you paint it. Developer example: "A personal finance web app for freelancers. Early MVP, solo developer." Workshop example: "A one-page dashboard for my landscaping business. I'm not a developer — explain everything in plain English."
How the project is built and the style rules the AI must follow every time. Developer example: "Python 3.11, Flask, SQLite. Type hints. No frontend frameworks." Not technical? Describe the shape instead: "One HTML file I open in my browser. My numbers live in a spreadsheet next to it. No installs, no accounts, plain English only."
The rules the AI is not allowed to break. This is where you take control. Developer example: "ALWAYS run the tests before saying done. NEVER commit secrets." Workshop example: "NEVER make up numbers — if data is missing, say so. ALWAYS show me the page in the browser before saying done. Ask before deleting anything."
Exactly how to run, check, and ship the project — so the AI never has to guess. Developer example: "Run: python app.py. Test: pytest. Deploy: git push." Workshop example: "To view: open dashboard.html in the browser. To update: I edit data.csv, then you refresh the charts."
Your CLAUDE.md
Fill in the 4 steps above to generate your CLAUDE.md...
CLAUDE.md in your project root. Every new session reads it first — so you never have to explain yourself twice.Prompt Cookbook
Ready-to-use prompts on two shelves: one for work, one for life. Pick a category, grab a recipe, paste it in — each one works out of the box. Stuck on what to build next in the workshop? This is also your idea menu: browse the life shelf and pick whatever sounds fun.
AI BS Detector
Twitter threads promising "10x your income with this one prompt." LinkedIn posts claiming a "secret formula" that makes AI do everything. We see this stuff daily. Paste it here — we'll tell you what's actually useful and what's pure BS.
How I learned to prompt — and run agents
Every note here is a real mistake I made building and running AI systems — turned into a rule so I don’t repeat it. Here’s the one that kicked off this whole page; the full shelf lives in the Library.
Front-load everything in your head
The #1 reason AI feels slow: you correct it one detail at a time instead of saying them all up front. Every fix below was a fact Peter already knew — the task sentence just didn’t carry it.
The task was one sentence: “Recreate my edit.” It took seven rebuilds — not because the work was hard, but because the editing recipe lived in his head and surfaced one correction at a time. Before you reveal each row, guess: what did the agent have no way of knowing?
Built: zoom, volume, and overlays laid over Peter’s continuous voice.
Built: cut voice → agent → resume at the next callout, dropping the takes in between.
Built: full voice-over kept, hold-and-resume.
Built: overlapped the callouts and trimmed Peter’s takes to hit length.
Built: full takes, cut only at callouts.
Built: full agent voice + buffer for every take.
Built: all of the above.
And three things that never made it into the sentence at all: the b-roll was a folder of clips in dogelord/backgrounds he hadn’t mentioned, the captions he’d redo himself, and the fixed crop couldn’t follow his face (it needed face-tracking). None were discoveries about the world — all were facts already in his head.
You don’t have to write it alone — build the spec with the agent
Here’s the honest part: nobody holds the whole spec in their head on demand. You don’t have to. That’s the heart of spec-driven development — you agree the spec before the build, and you draft it together. A good agent is unusually good at finding the holes in your thinking, so let it. Tell it to interview you first:
The discovery still happens — but as a two-minute conversation instead of seven rebuilds. A question is cheap: it’s text, answered in seconds, before a single frame is rendered. A wrong build is expensive: a full render, your review, and a context switch, every round. The same facts come out of your head either way — the agent asking just pulls them out before the work is wasted. Better still, the spec you agree on becomes a reusable artifact: drop it into the project’s CLAUDE.md or memory and you never brief it twice.
The manual version: the 90-second brain-dump
Whether the agent interviews you or you do it solo, make sure these get covered. Miss one and you’ll discover it by correction, the expensive way.
- Shape of the output. Length, format, structure, style — the target you’re actually aiming at. (Here: ~2:48, vertical.)
- The whole workflow. What happens at each step, especially the handoffs. (Here: your voice cuts, the agent takes over, your voice resumes.)
- Per-item exceptions. Which cases get treated differently, and how. (Here: 1/2/5 full, 3/4 trimmed & bled.)
- Where the inputs live. Files, folders, sources — don’t make the agent guess. (Here: b-roll in
dogelord/backgrounds.) - What “done” looks like. The acceptance test — how you’ll know it’s right. (Here: captions spelled right, face tracked, buffer intact.)
- The stuff that’s obvious to you. Say it anyway. The “obvious” details are exactly the ones that cost seven rounds.
The honest tradeoff. Memory amortizes; specificity prevents. An agent remembering “Peter trims 3 & 4” next time lowers the cost of the next task — but that memory only exists because the iteration cost got paid once. Specificity up front lowers the cost of this one. The brain-dump is the lever you control today. You saw the atomic version up top — Vague vs. Specific. This is that same lesson at full scale, from a real build.
Peter records the take — the script, research, agent voiceovers, captions, and the entire edit are produced by AI. That’s the spec-first workflow above, running in production.
More field notes
That was the interactive deep-dive. The rest of the shelf — 20 more real lessons from live builds and a shelved experiment, the receipts behind the Tips — lives in the Library. Browse all the Field Notes →