The night the prompt became a production system
I wanted to make a music video. Instead, I built software for making music videos. This is how I know I work in technology.
I wanted to test metaprompting outside a programming demo. My early process was twenty tabs, scattered prompts, characters who kept changing, forgotten settings and a folder structure based on optimism.
So I handed the AI the busywork. It wrote the prompts, generated images, called the video tools, pulled sample frames from each clip and judged whether a shot should move forward. It even helped write the song. That night, a prompt became a production system.
Then the bill came. It was aggressive. A loop that can make many attempts is also a loop that can quietly set money on fire.
The lesson was worth it, and it isn't about video. The same create, review, revise loop works on an email, a flyer or a lesson plan, and costs almost nothing when you're making text. (For a later music video where I coordinated a whole stack of AI tools, read how I made a Phantasy Star music video.)
Metaprompting: ask for the instructions, not just the answer
A normal prompt asks for an output: "Write an email inviting parents to the school bake sale." A metaprompt asks the AI to write or improve the instructions that produce the output: "Before you write anything, write the best possible prompt for this email. Ask me what you need to know first."
It will come back with questions you forgot to answer: who's reading, how formal, how long, and must it include the date, time and what to bring? You answer, it writes clear instructions, and those produce a much better email than your one-liner.
You also keep something reusable. Next year's bake sale email is a paste and a date change.
Give the reviewer a checklist, not a vibe
Ask an AI "is this good?" and you'll mostly get flattery. Ask it to check the work against a short written checklist (a rubric) and you get a real review. It's the same kind of yes/no checklist as in AI does the work faster. Judging it is still your job. The new part is handing the ticking to the AI.
For the bake sale email, the rubric might be five lines:
- The subject line says what and when.
- Date, time and place are in the first two sentences.
- It says exactly what people should bring.
- It is under 150 words.
- The tone is friendly, not guilt-tripping.
Write it before you see a draft, so it measures what you wanted, not what you happened to get.
Then ask for a decision, not an opinion: PASS (good to go), REVISE (fix something specific) or REJECT (wrong approach, start over). Require evidence: the reviewer has to quote the line that fails. Do the review in a fresh chat so the AI isn't grading its own homework with a proud parent's eyes.
A sample review (invented)
- FAIL Subject says what and when: "A sweet opportunity!" has no event or date.
- PASS Date, time and place up front: "Friday, 3 to 5 pm, in the school hall."
- PASS Says what to bring: "one tray of cookies or cupcakes, labeled with ingredients."
- PASS Under 150 words: 112 words.
- PASS Friendly, not guilt-tripping: "We'd love to see you there."
Decision: REVISE. Biggest mismatch: the subject line. One prompt change: add "The subject line must name the event and the date." Keep everything else the same.
My video checklist covered story clarity, whether a character still looked like the same character, composition, realistic motion, safety and cost. The reviewer didn't judge beauty. It caught obvious problems, like the wrong character or garbled text, before I spent time looking.
Fix one thing at a time
This is the one-change rule from deliberate practice, with the AI doing the spotting. When the review says REVISE, ask for the single biggest mismatch and change only the part of the prompt that causes it. Everything that already works stays fixed.
Say you're making a yard-sale flyer and the reviewer reports that the headline is too small to read from across the street. Change the headline instruction and leave the colors, layout and wording alone. If the next draft gets worse, you know exactly what to undo.
Set limits, and keep the final call
Loops love to keep going. Decide the limits before you start: a maximum number of attempts (one draft and two revisions is plenty) and, for any tool that charges per use, a maximum spend.
Make cheap drafts before expensive ones. Next time I'd test with low-resolution still images and only turn approved stills into video. The everyday version: an outline before the full lesson plan, a written description of the flyer before generating images.
When you hit the limit, the loop stops and a human chooses. The AI can make, check and rewrite; you own the meaning, the surprise and the final pick. Automate the loop, not the taste, and never let the reviewer publish, send or spend on its own. (Knowing why a draft is weak is your own skill to build; that's in AI does the work faster. Judging it is still your job.)
Try it
Try it in 15 minutes
All you need is any chat assistant (ChatGPT, Claude, Gemini, Copilot) and a small made-up writing task, like an email inviting neighbors to a block party.
- Write a two-line brief
Say what you're making, who it's for and what it must do. Example: invite 20 neighbors to a Saturday block party and ask each household to bring one dish.
- Write your checklist before any draft
Before the AI writes anything, list five yes/no checks a good version must pass, plus one deal-breaker. For the block party: the date, time and place are in the first two lines; it asks each household to bring one dish; it says how to reply; it is under 150 words; it sounds friendly, not pushy. Deal-breaker: the date is missing. Stuck? Ask the AI what a good one usually needs and keep the five you care about.
- Ask for the prompt, not the email
Paste the Prompt architect prompt below with your brief and checklist. Answer its questions, then save the prompt it writes in a note.
- Make the first draft
Open a fresh chat, paste the saved prompt and get a draft.
- Get a review with a decision
In another fresh chat, paste the Quality reviewer prompt with your checklist, the prompt you used and the draft. You'll get PASS, REVISE or REJECT, with quoted evidence and the one prompt change to make.
- Revise one line, then decide
Change only the prompt line the reviewer named, run the new prompt in a new chat, then review the new draft the same way. Stop after two revisions at most. Then you decide whether it's ready.
You’ll know it worked when
You have a draft you'd actually send, a reusable prompt and checklist, and you can say exactly what each revision changed and why.
When you’re ready
Going further: let agents run the loop
Once the loop works by hand, you can give it to agents: AI programs that take steps on their own. I connected Sora (video), ElevenLabs (AI voice, for a changed voice) and the OpenAI Agents SDK (a toolkit for building agents) so the loop could write the prompt, render, inspect, describe what was off and re-prompt. Add guardrails:
- Track what state each job is in: waiting for approval, rendering, in review, accepted, rejected or stopped.
- Save every prompt version with a one-line note on why it changed. Without that history, an excellent result becomes a magic trick you can't repeat.
- Never resend a paid job after a timeout until you've checked whether it already ran. Otherwise you may pay twice.
- Get permission for any changed or cloned voice, and disclose synthetic audio.
- Require a human yes before higher-cost output or publishing.
Avoid these
Common mistakes
- Asking "is this good?"
You'll get compliments, not a review. Use a checklist and demand quoted evidence.
- Changing everything at once
Rewrite tone, length and structure at once and you can't tell what helped.
- No stop rule
With paid tools, a loop without a limit is a bill without a limit. Set attempts and spending caps first.
- Letting the loop make the final call
The reviewer catches obvious problems. It shouldn't decide what's meaningful, or send, publish or pay without your yes.
Copy, adapt, run
Prompts to try
Paste one into any chat assistant and replace anything in [brackets].
Prompt architect
You are a prompt architect. I want to create [what you're making, e.g. an email inviting neighbors to a block party] for [who it's for]. Before writing it, ask me up to five questions about anything you need to know. Here is my checklist for a good result: [your five checks]. Make sure the prompt covers every item. Then write a standalone prompt I can paste into a fresh chat to produce it. Include audience, goal, tone, length, format, must-include facts and things to avoid. Keep every fact I gave you exactly, and label anything you assumed.
Quality reviewer
Review the [draft type] below against this checklist: [item 1], [item 2], [item 3], [item 4], [item 5]. For each item, say pass or fail and quote the exact part of the draft that shows it. Then give one overall decision: PASS, REVISE, or REJECT. If REVISE, name the single biggest mismatch. Quote the line of my prompt that caused it, or say which instruction is missing. Give the exact replacement or added line, and say what must stay the same. Don't flatter me; I want the problems. Do not PASS if [a deal-breaker, e.g. the date is missing or wrong]. Prompt I used: [paste the prompt] Draft: [paste draft]
Cost guard
Before we start, here are the limits for this task: one draft and no more than [2] revisions, and a maximum spend of [amount, if the tool charges per use]. Make cheap test versions first ([e.g. an outline before the full lesson plan, or a still image before any video]). Before each attempt, tell me the attempt number and any cost. If we hit a limit, or you're unsure what something will cost, stop and ask me to choose. Never send, publish or pay for a higher-cost version without my yes.
The tool kit
Tools and links
- Any chat assistant (ChatGPT, Claude, Gemini, Copilot)Enough to start. The whole loop works in a normal chat window.
- Claude CodeA coding agent; I used it to run the whole workflow.
- OpenAI image generationCheap still-image drafts before paying for video.
- SunoAI music; helped write the song.
- SoraAI video generation.
- ElevenLabsAI voice; get permission for any changed voice.
- OpenAI Agents SDKFor building agents that run the loop for you.
- PythonFor scripting the automated version.
- FFmpegPulls sample frames from video for review.
The short version
What to remember
- Ask the AI to write the instructions first, then use those instructions to make the work.
- Review against a short checklist and demand PASS, REVISE or REJECT with evidence.
- Fix the biggest mismatch, one change at a time.
- Set limits on attempts and spending, draft cheap first, and make the final call yourself.
patrickz
