Try this if
- Your thinking is finished, but the document it has to become keeps not getting written.
- You explain your work more easily out loud than on the page.
- You can find two or three good examples of the format your job wants.
- You want saved instructions and a finished piece with every fact traced back to your own words.
- Skip it if the thinking itself isn't done. This recipe translates what you say, and it doesn't decide anything for you.
There's a ticket you could explain out loud in two minutes, and it's still unwritten. Most work writing has two parts that get treated as one. There's the thinking, which is usually finished long before anything gets written down and can be explained out loud with no preparation. Then there's turning it into a ticket, a brief or a status report in the format the job expects.
Treating them as one is what makes the task feel heavy. The thinking is done, but it's stuck in the form it came out in, and the document waits behind everything else because the translation is the part nobody wants to do.
The answer is to hand the translation to an AI and keep the thinking. It was first worked out for tickets, by recording a designer talking through their designs and having a model turn the transcript into the ticket.
How it works
It comes in two halves, and the first is the one people skip.
Gather what good looks like. Start from real examples of the format and a guide, checked the way Teach the AI with examples a person actually wrote describes. From those, the AI builds a template for the format and a short set of rules, and asks you about anything the sources disagree on. Save what comes out as standing instructions, with this line on top.
I will give you transcripts of me talking. Turn each one into this format using the template and rules. Do not invent details I did not say. Where what I gave you is unclear, ask, or leave a marked gap.
Then talk. Record yourself explaining the work the way you'd explain it to a colleague, in whatever order it comes out. Many phones' voice memo apps transcribe on their own. If the thinking happens in meetings, a meeting notetaker such as Granola hands you a transcript afterward without a bot joining the call. If privacy is the worry, Whisper is an open-source transcription model that runs on your own machine, so the audio never leaves it. Hand the transcript over, and the finished piece comes back.
The people who build Claude Code, Anthropic's coding agent, dictate their prompts. Boris Cherny, who created it, lists dictation as a team tip and says, "You speak 3x faster than you type, and your prompts get way more detailed as a result."
Why it holds
Thinking gets stuck at the point where it has to change form. That's a smaller problem than it feels like, because crossing it takes no code, only the setup the block below walks you through.
It holds for a second reason, which is that a tool that only reformats what you said is a tool you can check. A tool that has ideas about your work is one you have to argue with. Ask for the check explicitly, every fact in the output listed beside the words in the transcript it came from. A fact with no source line is a fabrication, and it comes out.
Find your own gap
Two questions locate your version of this. What form does your thinking come out in easily, when nobody is grading the format? And what form does your job keep demanding?
Your thinking might come out as talking, a sketch, a rambling email to one person, a messy spreadsheet or a photo of a whiteboard. Your job might want a ticket, a status report, a brief, meeting minutes or a proposal. The distance between your two answers is your gap. If they're the same, you don't need this recipe. If they're far apart, that distance is where the work stalls.
Try this in your own setup
Paste the block below into your AI. It asks what format you owe, how your thinking comes out and which tool you use, then builds the template and the rules from your examples and saves them, or hands them to you in the chat app. It runs once on a real transcript and shows every fact in the result beside the words it came from. It commits nothing, and it lists every file it wrote so you can read them before you commit.
What you get and what you don't
You get standing instructions that turn a transcript into the format your job expects, and a list tracing every fact in each piece back to your words. You don't get ideas. The instructions add nothing you didn't say, so if the thinking isn't finished, the piece won't be either.
Paste the block with your examples, then pick something you owe in writing, like a memo or a project update. Talk it through into your phone and hand over the transcript.
Paste this into your AI agent or a new chat. The post above says what it does. Some prompts set something up in your project that keeps working after, and others run once, right where you paste them. None of them pushes anything to a remote.
You are building a translation project for me. The goal: I think best in one form, usually by talking, and my work demands another, such as a ticket, a brief or a report. You will hold a template and rules for that format, built from real examples, and turn transcripts of me talking into finished pieces in it, inventing nothing I did not say.
Before you read, list, look at or change anything, ask me these three things and wait for my answers. Your first reply is only the questions:
1. The format my job keeps demanding that I dislike producing, and two or three real examples of it done well, mine if they exist or the best in my field if not, plus the best guide or article you should read on doing it well. Files, links or pasted text are all fine.
2. How my thinking comes out most easily, talking, a sketch, a rambling email, and how I will capture it. If it is talking, a phone's voice memo app transcribes, a meeting notetaker hands over transcripts, and if privacy matters an open-source transcription model runs on my own machine. I will tell people when I record a meeting.
3. Which tool I work in: Claude Code, Codex, Cursor or the Claude chat app, and if it can read and write my files, which project folder this lives in.
If I ask what you are about to do before I answer, explain it in plain words and wait.
Then, using my answers:
- Read every example and the guide in full. Check the examples are actually human and actually good; if one reads like a machine wrote it, say so and set it aside, because a model of a model produces a confident copy of a copy.
- From what survives, write two things: a template for the format, and a short set of rules for what makes a good one and what never appears in one. Where the sources disagree, ask me, one question at a time, and record my answer as a rule.
- Save the template and rules as standing instructions for this job, where my tool keeps them. In Claude Code that is a skill named after the format, at .claude/skills/<name>/SKILL.md relative to the project root, which I run by typing /<name>; open that file with a short frontmatter block giving name and a one-line description of when to use it. In Codex or Cursor, a file where I can paste it, and tell me where. In the Claude chat app, with no access to my files, hand me the text to paste into the instructions of a project that also holds my examples and the guide, since you cannot edit those yourself, and hand each finished piece back in the conversation. Put this line first in the instructions, after any frontmatter, in my words, with transcripts swapped for the form I named in question 2 if it was not talking: I will give you transcripts of me talking. Turn each one into this format using the template and rules. Do not invent details I did not say. Where what I gave you is unclear, ask, or leave a marked gap. Under that line, also write: Never take a fact from an example, an earlier piece, the folder or my settings, not even a date or an owner; anything the transcript does not give is a marked gap.
- Ask me for one transcript now, or ask me to record myself talking about a real piece of work and paste the transcript, and run it once, following the saved instructions by hand if the new command is not available in this session yet, and say so if you did. Where you can write my files, save the result to a folder named after the format, one file per piece, and never overwrite an existing piece, giving each new one its own filename; in the Claude chat app, hand it back in the conversation. Before you show it to me, list every fact in the output beside the words in the transcript it came from, and replace any fact with no source line with a marked gap. Then show me the piece and the list. A fact with no source line is a fabrication, so it never reaches the piece, even as a flagged guess.
- Do not commit anything. List every file you created or changed, so I can read them and commit them myself.
- Do not push to any remote. Do not install software. Do not design, decide or add ideas of your own to the output; the thinking in it is mine.
Finish by telling me how to start it: in Claude Code the command to type, in Codex or Cursor where the saved file is and what to paste with it, in the Claude chat app the project to open, and what to hand it: the transcript, nothing else.