Requirement Ledger — AI skill for Claude Code

requirement-ledger

**From vibe coding to knowing what you actually wanted.** Point it at a project you already built with an AI agent and it reads the conversation back: what you really needed, what went wrong, and which repeated work is worth turning into a Skill.

[中文说明](README.zh-CN.md)

Who this is for

You built something by talking to an agent. It mostly works. You could not have written a spec for it before you started, and you still could not write one now.

That is not a discipline problem. Nobody can specify a thing they have not seen yet. But it has a real cost: the agent filled every gap you left silently, some of those guesses were wrong, you spent a dozen rounds correcting them, and none of it was written down. Next project, same gaps, same dozen rounds.

If you are a product manager with a signed-off requirements document, you do not need this.

You do not have to know what you want

Skip the spec. Build the thing badly, then let this read your own words back to you.

Here is the trick it is built on. Every time you told the agent it got something wrong, you described what you actually wanted — precisely, in your own vocabulary, without meaning to:

「不是让你重写 我就想让它别每次都问我一遍」

That one sentence contains the real requirement, the wrong guess the agent made, and how you would test it. The message you *opened* that task with was 「优化一下这个流程」.

So the requirement is not extracted from what you asked for. It is extracted from **where you had to correct it** — which is the one place in the transcript where you were specific.

How it works

Three things you can ask for, in plain words: 总结我的真需求 · 总结一下错误 · 哪些能自动化.

**Step 0 always runs a script.** Ask an agent to "summarise the project" and you get numbers guessed by eye, the opening message treated as the requirement, and a summary of the last two rounds passed off as a summary of the whole thing. So the counting is mechanical: `scan_transcript.py` reads the raw session files and reports real user tur