Andon

Andon

Stop fixing the same AI mistake twice.

A Lean *quality system* for AI-assisted work — operators turn recurring defects into countermeasures. Built by a manufacturing operator, for people who actually ship.
*Andon* is the cord on a Toyota line you pull to stop production when something's wrong. This is that cord for AI work: catch the defect, investigate it, and add a countermeasure.

*Not affiliated with Andon Labs (the AI-agent evaluation company). Same Lean word, different project — this is a solo, open-source operating kit.*

The problem

If you work inside an AI agent every day, you know the failure modes:

- It forgets what you decided last week.
- It repeats mistakes you already corrected.
- It says "done" when it isn't.
- Projects drift — two sessions stomp the same file, context evaporates, you re-explain the same thing.

Most "AI tips" make the *output* a little better. Recurring defects also need explicit checks and follow-through.

The fix isn't a better prompt. It's an operating discipline — borrowed from a factory floor. Use the same discipline in AI work: record a defect, investigate its cause, and add a countermeasure. Where practical, add a mechanical check and test both the failure and a legitimate passing case. Checks have limits; keep those visible.

The aim: retain what you learned and check whether the countermeasure works.

Quickstart — your first win in 5 minutes (no code)

The smallest useful piece is a structured memory file — so you can come back to a project after three months and your agent knows exactly where you left off, instead of rebuilding context from scratch.

0. Clone this repo first (you'll copy a couple of files out of it):

1. In your project, make a memory folder and an index:

2. Add 2–3 lines to memory/MEMORY.md — the…