| name | kaizen |
| description | Captures learning moments as short, permanent markdown notes in the project's kaizen/ directory, and registers them in AGENTS.md for progressive discovery.
Invoke this skill automatically โ without being asked โ whenever ANY of these signals appear: the user corrects a mistake you just made ("no", "wrong", "not like that", "undo", "you missed", "that's not right"); the user redirects you after you went down the wrong path; you realize mid-task you made a significant error and had to backtrack; the user has to re-explain something you already got wrong once. This skill is compulsory โ every learning must be captured. Do not skip it because the fix is already done. Do not wait to be asked. Invoke it as part of the same response where you acknowledge the correction.
|
Kaizen
Every mistake or correction is signal. This skill captures it immediately โ before the turn ends and the context is lost โ as a short, permanent note in kaizen/.
The philosophy (from Japanese manufacturing): small, continuous improvements compound. A 5-10 line note written now can save hours later.
When to invoke
Invoke when you detect any of these signals:
| Signal | Examples |
|---|
| Explicit correction | "no", "wrong", "not like that", "undo that", "that's incorrect" |
| Path redirect | User asks you to stop and restart differently after you've made progress |
| Self-correction | You catch an error mid-task and have to backtrack significantly |
| Implicit correction | User re-explains something you already got wrong; visible frustration |
Do not invoke for: requirement changes the user hadn't decided yet, minor typos, or one-off context-specific mistakes with no reusable lesson.
How to write the note
1. Name the learning precisely
One sentence. Specific, not generic.
- โ "Be careful with git commands"
- โ
"Always pass
--repo OWNER/REPO to gh issue create when not inside the target repo"
2. Create the file
Path: kaizen/YYYYMMDD-slug.md where slug is 2โ4 hyphenated words from the title.
# [Short title โ what was learned]
**Date**: YYYY-MM-DD | **Context**: [task / file / command where this happened]
## What happened
[1โ2 sentences: the concrete mistake]
## Why it was wrong
[1โ2 sentences: the underlying reason this matters]
## What to do instead
[2โ3 sentences or bullet points: the correct approach going forward]
Keep it to 10โ15 lines total. Scannable in 10 seconds is the goal.
3. Register in AGENTS.md
Add one line to the ## Kaizen Learnings section at the bottom of AGENTS.md.
If the section doesn't exist yet, create it:
## Kaizen Learnings
<!-- One line per entry: bold date + title, arrow, relative link -->
- **[YYYY-MM-DD] Short title** โ [kaizen/YYYYMMDD-slug.md](kaizen/YYYYMMDD-slug.md)
For subsequent learnings, append a new line inside the existing section โ never rewrite what's already there.
4. Then continue
After writing the note and updating AGENTS.md, proceed with fixing whatever went wrong. The note takes ~60 seconds to write.
What makes a good note
- Specific: future you knows exactly what to do differently
- Causal: explains why the mistake happened, not just what happened
- Actionable: states the correct approach clearly enough to follow without re-reading the original conversation
