SLUG=$(echo "<target>" | tr '[:upper:]' '[:lower:]' | tr ' ./' '---')
DIR=/praxis/thinking/kaizen
mkdir -p "$DIR"
FILE="$DIR/$SLUG.jsonl"
NOTE_JSON=$(printf '%s' "<note>" | python3 -c 'import json,sys;print(json.dumps(sys.stdin.read()))')
printf '{"ts":"%s","cwd":"%s","note":%s}\n' "$(date -Iseconds)" "$(pwd)" "$NOTE_JSON" >> "$FILE"
wc -l < "$FILE"

Then respond: 🏷 kaizen:<target> logged (total: N) — thanks and resume prior conversation immediately. No commentary, no analysis, no reformulation.

Use python3 json.dumps for escaping — never hand-roll JSON.

Reserved targets

Target	Meaning
`praxis`	Top-level meta (workflow, methodology, portfolio)
`claude.md`	Instruction drift in any CLAUDE.md. All levels collapse to `claude-md.jsonl`; disambiguate via `cwd` field at review time
`self`	Note to future-self about personal workflow habits (distinct from `introspect/self.md` identity profile)
`<skill-name>`	Friction with a specific skill — use real name, not `lazy-*` prefix

Free-form otherwise; conventions unenforced.

Examples

/kaizen troubleshoot "asks same diagnostic twice in a row"
/kaizen claude.md "git commit rules ambiguous on scope field"
/kaizen convert-pdf "OCR re-runs on cached pages"
/kaizen convert-pdf "VLM mode is great for scanned docs"
/kaizen self "forgot to check benchmarks/inventory.yaml before starting"

Praise and friction share the bucket — natural language carries the sign. No type field, no --priority, no --undo, no auto-dedupe (repeats = valid intensity signal). Manual jsonl edit if needed.

Key rules

Note required. Missing note = useless for clustering.
After-only. Works after a skill returns, not mid-skill.
Verbatim storage. No cleanup, categorization, or reformulation.
One-line response. Never comment on note content. Resume immediately.

kaizen

Kaizen — Friction Capture

Usage

Behavior