Retrospective

Review session transcripts, list mistakes, propose improvements.

1. Resolve scope

Parse $ARGUMENTS:

• empty → current session — analyze from your existing conversation context. Do not read any jsonl file; the conversation is already loaded.
• contains a number + d/day(s) → past sessions modified within that many days. Read jsonl files from ~/.claude/projects/<encoded-cwd>/:

  SESSION_DIR=~/.claude/projects/$(pwd | sed 's|/|-|g')
  find "$SESSION_DIR" -name '*.jsonl' -mtime -N
  

• remaining text → extra focus/context to bias the analysis (e.g. "focus on test failures")

2. Analyze

Current session: scan your own conversation context. Extract user corrections, retried tool calls, reverts, "no"/"don't"/"stop", failed builds/tests, repeated identical commands, dead-end paths.

Past sessions (N days): spawn one Explore subagent per jsonl in parallel (single message, multiple Agent calls). Brief each:

Read <path>. List only the mistakes Claude made: user corrections, wrong assumptions, repeated failures, wasted tool calls, friction points. One bullet each, ≤15 words. No fixes, no praise. Under 200 words.

Then reconcile: dedupe, group by theme, keep frequency counts.

3. Report

Output to the user in this shape (no preamble):

## Mistakes
- <mistake> (×N)
- ...

## Improvements
**CLAUDE.md** (general, always-on):
- <rule> — <one-line why>

**Skill: <name>** (specific workflow):
- Trigger: <when to invoke>
- Body: <2-3 bullet behavior>

## Apply?

Routing rule:

• CLAUDE.md = always-relevant rules (style, conventions, repo invariants).
• Skill = workflow that fires on a specific trigger (slash command, task type).

Wait for confirmation before editing CLAUDE.md or creating skill files.

4. Apply (on confirmation)

• CLAUDE.md edits: append/modify in place, keep terse.
• New skill: write .claude/skills/<name>/SKILL.md with frontmatter (name, description, optional argument-hint).

Stop. Do not auto-commit.