← ClaudeAtlas

retrolisted

Monthly or end-of-milestone review of Sarah's Claude setup across all projects — repeated prompts, lessons stuck in one folder's memory, what the hooks blocked, the costliest helpers — ending in proposals for where each fix belongs. Use on "retro", "what am I repeating", "review my setup", "why is this expensive", or when ~/.claude/PLAN.md says a retro is due.
schen22/claude-config · ★ 0 · AI & Automation · score 70
Install: claude install-skill schen22/claude-config
# Retro Plan, then wait: this skill proposes. Nothing changes until Sarah agrees. ## Steps 1. `python3 ~/.claude/skills/retro/scan.py` (add `--days N` to change the window). Its **Sessions** section is the main evidence: one line per session (commits, tokens, helpers, denies, and the `/end` note: goal, what went wrong, lessons). Fewer than half the sessions with a note → `/end` isn't being run; raise it. 2. For token usage by project, session and skill, run the `session-report` skill's analyzer in JSON mode and read only `overall`, `by_project` and `by_subagent_type`. Build the HTML report only if Sarah asks for it. 3. Read `~/.claude/PLAN.md` (inventory, Now/Next, Decisions) so proposals don't repeat something already decided. ## Report (≤ 30 lines) | Finding | Evidence | Proposed fix | Rung | |---|---|---|---| Rungs, strongest first (from `~/.claude/CLAUDE.md`): 1 script/test/hook · 2 template · 3 skill step · 4 CLAUDE.md line. Pick the strongest that fits. - A repeated phrase across projects → a line in `~/.claude/CLAUDE.md` or a skill. - A local memory holding a workflow lesson → move it up, then delete the memory. - A hook's would-deny that was right → propose enforce. One that was wrong → fix the rule. - A costly helper → name the cause (items per helper, pointer brief, unfiltered output). ## After Sarah agrees Apply the changes, then update `~/.claude/PLAN.md`: inventory, status, a dated Decisions line and a Lessons line where something was l