session-reflect

name	session-reflect
description	End-of-session maintenance — audits recent work (git log, memory-file mtimes, CLAUDE.md diff) for rules earned, contradicted, or now surplus, then proposes promote / remember / trim / skip decisions for CLAUDE.md. Output-only — the user confirms each and the main agent applies. Trigger when wrapping up a session ("done for today", "/exit") or after a notable stretch of changes; not for start-of-session triage — use /whats-next for that.
tools	Bash, Read, Grep

Skill — session-reflect

A short end-of-session check: did anything emerge worth promoting, contradicting, or trimming?

Output-only. The skill proposes; the user confirms each edit, and the main agent applies it. The skill never auto-writes to CLAUDE.md (memory entries follow the existing auto-memory write path).

---

When to use

• End of session — user signals good night, done for today, next task tomorrow, /exit, or otherwise wraps the day.
• After a notable stretch — reviewer findings that produced new conventions, repeated user corrections, an emergent pattern across multiple PRs.
• Periodic — once a week if not naturally triggered, or whenever CLAUDE.md crosses its length budget (~250 lines).

Bypass the skill entirely when the session was routine and no rule shifts are visible in the artifacts (Step 1). Just say "No CLAUDE.md changes — routine session." and stop.

---

Required tools

Bash, Read, Grep.

---

Compact-resilience

This skill is designed to survive an in-session /compact. After a compact, Claude loses verbatim user quotes and the texture of earlier reviewer findings, but persistent artifacts remain. Signal priority order:

1. git log --since="1 day ago" and git log main..HEAD — commits authored this session (never lost).
2. ls -lt {memory_dir}/feedback_*.md — memory entries added/updated (file mtimes reveal session activity). Skip if the project has no auto-memory directory.
3. git diff main..HEAD CLAUDE.md — rules already added/changed this session.
4. docs/techdebt.md entries dated today — tech-debt filed (if the project tracks tech debt there).
5. Conversation context (verbatim quotes, reviewer texture) — supplement only, not required.

Throughout, main stands for your repo's default branch — substitute if it differs (master, trunk, …).

If a candidate is only visible in conversation context (signal 5) and not in any persistent artifact, treat it as low-confidence — likely belongs in Remember, not Promote.

---

Execution

Step 1 — Gather artifact signals

Run in parallel (substitute for your environment: {memory_dir} → your auto-memory directory from context, typically ~/.claude/projects/<project-slug>/memory/; main → your repo's default branch; {today} → today's date as YYYY-MM-DD. Skip the memory signal if the project has no auto-memory):

wc -l CLAUDE.md

git log --oneline --since="1 day ago"

git diff --stat main..HEAD CLAUDE.md

ls -lt {memory_dir}/ 2>/dev/null | head -10

grep -n "^## {today}" docs/techdebt.md 2>/dev/null

From the conversation (supplement only): user preferences expressed verbatim ("I will NEVER…", "from now on…", "always…", "stop doing…"), reviewer findings applied vs rejected, repeated decisions across multiple tasks.

Step 2 — Bucket each candidate

For every signal, decide:

• Promote (→ CLAUDE.md) — high bar; rule meets promotion criteria:
  1. Appeared in this session AND at least one prior session (git log --grep="<keyword>" for related commits), AND
  2. Project-wide generality (binds every future session, not just one feature's behavior).
• Remember (→ memory/feedback_*.md) — low bar; signal worth keeping but premature for CLAUDE.md: single-session preference, behavior-only, or not yet re-applied across sessions.
• Trim (→ CLAUDE.md) — existing rule was contradicted this session, redundant with another rule, replaced by a hook/lint/agent, or unexercised this session despite clear opportunities. Evidence required.
• Skip — routine signal, no action. Default outcome.

If your project sets a CLAUDE.md length budget (the kit targets ~250 lines) and CLAUDE.md is above it, additionally include a per-section line count (grep -n '^## ' CLAUDE.md) and flag the largest sections as trim candidates even without specific contradiction signals.

Step 3 — Emit proposals

Emit the result per ## Output format below: the routine-session one-liner when every signal buckets to Skip, otherwise the proposal table closed with Confirm Promote/Trim entries to apply.

Step 4 — Hand off for application

This skill is output-only: it does not edit CLAUDE.md or write memory itself. After the user accepts or rejects each row, the main agent applies the confirmed entries:

• Promote → main agent edits CLAUDE.md (locate the appropriate section; add as a new bullet or sub-bullet). Use the established section structure; don't invent new top-level sections without flagging it.
• Trim → main agent edits CLAUDE.md (delete the rule, or consolidate with another).
• Remember → main agent uses the auto-memory write path ({memory_dir}/feedback_*.md + MEMORY.md index update) — same mechanism the auto-memory system uses every session. If the project has no auto-memory directory, Remember degrades to propose-only — surface the rule for the user to record manually.

---

Output format

When every signal buckets to Skip:

No CLAUDE.md changes — routine session.

Otherwise, a single proposal table — keep each cell terse (an evidence cell is a reference, not a sentence; ≤40 chars):

#	Bucket	Item	Evidence
1	Promote	One-line rule statement	commit sha + prior-session ref
2	Trim	Rule to remove or rewrite	contradiction/redundancy cited
3	Remember	Rule going to memory only	why premature for CLAUDE.md

Close with: Confirm Promote/Trim entries to apply.

---

Critical rules

1. Output-only — propose, don't auto-edit CLAUDE.md; the main agent applies confirmed rows (Step 4). The user confirms each Promote/Trim.
2. Specific evidence required — every proposal cites something concrete from this session (a commit, a memory entry, a reviewer finding, a verbatim user quote). No theoretical "we should add this".
3. Brief output — one line when nothing applies. Short table otherwise. No prose preamble.
4. Honor promotion criteria — Promote proposals require evidence in ≥2 sessions AND project-wide generality. Single-session signals go to Remember.
5. Compact-resilient signals only for Promote — if the only evidence is conversation context (no commit, no memory entry, no CLAUDE.md diff), downgrade Promote proposals to Remember. A rule promoted to CLAUDE.md must be visible in persistent artifacts so a future audit can verify its origin.
6. Length budget as proactive trim signal — when CLAUDE.md exceeds the project's length budget (the kit targets ~250 lines), surface a per-section line distribution and propose trim candidates even without contradiction evidence.

---

Notes

This skill complements /whats-next (start-of-session: triage pending work). Together they bracket a day: pick task → execute → reflect.

The four memory categories (user / feedback / project / reference) are unchanged — Remember outputs typically land as feedback entries. The auto-memory system's normal write criteria apply.