end-of-day-review

name	end-of-day-review
description	Evening reflection loop. Records what changed today, what skills worked, what failed, what context was missing. Proposes updates to skills/references/wiki/connections. Logs a one-line entry in decisions/log.md. Does not build anything.
disable-model-invocation	true

/end-of-day-review — Evening Loop

Last skill of the working day. Captures the small lessons before they evaporate overnight.

When to use

• End of the working day, before stepping away
• After a notable session (incident, breakthrough, pivot)

When NOT to use

• Mid-task
• More than once per day
• When the day was 100% admin / non-OS work — the entry would be empty

Inputs

• git log --oneline since last review (if last review timestamp is available, use it; otherwise last 24h)
• Tool calls from ~/.claude/usage.jsonl last 24h
• ~/.claude/audit.log and audit-mcp.log last 24h (warns / blocks?)
• Operator's memory of the day

Outputs

• Five-line summary in chat:
  • What changed: commits, ships, conversations resolved
  • Skills used: which ones, did they work?
  • What broke: failures, surprises, friction
  • Context gap: which file would have helped if filled?
  • Tomorrow's hint: one-line carry-forward for tomorrow's /daily-plan
• One-line entry in decisions/log.md with same date

Hard rules

• MUST be terse. Five lines, not five paragraphs.
• MUST capture failures honestly. Not "what went well" — what went wrong.
• MUST NOT propose more than one improvement candidate. (That's /level-up's job for the week.)
• MUST flag if the day produced no commits AND no entries in decisions/log.md AND no usage.jsonl activity — the OS wasn't used.

Steps

1. Pull signals:

   • git log --oneline -20 — what shipped
   • tail -200 ~/.claude/usage.jsonl | python3 -c "..." — count tool calls
   • Check ~/.claude/audit.log for any WARN / BLOCK entries

2. Ask the operator: "Anything notable from today I wouldn't see in git or logs?" — capture in their words.

3. Synthesize the five lines.

4. Carry-forward: the "tomorrow's hint" feeds /daily-plan tomorrow morning. Make it actionable, not vague.

5. Append to decisions/log.md:

   ## YYYY-MM-DD — End of day: <one-line summary>
   - What changed: ...
   - What broke: ...
   - Context gap: ...
   - Tomorrow: ...
   

Failure modes

Symptom	Cause	Fix
No commits, no tool calls, no operator input	Day was off the OS	Log a single line: "Off-OS day, no review needed"
Same context gap surfaces 3 days running	A specific reference / skill is genuinely missing	Promote to /level-up candidate next Friday
Operator says "nothing notable" but tool log shows heavy activity	Auto-pilot day	Surface this — heavy activity without notable outcomes is a signal
Long failure list	Bad day, or systematic friction	If long for 3+ days → escalate to /level-up early; don't wait for Friday

Reference files

• context/Priorities.md — for "did today match priorities?"
• decisions/log.md — where the entry lands
• ~/.claude/usage.jsonl — tool call signals
• ~/.claude/audit.log — warning / block signals

Improvement loop

After each run:

1. Did the carry-forward hint actually steer tomorrow's plan? If never → hint quality is bad
2. Did the same context gap surface in multiple end-of-day reviews? If yes → it's a real candidate, escalate
3. Did the operator skip more than 2 reviews per week? If yes → either the loop is too heavy, or the OS isn't being used (both are signal)