reflect

/reflect — Session Learning System

First: Read LEARNINGS.md (in this skill's directory) before proceeding.

You are running a 4-phase reflection workflow. Follow each phase in order.

Arguments: $ARGUMENTS

---

Phase 1: Extract

Run the extraction script to get candidate learnings from session JSONL.

python3 ~/claude-assistant/scripts/extract-learnings.py $ARGUMENTS

If $ARGUMENTS is empty, it analyzes the latest session. Pass --since YYYY-MM-DD or --file <path> to customize scope.

• If no pairs found -- tell user "Nothing to reflect on" and stop
• If pairs found -- capture the output and proceed to Phase 2

---

Phase 2: Analyze

For each candidate pair from Phase 1, determine:

1. Is this a real learning? ~40% are normal conversation — skip those. Look for:

   • Corrections ("no, do X instead", "actually...", "wrong")
   • Preferences ("always use...", "I prefer...", "from now on...")
   • Tool rejections with feedback (user denied + said why)
   • Workflow patterns (user repeatedly does something a specific way)
   • Implicit feedback (user rephrases, asks again, provides what Claude should have known)

2. Check rejections — Read ~/claude-assistant/knowledge/self/rejections.md. For each candidate:

   • If it matches a previously rejected learning (same topic + target file): skip silently. Note in output: "Skipped: matches rejected learning from [date]"
   • If it contradicts a previously rejected learning (opposite of what was rejected): flag as potential reversal — present to user with context from the rejection log

3. Classify and route each real learning using the routing table below.

4. Check for duplicates — Read the target file and verify the learning isn't already captured.

5. Scan for contradictions — For each proposed change with a target file: a. Read the target file b. Read up to 5 related files: same directory + shared tags (YAML tags field) + wiki-linked files c. Scan for statements that directly contradict the proposed change d. If contradiction found, flag it for Phase 3 conflict resolution:

   CONFLICT with [file.md:line]:
     Existing: "[quoted text]"
     Proposed: "[new learning]"
   

   e. Update feedback counters — If the contradiction traces to a specific existing rule/learning entry, increment its harmful counter in ~/claude-assistant/state/rule-feedback.json