extract-approachlisted

# Extract Approach — the reasoning recorder Every hard solve contains two products: the fix, and the *approach that found the fix*. The fix ships; the approach evaporates when the session ends. This skill saves the approach — the more valuable half — as a permanent note any future model reads before touching similar work. **Why this exists:** frontier models get repriced and retired. An answer helps once; a written-down approach upgrades every answer after it, at any model tier. The recipe requires top-tier judgment to *author* but only ordinary intelligence to *follow*. ## Setup (one-time) Pick a learnings directory in your knowledge base (e.g. `{KNOWLEDGE_BASE}/learnings/`) and keep a `## Learnings` index — one line per lesson — in a file your sessions always load (CLAUDE.md is the natural place). Everything below assumes those two locations exist; adjust paths to your setup. ## Trigger test (all three must hold) 1. **It was actually hard.** At least one of: multiple failed attempts before the solve; a non-obvious insight was the turning point; a judgment call between 2+ viable approaches; a quality bar or heuristic was invented on the spot. 2. **It will recur.** The problem *shape* (not the exact problem) will plausibly show up again in the user's work. 3. **The approach is transferable.** You can state a recipe someone else could follow without this session's context. If any fail, don't write a note — noise kills the learnings system faster than gaps do. Routine bu