self-learn
Self-learn
A lesson learned twice and written nowhere will be learned a third time. This
skill is the shared loop over the repo-declared memory store (project memory,
session checkpoint, recall, memory-hygiene): where the active repository
declares its own learning spec or store, that declaration is authoritative for
categories, qualification, and file paths, and this skill defers to it. The
loop is engine-agnostic — it must work identically for any capable agent
(Claude Code, Codex, Copilot, custom harnesses). Harness-provided artifacts
(session checkpoints, notes files, trajectory stores) are optional inputs,
resolved at runtime: use them when the harness exposes them, fall back to
portable sources (git log, session notes, the repo tracker) when it does
not. Never hardcode another harness's paths, schemas, or store names.
The loop has three modes. Run only the mode the situation calls for.