learnlisted

# crew Learn The harness should get better with every project. After completed work, distil what is worth reusing — but propose, never auto-impose. ## What to look for - **A reusable procedure** → a new **skill** (e.g. "how we wire BullMQ workers"). Draft it in the skill format (frontmatter `name`/`description`/`origin: crew` + concise body). - **A capability that should pull in skills/rules** → a new or extended **tag** in the global registry (`project-types.json`). E.g. a `realtime` tag that activates a websocket skill. - **A decision worth remembering** → an entry in the project's `PROJECT.md` (the *why*), so it isn't relitigated. ## Rules - **High-signal only.** Promote things that will plausibly recur. Don't manufacture learnings to look productive. - **Propose with the artifact.** Show the actual draft skill / tag mapping / decision text; the user confirms each. - **Write on approval.** Skills/tags go to the global layer (`~/.claude/crew/`) so all projects benefit; decisions go to the project's `PROJECT.md`. - **Close the loop.** A learning that becomes a tag should, on the next `/crew:init` of a matching archetype, activate automatically. ## In the finish strand `/crew:learn` is the **Learn step** of the `/crew:finish` close-out strand (`config.workflow.learn.run`, additionally gated by `config.workflow.learn.enabled`). Cadence there is **per milestone** — one learn per finished milestone, never per phase. Run standalone any time, or let `/crew:finish` offer/run