retrolisted

Focus: $ARGUMENTS The template is meant to be a compounding asset. This command is how it compounds: turn what was learned into configuration, so it is not re-learned next time. ## What to look for Go back over this session and find: 1. **Repeated corrections.** Anything the human had to say more than once — a convention, a preference, a constraint, a "no, not like that". Every repetition is a missing line of configuration. 2. **Wrong assumptions.** Where an agent guessed about the codebase and guessed wrong. That is usually a gap in `CLAUDE.md` or in `docs/QUICK_REF.md`. 3. **Friction.** Steps that were manual and did not need to be. Permission prompts that fired repeatedly for something obviously safe. 4. **Role failures.** Cases where an agent did something outside its remit, or lacked a tool it genuinely needed. ## Where each fix belongs | Finding | Goes to | |---|---| | A project fact, always true, needed every session | `CLAUDE.md` — but keep it under 200 lines | | A procedure with steps, needed occasionally | a new skill in `.claude/skills/` | | How one role should behave | that role in `.claude/agents/` | | A safe command prompting every time | `permissions.allow` in `.claude/settings.json` | | A decision and its reasoning | a new ADR in `docs/decisions/` | | A recurring gotcha | `docs/QUICK_REF.md`, if it stays under 50 lines | Prefer skills over `CLAUDE.md` when there is any doubt: `CLAUDE.md` costs context in every session, a skill costs nothi