← ClaudeAtlas

remembering-and-learninglisted

Keep what you learned so future runs start ahead — what belongs in remember, what belongs in a skill written with learn_skill, and what belongs nowhere. Use at the end of a run where you worked something out the hard way.
mirzaaghazadeh/StandBye · ★ 6 · AI & Automation · score 75
Install: claude install-skill mirzaaghazadeh/StandBye
# Remembering and learning You start most runs with no recollection of the last one. What survives is what you wrote down: your memory notes, and your skills. Both are loaded into or listed in every future prompt, so they are valuable and they are not free — a memory full of noise makes every later run worse. ## Which one - **`remember`** — a single durable fact, in one sentence. A convention, a gotcha, a preference of the owner, a decision and its reason. It is loaded into every one of your future runs. - **`learn_skill`** — a procedure with steps, worth following again. Only its name and description sit in the prompt; the body is read on demand with `use_skill`. Anything longer than a sentence belongs here, not in memory. - **Neither** — task progress, run summaries, things already written in the repo, things that will be false next week. ## Writing a memory note Facts a teammate could act on, not impressions. - Good: "The typecheck needs `pnpm --filter @crew/shared build` first or it reports stale errors." - Good: "The owner wants release notes written before the tag, not after." - Bad: "Working on the auth refactor." — that is progress, and it will be wrong tomorrow. - Bad: "The codebase is quite complex." — unactionable. One fact per call. If you find yourself writing "and", it is probably two notes, or a skill. ## Writing a skill Write one when you had to work something out and would have to work it out again. Command sequences for this repo, a checklis