← ClaudeAtlas

record-what-you-learnedlisted

Use when a run is finished and the report is written, after the report is written, to record one reusable lesson for the next run in this field. Covers what counts as a lesson worth passing on, what must never be passed on, and how to write it.
tangxiangru/AutoR · ★ 805 · AI & Automation · score 77
Install: claude install-skill tangxiangru/AutoR
# Leave one note for the next run in this field You hit things this run that were in no prompt: an archive that stores its axis in an unexpected order, a reference implementation needing an undocumented flag, a check that caught a mistake you would otherwise have shipped, a step the field treats as obvious and no instruction mentioned. The next run in this field hits the same thing unless you write it down. ## How Run this once, at the end: ``` python3 -c " import sys; sys.path.insert(0, '<AUTOR_ROOT>') from src.skill_evolution import record_note note, problems = record_note( discipline='<the field, e.g. earth>', title='<short, routable, what the lesson is about>', body='''<what you hit, and what to do instead next time>''', learned_in='<this task id>') print(problems or 'recorded') " ``` A good note is one paragraph answering: what surprised you, how it shows up, and what to do instead. Write it for someone competent who has not seen this corpus. ## What must never go in **No results.** Not your numbers, not the paper's, not "it came out around X". A note travels to a *different task*, and a finding that travels is contamination — it invites the next run to expect an answer instead of measuring one. The recorder refuses notes containing measured values, and that refusal is not an obstacle to work around. **Nothing you did not hit.** A guess about what might help is prose, and prose accumulates until nobody reads the pool. A run that learned nothing t