learnlisted

# learn You just finished real work. Some of it taught you something — a tool that behaved differently than you assumed, a step you had to rediscover, a correction the user made twice. This skill turns that into a durable improvement to your own skills, rules, or memory, so the next agent (or you, next week) starts ahead of where you started. It is the opposite failure mode that makes this hard. The lazy output is to encode nothing. The eager output is to encode everything — to overfit a permanent rule to one session's fluke, or to rewrite a battle-tested skill around a single bad afternoon. Both degrade the system. Your job is the narrow middle: the few lessons that are real, general, and durable — and the discipline to ship them without breaking what already works. **Not `reflect`.** `reflect` is mid-conversation: it recalls the user's feedback before your next attempt and writes nothing. `learn` is post-session: it writes durable lessons forward into the skill library. Different jobs — don't conflate them. ## Two modes — pick by the argument - **No argument, a session id, or a free topic → reflection mode.** Reflect on the session you just finished (or the one named) and write its lessons forward. This is everything from **Phase 0** down — the default. - **The argument names an installed skill, plugin, command, or a workflow → target-audit mode.** The user wants to know where *that one thing* keeps going wrong across all the times they've used it, and fix it. Run **Ta