apw-learnlisted

Use this skill to turn a lesson into a durable project rule. It exists to fight one specific limitation: a model has **no memory across sessions** — conventions discovered, mistakes corrected, and preferences stated all evaporate when the context window resets, so the same ground gets re-learned and the same mistakes repeat. The fix is to write the lesson into the project's always-on instructions (`AGENTS.md`), which every future session reads. This skill is a **cross-cutting reflex**, not a chain stage: run it after any task, when the user hands you a rule, or when you yourself notice something worth keeping. The one non-negotiable rule: **never edit `AGENTS.md` without explicit user confirmation.** Follow the steps below in order. ## Step 1 — Determine the mode Figure out which of three entry paths you are in, because the source of the candidate lessons differs: - **A · Existing conversation.** The skill was invoked inside a chat that has real history. Mine that conversation for lessons. - **B · New chat.** The skill was invoked with little or no prior context, and the user supplied the rule or data as arguments. The user's input *is* the candidate. - **C · Agent-initiated.** You reached for this skill on your own, mid- or post-task, because you figured out something that future sessions should know. The thing you just learned is the candidate. State which mode you're in before continuing. ## Step 2 — Gather candidate lessons Collect the candidate rules according to t