teach-agentslisted

# Capture Session Learnings into AGENTS.md The goal: every session that fights through a real bug, a tooling quirk, or a non-obvious architectural constraint should leave that knowledge behind for the next agent, instead of making them rediscover it. This skill runs at the end of (or during) a session, over the conversation you already have full context of — do NOT delegate this to a subagent via the Agent tool, since a fresh agent has none of this session's context and would have nothing to review. ## Process ### 1. Mine the session for candidate nuances Review the conversation from the start. Look specifically for things that were **surprising, costly, or non-obvious from reading the code alone**: - A bug whose root cause wasn't where it looked (wrong layer, wrong assumption, a library/runtime limitation rather than a logic error). - Something that took more than one attempt to get right, or that you had to verify empirically (a script, a repro, a real render) rather than reason out from the source. - A tool/library/framework limitation discovered the hard way (e.g. "mpdf ignores `transform`", "this test runner needs X installed first"). - A convention, seam, or existing mechanism in the codebase that turned out to be the "right place" for a fix, especially if you almost put the fix somewhere else first. - A class of bug that could recur elsewhere (e.g. "never key rows by column label — labels aren't unique") rather than a one-off typo fix. - Anything the