Set up a learnings inbox

This skill bootstraps a lightweight system for capturing what the agent learns while working in a repo, and wires the repo's agent router file(s) — CLAUDE.md and/or AGENTS.md — to read the accumulated learnings before work and record new ones into a staging file (the inbox).

The design deliberately splits two jobs:

• Capture (cheap, ambient, automated): while doing any task, the agent appends surprising or non-obvious findings to docs/learnings/INBOX.md. This is driven by an always-loaded instruction in the router file(s), because the trigger ("huh, that wasn't what I expected") can fire during any task and can't be predicted in advance.
• Curation (deliberate, human-in-the-loop): promoting good inbox entries into the curated docs/learnings/LEARNINGS.md, pruning stale ones, merging duplicates. That is a separate, invoked process — out of scope for this skill.