capture-learnings

Add a short entry to learnings.md at the project root whenever you figure out surprising behavior or hit the same issue twice. Use a heading per topic with two to three sentences explaining what you learned and why it matters.

Capture surprising behavior, repeated friction, architectural decisions and their rationale, undocumented API or library quirks, and performance insights. Skip anything obvious from reading the code, standard framework behavior, or temporary debugging notes.

When a learning is referenced repeatedly, it has outgrown learnings.md. Propose adding it to an existing relevant skill or create a new one via the create-skill skill, treating the update as a Tier 2 source modification requiring verification.

Match the existing file format, typically a Markdown heading per topic followed by a brief explanation of two to three sentences covering what was learned and why it matters. Keep entries concise and focused on non-obvious knowledge.

No, it is background knowledge with user-invocable set to false. The agent reads learnings.md before significant work and updates it when it discovers something worth remembering, rather than being triggered manually.