When you encounter a failure, correction, or new pattern, write a learning and optionally propose a skill.
Write a learning to the appropriate directory:
learnings/failures/{YYYY-MM-DD}-{project}-{summary}.md for failureslearnings/corrections/{YYYY-MM-DD}-{summary}.md for human correctionslearnings/discoveries/{YYYY-MM-DD}-{summary}.md for new patternsInclude: what happened, what you expected, what actually occurred, what you learned, what should change.
If the learning suggests a reusable skill (something that would help ANY agent in similar future situations):
skills/proposed/{skill-name}/SKILL.mdIf you surface friction, tooling limitations, or architectural gaps during work, write an observation entry to memory/{your-agent-name}/needs.md using the human-action-needed skill (category: observation). This makes the observation durable — prose in your response text is ephemeral and the factory cannot act on it. The kernel post-run pass automatically promotes observation entries from needs.md to specs/factory-internal/ for structured operator triage via factory triage.
Update your daily log with a brief note about the learning.
skills/shared/ directly — always propose via skills/proposed/skills/shared/ first.