learn

# Learn Use this skill when the user correction should change future Project Khepri agent behavior. ## Steps 1. Identify the correction, preference, or repeated mistake from the user's message. 2. Convert it into a succinct generalization that future agents can apply. 3. Do not store secrets, credentials, private data, or a long transcript. 4. Update `STEERING.md` with the generalized guidance. 5. Preserve that future agents should read before work by checking `STEERING.md` before phase tasks. 6. Keep existing steering intact unless the user explicitly replaces it. 7. If the correction implies a reusable workflow, ask `khepri-evolution` to create or improve the relevant Agent Skill, hook, or AgentV eval. ## Format Add entries under `Learned Corrections` using this shape: ```markdown - YYYY-MM-DD: Generalized correction. Evidence: brief source phrase or context. ``` Prefer a short principle over a quote. For example, convert "you only inspected the files, I asked you to run the eval" into "When asked to verify agent behavior, run the AgentV eval gate instead of relying on file inspection alone." ## Edge Cases - If the user is only giving task requirements, do not add steering. - If the user is angry but gives no reusable correction, acknowledge the problem and ask what behavior should change. - If the correction conflicts with existing steering, preserve both and flag the conflict for `khepri-evolution`. - If automation captured an imperfect entry, revise `STEERING.md` manually before continuing.