What problem does it solve? When you manually fix code an AI agent wrote, that feedback is usually lost — the agent repeats the same mistake next session. This Skill captures those manual edits, infers the general coding preference behind each one, and records them as durable conventions in the project's AGENTS.md or CLAUDE.md so corrections never need to be made twice. ## Core Features & Use Cases - Edit Attribution: Separates your manual edits from the agent's own uncommitted changes in the working tree by reconstructing what the agent wrote and diffing against current content, asking you when attribution is ambiguous. - Preference Generalization: Filters edits to keep only those that generalize — style, structure, naming, error handling, testing patterns — while excluding one-off bug fixes and business-logic corrections. - Persistent Convention Log: Maintains a dated, deduplicated "Conventions Learned from Manual Edits" section in CLAUDE.md or AGENTS.md, refining or replacing rules as preferences evolve. - Use Case: After a long coding session, you tell the agent you tweaked its output. The skill identifies your hunks, records rules like "wrap errors with fmt.Errorf including the failed operation", and applies them to all subsequent code. ## Quick Start Tell the agent: I manually changed some of the code you wrote — learn from my edits and record the conventions.
