What problem does it solve? Creating a domain-focused coding agent that actually improves over time requires more than a prompt — it needs a bounded spec, persistent knowledge, and a mechanism that forces every session to leave the system better than it found it. This Skill scaffolds that entire structure so agents compound learning instead of starting from zero each session. ## Core Features & Use Cases - Agent scaffolding: Generates .claude/agents/<name>.md with a domain boundary, operating loop, guiding principles, and a scoped tool allowlist, plus knowledge.md and logs.md memory files. - Mechanical self-improvement enforcement: Installs a SubagentStop hook that blocks a domain expert from ending its session until it appends a log entry containing a Bottleneck, Counterfactual, and Patch. - Deterministic validation: Ships validate-scaffold.mjs to check frontmatter fields, unique colors, tool allowlists, memory files, and required spec sections before the agent is considered ready. - Use Case: You want an agent that owns only your chat backend. Run the skill with chat-backend-developer, define what it owns and does not own, and get a validated agent that patches its own knowledge after every task. ## Quick Start Ask the assistant to create a codebase expert for a specific subsystem, for example by invoking /create-codebase-expert chat-backend-developer and answering the boundary question when prompted.
