skill-registry

To generate a compact catalog of project skills, you can use a skill registry to scan SKILL.md files and pre-digest them into 5-15 line actionable rule blocks. This prevents sub-agents from needing to read many individual files by injecting resolved rules directly into prompts.

Preventing prompt injection during agent delegation involves using pre-resolved constraint mappings extracted from project conventions. Generating a skill registry isolates actionable rules from raw SKILL.md files, providing a secure, compact rule set for injection.

You can deduplicate project conventions and skill instructions by scanning known user-level and project-level skill directories, keeping only the most specific version of each skill. This automated process ensures your sub-agents use the correct, resolved rules.

The skill registry explicitly skips workflow-only skills and processes YAML frontmatter from discovered SKILL.md files. It focuses on extracting delegation-ready trigger mappings and constraints to persist in the .atl/skill-registry.md file.

You collect project conventions for agent configuration by detecting convention index files like agents.md or AGENTS.md. The registry automatically includes both the index file and all referenced paths so sub-agents avoid extra file reading hops.

You should update the skill registry artifact whenever you add or remove skills, prompting the AI to regenerate the .atl/skill-registry.md file. This ensures current compact rules are persisted and optionally saved to engram for cross-session reuse.