Claude Skills

Claude skills are lightweight descriptors that surface documentation-like responses without executing any code or calling the LLM. They act as reference cards for teams and allow the agent to return a formatted summary and body when users ask for quick guidance.

Where They Live

Each Claude skill lives under .AchillesSkills/<domain>/<skill_name>/skill.md. There is no JavaScript entrypoint to write—the Markdown alone is enough for discovery and execution by ClaudeSkillsSubsystem.

# Runbook: Reset Database Passwords

## Summary
How to reset a customer database password in production.

## Body
Use this guide when a customer cannot authenticate to their managed database cluster.
1. Verify caller identity via support ticket.
2. Rotate credentials through the admin portal.
3. Confirm the rotation and note the timestamp in the incident log.

The first heading becomes the title; Summary and Body sections populate the metadata that the subsystem returns when the skill is invoked.

Execution Behaviour

When invoked, Claude skills bypass the language model and simply return the descriptor content:

Result shape: { summary, details, type: "claude" }
Memory: No session memory is created.
Side effects: None—perfect for quick references, onboarding notes, or capability catalogs.

Because they are deterministic and cheap to run, Claude skills make excellent landing points in orchestrators that need to explain a capability or expose help text before branching to heavier flows.

Authoring Tips

Use concise summaries; keep the body instructional and scannable.
Group related skills under domain folders to keep discovery organised.
Pair with orchestration skills when you want a “help” intent that returns documentation before collecting inputs.