The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates environment setup by creating symbolic links (ln -s) between the repository and the local agent directories (~/.claude/skills and ~/.agents/skills). This is a standard procedure documented in the repository conventions to enable cross-tool skill discovery.
[COMMAND_EXECUTION]: The diagnose mode utilizes git apply to patch the source code of other skills. This functionality is intended to harden skills against previously observed failures and is protected by a mandatory 90% confidence gate and explicit user confirmation requirements.
[INDIRECT_PROMPT_INJECTION]: The diagnostic workflow represents a potential surface for indirect prompt injection as it ingests untrusted artifacts produced by other skills.
Ingestion points: Step 2 of the diagnose mode (detailed in rules/diagnose-mode.md) reads external logs, session transcripts, and files like plan.md and walkthrough.md into the agent's context.
Boundary markers: While the skill utilizes progressive disclosure to load rules on demand, it does not specify explicit boundary delimiters (e.g., XML tags or specific 'ignore instructions' prefixes) when interpolating the content of these external artifacts.
Capability inventory: The skill possesses the ability to execute file system operations (ln -s, readlink), network operations (WebSearch, WebFetch), and source code modifications (git apply).
Sanitization: Security is addressed through a mandatory confidence(analysis) gate that requires a score of 90% or higher before any changes are proposed, accompanied by a multi-iteration refinement loop to verify evidence consistency.

create-skill