The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Bash tool to manage skill snapshots and verify changes. Specifically, it uses mkdir, cp, and diff for file operations. It also executes uv run to invoke a local validation script located at skills/create-skill/scripts/quick_validate.py.
[INDIRECT_PROMPT_INJECTION]: The skill ingests existing agent skills as untrusted input to be audited and rewritten. This creates a vulnerability surface where malicious instructions within a source skill could attempt to influence the agent's behavior during the rewrite process.
Ingestion points: Existing skill files (SKILL.md and referenced files) located in the skills/ directory.
Capability inventory: Write, Edit, Bash(uv run *), and the Task tool for orchestration.
Boundary markers: The instructions do not define specific delimiters or "ignore" warnings for the external skill content being processed.
Sanitization: The process relies on an authoring rubric and manual review but lacks automated sanitization or filtering of instructions contained within the input files.
[DYNAMIC_EXECUTION]: The benchmarking workflow involves executing evaluation prompts against the skill configuration to verify performance. This process executes logic and instructions defined in evals/evals.json at runtime to generate output for comparison.

improve-skill