The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process other SKILL.md files from the user's filesystem (~/.claude/skills/). These files are used to generate test prompts and evaluate agent behavior. If a targeted skill contains malicious instructions, they could influence the agent during the optimization or evaluation process.
Ingestion point: Reads content from all files matching .claude/skills/*/SKILL.md.
Capability inventory: The skill has the ability to write to files, execute git commands (commit, revert), and spawn sub-agents to run test scenarios.
Boundary markers: The instructions do not specify the use of delimiters or 'ignore' warnings when interpolating content from external skills.
Sanitization: No validation or sanitization of the content from other skills is described before processing or execution.
[EXTERNAL_DOWNLOADS]: The README provides a link to download the skill as a ZIP file from a remote R2 bucket (pub-161ae4b5ed0644c4a43b5c6412287e03.r2.dev). While R2 is a Cloudflare service, the specific bucket is user-defined and not a verified official source.
[COMMAND_EXECUTION]: The skill uses shell commands to manage version control (git add, git commit, git revert) and to generate visual reports. It specifically references npx playwright screenshot and executes a local Node.js script (scripts/screenshot.mjs) which internally uses execSync to open generated files. The script also contains a hardcoded local absolute path (/Users/alchain/...) for playwright-core.
[DYNAMIC_EXECUTION]: The skill's workflow involves spawning 'sub-agents' to execute and compare outputs of different skill versions. This effectively executes instructions derived from other skills in an autonomous loop.

darwin-skill