The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides a specific workflow in motion.md that involves generating an HTML file and executing a TypeScript snippet via Playwright to capture frames, followed by encoding them with ffmpeg.
Evidence: motion.md contains a code block for Playwright (await page.goto('file://${htmlPath}')) and a shell command for ffmpeg (ffmpeg -framerate 30 -i frames/f%05d.png ...).
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from the repository (diffs, code paths, and console outputs) and incorporate it into pull request and issue bodies, creating a vulnerability surface for malicious instructions embedded in code to influence the agent.
Ingestion points: SKILL.md (Step 1: Gather the ground truth... diff... code path) and pr.md (Step 2: Copy its real output).
Boundary markers: The skill instructs using a heredoc (cat <<'EOF') to wrap markdown for shell execution, which provides basic shell safety but does not isolate the LLM from instructions inside the diff.
Capability inventory: The skill utilizes git (read), gh (write/create), and node (Playwright execution).
Sanitization: Uses shell heredocs to prevent markdown content from being interpreted by the shell as commands.
[COMMAND_EXECUTION]: The skill relies heavily on executing shell commands to interact with the environment and GitHub API.
Evidence: Multiple files use git diff, git rev-parse, gh pr create, gh issue create, and curl for connectivity checks.

pr-issue