The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a 'seed command' provided by the user or project documentation to prepare the environment for verification. It also executes ffmpeg for media processing and a bundled shell script (verify-assets.sh) to perform Git plumbing operations.
[DATA_EXFILTRATION]: The skill captures screenshots and GIFs of the application UI and pushes them to a hidden Git reference (refs/verify-assets/*) on the remote origin. Because the skill requires the repository to be public for its automated inline-rendering feature to work, any sensitive information captured in the UI (e.g., tokens, PII) will be publicly accessible via raw.githubusercontent.com URLs.
[INDIRECT_PROMPT_INJECTION]: The skill reads untrusted external data such as git diff and issue descriptions to determine the scope of verification. This introduces a surface for indirect prompt injection if the source code or issue comments contain malicious instructions.
Ingestion points: git diff, linked issues, and branch names in SKILL.md.
Boundary markers: The skill does not define clear delimiters or instruction-ignore blocks when processing these external inputs.
Capability inventory: Includes Bash tool usage, Git push operations, and browser automation tools.
Sanitization: There is no explicit sanitization of content extracted from code diffs or issue descriptions before it influences the agent's actions.

verifykit