The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill sets up a CI pipeline that processes untrusted data from pull request diffs using an AI agent.
Ingestion points: The GitHub Actions workflow (.github/workflows/checks.yml) reads content from the PR diff and custom check files to build a prompt for the agent.
Boundary markers: The prompt uses basic markdown headers (## Check, ## PR Diff) to delimit instructions from data, which may be insufficient to prevent a motivated attacker from overriding agent behavior.
Capability inventory: The execution environment has access to the ANTHROPIC_API_KEY secret and the GH_TOKEN, allowing interaction with the GitHub API.
Sanitization: There is no evidence of sanitization or filtering applied to the PR diff content before it is processed by the AI.
[EXTERNAL_DOWNLOADS]: The generated GitHub Actions workflow installs the @anthropic-ai/claude-code package from the official npm registry during execution.
[COMMAND_EXECUTION]: The skill utilizes local shell commands including git for repository state management and the gh CLI for retrieving pull request comments and managing GitHub Secrets.
[DYNAMIC_EXECUTION]: The skill generates and deploys a GitHub Actions workflow script that dynamically constructs and executes agent prompts based on the repository's current state and incoming pull requests.

setup-checks