The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from untrusted external sources, such as GitHub PR descriptions, diffs, and web search results, to generate content. This creates a surface for indirect prompt injection where malicious instructions could be embedded in PR comments or search results to influence the agent's behavior.
Ingestion points: Phase 1 reads input from marketing briefs, gh pr view, gh pr diff, and codebase files. Phase 2 performs web research on external sites.
Boundary markers: The instructions do not define explicit boundary markers or delimiters to isolate untrusted input from system instructions.
Capability inventory: The skill executes shell commands (git, gh), performs network requests (web search), and writes files to the local filesystem.
Sanitization: No sanitization or validation mechanisms are specified for the external data before it is processed by the model.
[COMMAND_EXECUTION]: The skill uses local CLI tools to gather context.
Details: It executes gh pr view and gh pr diff to fetch Pull Request data, and uses git diff and git log to analyze code changes.
[EXTERNAL_DOWNLOADS]: The skill performs research using external network operations.
Details: Phase 2 uses a web search capability to analyze competitor blog posts and market trends from arbitrary third-party websites.

blog-post