The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements strict behavioral constraints that override standard AI conversational patterns, specifically banning polite acknowledgments and expressions of gratitude (e.g., "You're absolutely right!", "Great point!"). It also establishes a specific signal phrase ("Strange things are afoot at the Circle K") for the agent to use if it feels unable to technically push back against feedback.
[COMMAND_EXECUTION]: The instructions direct the agent to utilize specific tools and APIs to verify feedback, including gh api for interacting with GitHub Pull Request comments and internal tools like deep_dive(symbol) and fast_refs(symbol) for codebase exploration.
[DATA_EXFILTRATION]: While the skill does not explicitly exfiltrate data, its reliance on gh api and interactions with external reviewers provides a potential path for data flow to external platforms if the agent is manipulated via the review feedback.
[INDIRECT_PROMPT_INJECTION]: The skill is intended to process "External Reviewer" feedback, which is an untrusted external data source that could contain malicious instructions designed to influence the agent's behavior during implementation.
Ingestion points: External review feedback received via GitHub or other review platforms (referenced in SKILL.md).
Boundary markers: The skill does not define structural delimiters for the external feedback but relies on a multi-step "Response Pattern" (Read, Understand, Verify, Evaluate) to mitigate risks.
Capability inventory: Access to GitHub API (gh api), code analysis tools (deep_dive, fast_refs), and the ability to modify the codebase.
Sanitization: Relies on manual technical verification against the existing codebase and "architecture-quality" checks rather than automated input sanitization.

receiving-code-review