The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests findings from external sources (referred to as "Hub feedback findings") to generate documentation edits. This creates a surface where instructions embedded in the feedback could potentially influence the agent's behavior.
Ingestion points: Findings are read from ../review-hub-feedback/references/handoff.md and related artifacts.
Boundary markers: The skill uses a "handoff contract" for validation but lacks explicit instructions to ignore embedded prompts in the findings themselves.
Capability inventory: The skill has the capability to write to documentation files (e.g., PRODUCT-DESIGN.md, DESIGN-SYSTEM.md) and execute shell commands for documentation validation.
Sanitization: The skill relies on a robust human-in-the-loop approval process. Every proposed change must be presented as a diff and approved by the user before being applied, which serves as a primary defense against malicious injection.
[COMMAND_EXECUTION]: The skill instructs the agent to run documentation checks on the local repository.
Evidence: references/apply.md contains the instruction: "Run the repository's applicable documentation checks and inspect the final diff against approved proposals." This involves executing commands or scripts found within the consumer repository.

prepare-feedback-proposals