The Agent Skills Directory

[COMMAND_EXECUTION]: The skill identifies and executes build or test commands found in the project's configuration files (e.g., package.json, go.mod, pyproject.toml). This behavior is consistent with its role as a QA tool but involves executing strings provided by the local environment.
[PROMPT_INJECTION]: An indirect prompt injection surface exists because the skill reads project-specific files to formulate tasks for its sub-agents.
Ingestion points: The skill reads PLAN.md, CLAUDE.md, AGENTS.md, and files within the docs/ directory to determine project standards and success criteria.
Boundary markers: The sub-agent prompts defined in 'references/subagent-prompts.md' do not utilize delimiters or specific instructions to ignore potentially embedded directives within the ingested data.
Capability inventory: The skill can execute arbitrary shell commands, write files to the filesystem, and perform network requests via CLI tools.
Sanitization: No sanitization or escaping of the ingested file content is performed before it is interpolated into the prompts for sub-agents.
[DATA_EXFILTRATION]: The skill uses curl and Playwright for network-based verification of API endpoints and frontend reachability. While functionally necessary for its stated purpose, these tools provide a capability for outbound data transfer if combined with sensitive data access.

qa