The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes external, untrusted data (git diffs and contract files) by interpolating them into subagent prompts. This creates an indirect prompt injection surface where malicious instructions could be embedded in the code under review or the requirements documentation.\n
Ingestion points: .harnessed/contract.md, docs/superpowers/specs/, and git diff HEAD (Step 3).\n
Boundary markers: The prompt uses ---BEGIN/END EVALUATOR PROMPT--- delimiters and a "Trusted vs Untrusted Context" section to instruct subagents to ignore data-embedded directives.\n
Capability inventory: The skill executes local test suites, linters, static analysis tools (Semgrep, CodeQL, bandit), and curl smoke tests (Step 2b, 4b).\n
Sanitization: Relies on clear instructional separation and subagent role-definition rather than programmatic escaping.\n- [COMMAND_EXECUTION]: The skill facilitates the execution of project-specific verification commands and security tools as part of the quality assurance process.\n
Evidence: Step 2b and 4b in SKILL.md outline the use of local verification infrastructure, including unit tests and security scanners.\n
Context: These operations are the primary intended purpose of the skill and are performed on project-local artifacts.

independent-qa