The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several shell scripts from the repository's scripts/ directory, including lint-skills.sh, validate-spec.sh, refiner-rubric-hash.sh, refiner-history-compact.sh, and check-refiner-state.sh. These scripts are utilized for structural validation, rubric hashing, and history management during the refinement loop.
[REMOTE_CODE_EXECUTION]: The skill detects and invokes external AI CLI harnesses such as Claude Code, Codex, and Gemini CLI for peer review and smoke testing. This involves sending prompts to and processing responses from remote AI models. The skill includes logic to verify the presence of these binaries and run smoke tests in isolated temporary directories to validate their behavior before use.
[INDIRECT_PROMPT_INJECTION]: The skill processes SKILL.md files and reference files as untrusted data. Malicious content within these files could attempt to influence the agent's scoring or improvement decisions.
Ingestion points: Reads SKILL.md and referenced files during Phase 1 (Step 11a) and Phase 2.
Boundary markers: Uses tags such as <changed sections> and <git diff output> in the review prompt template to delimit untrusted data.
Capability inventory: Executes shell scripts, runs git commands (branch, commit, diff), and invokes external AI CLI tools.
Sanitization: Rule 14 explicitly requires treating candidate content as data, ignoring embedded scoring instructions, and hiding quality signals from the evaluation context.
[DATA_EXFILTRATION]: Repository content, including skill files and change diffs, is sent to external AI providers (e.g., Anthropic, OpenAI, Google) for peer review. The skill implements a privacy filter that classifies source material as public, private, or sensitive and requires explicit user authorization before sending private or sensitive content to a secondary harness.

skill-refiner