The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches best practices, model guides, and platform documentation from official domains belonging to Anthropic (platform.claude.com, code.claude.com) and OpenAI (developers.openai.com). These downloads are used to calibrate refactoring verdicts and are sourced from well-known, trusted providers.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from existing skill files and fetched web content, which could contain malicious instructions designed to influence the refactoring process.
Ingestion points: Reads user-targeted SKILL.md files, their sibling files, and remote documentation fetched via WebFetch (SKILL.md, references/claude-code.md, references/codex.md).
Boundary markers: The instructions do not define explicit delimiters to isolate ingested data from the agent's logic during the audit.
Capability inventory: The skill utilizes the Edit tool to modify local files and WebFetch for network access.
Sanitization: No specific sanitization or filtering of external content is described before interpolation into the audit logic.
Mitigation: The process includes a mandatory human-in-the-loop step, requiring the agent to present decisions and collect sign-off using AskUserQuestion before applying any edits.

refactor-skill