The Agent Skills Directory

[SAFE]: The skill follows a structured, multi-step process to audit codebases without making state-changing modifications. It explicitly instructs subagents to 'report findings only, editing no code and running nothing that changes state.'
[INDIRECT_PROMPT_INJECTION]: As a tool that analyzes external codebases, the skill has an ingestion surface for untrusted data. However, the analysis is limited to reporting findings and writing plans. It does not execute the audited code or provide a mechanism for external code to influence agent behavior beyond the analysis context. Mandatory evidence:
Ingestion points: Reads arbitrary source files, documentation, and configuration files (SKILL.md step 2 and 3).
Boundary markers: Subagent prompts include specific instructions on reporting format and limitations on actions.
Capability inventory: Writing local markdown files (redesign.md, plan.md). No network exfiltration or shell execution of audited content.
Sanitization: Uses structured templates for output synthesis, reducing the risk of raw content influencing downstream prompts.
[REMOTE_CODE_EXECUTION]: No remote code execution patterns or unverified dependencies were found. All external skill references are local relative paths (e.g., ../git-metadata/SKILL.md).
[DATA_EXPOSURE]: The skill accesses local repository metadata and configuration files (package.json, go.mod, .env is not mentioned, but standard project files are). This access is necessary for its primary function and is not combined with network operations to external domains.

redesign