The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses standard shell commands (test, mkdir, git) for repository state verification and scaffolding. These are used in a controlled manner to build the wiki structure.
[DATA_EXFILTRATION]: The skill reads repository metadata (paths, commit logs, and configuration files like package.json or go.mod) to understand the project structure. It does not perform any network operations and the data remains within the local environment's .repo-wiki/ directory.
[SAFE]: The skill processes untrusted repository metadata which is a theoretical surface for indirect prompt injection. However, it does not execute this data or perform network operations, and it adheres to a 'WHY-not-WHAT' principle that excludes source code content.
Ingestion points: Root configuration files (package.json, go.mod, etc.), git ls-files output, git log output, and README.md headings.
Boundary markers: None explicitly implemented for written output, though the skill is instructed to avoid implementation code.
Capability inventory: Local file system write access and git command execution; no network access.
Sanitization: None mentioned for the metadata incorporated into markdown files.
[SAFE]: The skill modifies the CLAUDE.md file in the repository root. This is an intended integration feature for the Claude Code platform to maintain agent context, not a malicious persistence mechanism.

init