The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill mines verbatim review comments from GitHub to synthesize a new 'Review as Me' skill. This introduces a vulnerability where malicious comments in the source data could manipulate the instructions or philosophy of the generated skill.
Ingestion points: GitHub review history (inline comments, review body, conversation comments) via gh api.
Boundary markers: None. The skill explicitly requests 'verbatim voice' and real comments for calibration without delimiting them from the instructions.
Capability inventory: File writing to ~/.claude/commands/, shell command execution, and gh api read/write access.
Sanitization: No sanitization, escaping, or filtering of the extracted comments is performed; they are used directly to build the 'Calibration Examples'.
[DYNAMIC_EXECUTION]: The skill's primary function is to generate a new functional markdown skill file at ~/.claude/commands/review-as-me.md. This generated skill includes instructions, core philosophy, and behavior profiles derived dynamically from external, untrusted data synthesized at runtime.
[COMMAND_EXECUTION]: Utilizes gh api and git log to discover user identities and fetch large volumes of repository metadata and review content from GitHub.
[DATA_EXFILTRATION]: Aggregates professional and organizational review history, including communication styles and priority signals, into a temporary file (/tmp/review-persona-{username}.md). This creates a significant local aggregation of potentially sensitive internal behavioral data.

build-review-persona