The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted text from codebase files and PR descriptions, presenting a surface for indirect prompt injection where malicious content could attempt to manipulate the agent's humanization logic.\n
Ingestion points: Scans local markdown and source code files, and fetches PR bodies via gh pr view as defined in references/review.md.\n
Boundary markers: Employs a structured JSON intermediary (ai-writing-review.json) and a dedicated review-verification-protocol.md that requires the agent to echo artifacts before judging them, mitigating confabulation and instruction override risks.\n
Capability inventory: Accesses the filesystem for reading and writing, executes shell commands for git operations and syntax checking (git, jq, python3, npx), and modifies codebase files in 'fix' mode as seen in references/fix.md.\n
Sanitization: Implements post-modification validation steps for Python and JavaScript/TypeScript files using ast.parse and acorn respectively to ensure that 'humanizing' edits do not introduce syntax errors.

do-humanize