The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the codebase, such as JSDoc comments, source code signatures, and test files, to generate README content. This creates a surface for indirect prompt injection where malicious instructions hidden in code comments could influence the agent's behavior or the output of the refiner skill.
Ingestion points: The skill reads package.json, source code entry points, test files (*.test.ts), and extracts JSDoc @example blocks. It also processes a findings: list from the invoking prompt.
Boundary markers: While the skill instructs the agent to treat prompt-level findings: as factual data, it lacks explicit boundary markers or instructions to ignore embedded commands for data extracted from the project's source files.
Capability inventory: The skill writes to README.md and invokes the accelint-english-manager sub-skill for prose refinement.
Sanitization: There is no evidence of sanitization or validation for content extracted from JSDoc or test files before it is interpolated into the documentation draft.
[COMMAND_EXECUTION]: The skill's logic depends on detecting and using package manager commands (pnpm, npm, yarn, bun) derived from local lockfiles.
Evidence: SKILL.md and the references/writing-principles.md file provide specific mappings for lockfiles to installation commands. A manipulated environment (e.g., a spoofed lockfile) could cause the agent to generate or potentially execute incorrect or malicious command strings during the discovery and documentation workflow.

accelint-readme-writer