The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to read, analyze, and summarize untrusted data from a target repository (source code, READMEs, configurations) to generate architecture and requirement documents. This process exposes the agent to potentially malicious instructions embedded in the target project's files.
Ingestion points: Project repository files including README.md, package.json, and source code as described in SKILL.md under 'Inputs'.
Boundary markers: The skill advises the agent to note 'Discovery Gaps' for unverified information but does not define explicit delimiters or 'ignore' instructions for the ingested repository data.
Capability inventory: The skill utilizes shell commands (ls, grep, cat, find), repository metadata tools (gh), and can write to the local .context/ directory.
Sanitization: The instructions explicitly direct the agent to 'cite each claim' and 'never fabricate', grounding the output in verifiable evidence.
[COMMAND_EXECUTION]: The skill leverages a wide range of shell commands to inspect the target repository's file structure, tech stack, and security posture. While these are read-only and localized to the target repository path, they represent a significant capability surface used for automated reverse-engineering.
Evidence: Comprehensive command lists for tech stack detection and codebase exploration are found in references/phase-1-constitution.md and references/phase-3-infrastructure.md.
[EXTERNAL_DOWNLOADS]: The skill facilitates the use of bun run api:sync, which may download OpenAPI or Swagger specifications from remote URLs to synchronize technical endpoint definitions. While a standard development practice, this involves fetching data from external network sources.
Evidence: This capability is detailed in references/context-generators.md and references/phase-2-srs.md.

project-discovery