The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's external review protocol defined in references/external-review.md constructs shell commands for the gemini and codex CLIs using command substitution ($(cat ...)) inside double-quoted strings. This pattern allows any shell metacharacters or commands embedded in the claude-plan.md file to be expanded and executed by the local bash shell.
[PROMPT_INJECTION]: The skill establishes an attack surface for indirect prompt injection by ingesting untrusted data from the internet during its research phase (references/research-protocol.md). Evidence: 1. Ingestion points: WebSearch findings are collected into claude-research.md. 2. Boundary markers: No delimiters or ignore instructions are used when combining research into the specification. 3. Capability inventory: The agent has bash execution capabilities used in Step 10 and file-write capabilities throughout the workflow. 4. Sanitization: No sanitization or validation of the external content is performed before it flows into the implementation plan and shell commands.
[EXTERNAL_DOWNLOADS]: The documentation in README.md and references/external-review.md references the use of CLI tools from Google and OpenAI to perform plan reviews.

gepetto