The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests user feature requests to generate a specification file at specs/<feature-name>.md, which is then used to drive code generation and verification. This introduces a surface for indirect prompt injection where malicious instructions in a request could influence the agent's implementation or testing behavior. 1. Ingestion points: User requests and existing artifacts are processed and stored in specification files. 2. Boundary markers: The instructions do not mandate the use of delimiters or 'ignore embedded instructions' warnings when creating the spec. 3. Capability inventory: The skill utilizes file-writing, code generation, and shell command execution. 4. Sanitization: No explicit sanitization or validation of the input content is described before interpolation.
[COMMAND_EXECUTION]: The workflow involves an implementation phase and a verification phase that may require executing shell commands to build or test code. The skill documentation includes safety notes advising the use of sandboxed environments and human gates for any external or production-impacting actions.

ai-loop