The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local script utilities/skill-builder/scripts/record_skill_feedback.py to persist feedback after execution, using arguments generated from the session data.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the ingestion of untrusted user input and existing documents.
Ingestion points: User feature descriptions and existing files in docs/brainstorms/.
Boundary markers: The skill does not define explicit delimiters for untrusted data interpolation, though it provides a general instruction to treat user input as untrusted.
Capability inventory: The skill can write files to the project directory and execute a local Python script.
Sanitization: Instructions require the redaction of PII and secrets from any generated artifacts.
[PROMPT_INJECTION]: Test cases in references/evals.yaml include prompt injection examples (e.g., 'ignore previous instructions') designed to verify agent safety and are benign in this context.

ce-brainstorm