The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external implementation plans, acceptance criteria, and repository content which could potentially contain malicious instructions.
Ingestion points: Acceptance Criteria, MECE Review, and Verification Plan sections extracted from the task context and repository (Step 1).
Boundary markers: The workflow mandates a strict structural schema (Step 1) for all input sections, which serves as a validation mechanism against unstructured or malformed injection attempts.
Capability inventory: The agent can execute repository scripts, database queries, and interact with rendered UI interfaces (Step 3, 4).
Sanitization: The instructions explicitly include guardrails, such as "Do not report sensitive contents" (Step 2) and a requirement that external mutations must be explicitly authorized by the live request (Step 10).
[COMMAND_EXECUTION]: The skill involves the execution of local repository tests and scripts as part of its primary verification function.
Evidence: Steps 3 and 4 instruct the agent to resolve direct observations using repository scripts, tests, and APIs.
Context: This activity is governed by a requirement to use existing mechanisms where possible and to verify that the repository state remains consistent via deterministic fingerprints (Step 2, 10), minimizing the risk of unauthorized command execution.

verify-plan