The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes feature maps and project source code to verify documentation accuracy, which introduces a surface for instructions embedded in data to influence agent behavior.
Ingestion points: Reads project-local SKILL.md files, feature map files, and various source code files across the project during the 'Source wave' and 'Reconcile' steps.
Boundary markers: Instructs sub-agents to operate in a 'read-only' mode for inspection, but does not define explicit delimiters or 'ignore' instructions for the interpolated source code data.
Capability inventory: The skill possesses capabilities for file editing (within specific directories), sub-agent delegation (delegate_task), GitHub CLI operations (gh pr create), and shell execution for harness scripts.
Sanitization: No explicit sanitization or filtering of content read from source files is mentioned before the data is processed by the LLM.
[DYNAMIC_EXECUTION]: The skill executes local 'harness scripts' that are part of the target project being audited.
Evidence: Step 4 ('Live pass') involves driving the application using the verification skill's own harness scripts, which may involve executing shell scripts or other binaries found in the local file system.
[COMMAND_EXECUTION]: Utilizes system tools and external CLI utilities to perform its maintenance cycle.
Evidence: Uses the GitHub CLI (gh pr create) to submit changes and executes local scripts as part of the verification process to exercise application features.

maintain-verification-skill