The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill "interviews" the repository codebase to generate verification instructions, which introduces a potential attack surface if the repository contains malicious content intended to influence the agent's behavior.
Ingestion points: The skill reads repository structure, documentation, README files, package scripts, routes, and menus to identify features and drive patterns.
Boundary markers: No specific delimiters or "ignore previous instructions" warnings are utilized when processing content from the codebase.
Capability inventory: The resulting verification skill has the ability to launch processes, execute shell commands, and interact with web browsers or CLIs.
Sanitization: There is no explicit requirement or process for sanitizing or validating repository content before it is interpolated into the generated skill's instructions.
[DYNAMIC_EXECUTION]: The skill follows a "generate then run" pattern where it creates a project-local skill and then immediately executes its instructions.
Evidence: Section 4 of SKILL.md instructs the agent to "Run its own instructions end to end once: launch, doctor, drive ONE mapped feature," which involves executing shell commands defined by the dynamically generated output.
[COMMAND_EXECUTION]: The skill relies on shell command execution for its primary verification tasks.
Evidence: The references (e.g., references/feature-map-example/README.md) utilize specialized CLI tools like control-notes to drive browser and terminal actions, which are executed directly on the host system to verify application behavior.

create-verification-skill