The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and follow instructions from various project-level files, which could contain malicious or conflicting directives.
Ingestion points: In SKILL.md, Principle 2 (Recon Before Judgment) directs the agent to read CONTRIBUTING.md, CODING_STANDARDS.md, AGENTS.md, CLAUDE.md, design system documentation, and Storybook docs.
Boundary markers: There are no explicit instructions to use delimiters or ignore potential injection attempts within these ingested files.
Capability inventory: The skill possesses the capability to read any file in the workspace and execute shell commands identified during the reconnaissance phase.
Sanitization: The instructions do not specify any validation or sanitization of the content read from the external project documentation.
[COMMAND_EXECUTION]: The skill allows the execution of project-defined scripts and tools during the verification phase.
Evidence: Principle 8 (Verify What Can Be Verified) instructs the agent to run "safe, relevant checks available in the project" and "preview or test commands." These commands are discovered by reading project metadata (e.g., package.json scripts) in the Recon phase. If a malicious actor has modified these local configuration files, the agent might execute harmful code under the guise of verification tasks.

better-interface