The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from documentation files (e.g., docs/, README.md) and the codebase. It uses the Task tool to spawn subagents (Explore type) and interpolates document content directly into their prompts. A malicious documentation file containing hidden instructions could potentially influence the behavior of these subagents.
Ingestion points: Documentation files (docs/, README.md, CONTRIBUTING.md) and codebase files via Read, Grep, and Glob tools.
Boundary markers: Absent; the skill does not wrap interpolated content in delimiters or provide 'ignore instructions' warnings.
Capability inventory: Bash (restricted to git and find), Read, Grep, Glob, and Task (for subagent creation).
Sanitization: Absent; no escaping or validation of document content is performed before interpolation into subagent prompts.
[DYNAMIC_CONTEXT_INJECTION]: The file references/verification-patterns.md contains several shell commands using the !command`` syntax which are executed when the file is loaded by the agent. These commands (e.g., !ls -1 docs/, !git log, !find docs -name "*.md") are used for benign repository discovery and history checks and do not involve sensitive data access or network exfiltration.

docs-check