The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data from an external codebase (source code, manifests, and configuration files), which could potentially contain malicious instructions intended to influence the agent's behavior.
Ingestion points: The skill reads project instructions, manifests, and source files across the repository to establish topology as described in SKILL.md.
Boundary markers: The Discovery Protocol mandates tagging every observation with certainty levels ([measured], [inferred], or [unresolved]), which helps distinguish between verified codebase facts and potentially misleading or injected content.
Capability inventory: The skill is restricted to read-only access and the use of discovery tools (e.g., find, grep) to generate an evidence packet; it does not have file-write or network capabilities that could be exploited via injection.
Sanitization: No explicit sanitization or escaping of file content is mentioned, though the skill relies on systematic verification to mitigate misinformation.
[SAFE]: The skill adopts a professional and security-conscious posture by explicitly stating that discovery is a read-only process unless the task is changed by the user. It focuses on uncertainty reduction through evidence-based mental modeling and systematic investigation, which minimizes risks associated with automated code analysis.

fable-discover