The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill instructs the agent to process untrusted repository data (source code, tests, configuration) to construct knowledge maps. Maliciously crafted content within these files could attempt to manipulate the agent's analysis or documentation tasks.- Ingestion points: SKILL.md and references/knowledge-schema.md define workflows for reading and analyzing source code, documentation, and configuration files.- Boundary markers: The skill lacks explicit prompt delimiters for external content, though it utilizes a claim classification system (observed, inferred, proposed) to distinguish between fact and interpretation.- Capability inventory: The agent possesses the ability to read repository files and write to durable documentation locations (e.g., READMEs, ADRs) upon user approval.- Sanitization: The references/update-policy.md defines strict rules against persisting secrets, raw command output, or personal data, requiring redacted summaries instead.- [COMMAND_EXECUTION]: Evidence Collection via Commands. The skill's workflow includes executing repository-local commands to gather evidence for architectural claims.- Evidence: SKILL.md and references/knowledge-schema.md highlight the use of command results to support 'observed' claims. The skill operates within the agent's standard execution environment and does not attempt to acquire elevated privileges.

graph-engineering