The Agent Skills Directory

[Indirect Prompt Injection Considerations]: The skill performs research by ingesting data from external web sources and repository files. This activity involves a surface for indirect prompt injection where instructions embedded in external content could attempt to influence the agent. The skill includes specific guardrails, instructing the agent to treat this data as inert and to flag potential injection attempts, which align with safety best practices.
Ingestion points: External research via web fetch and codebase search (SKILL.md, references/research.md).
Boundary markers: Explicit instructions to ignore embedded directives and treat data as inert.
Capability inventory: File writing (to tracking folder), user questioning (vscode_askQuestions), and subagent delegation.
Sanitization: Logical separation of fetched content from instructional logic.
[Credential Safety]: The instructions explicitly prohibit the handling of secrets such as API keys, tokens, or credentials, requiring they be excluded from all artifacts and logs.
[Scoped File System Access]: The skill is constrained to writing research artifacts only within the .copilot-tracking/research/ directory, preventing unintended modifications to the project source or other sensitive areas.

rpi-research