The Agent Skills Directory

[PROMPT_INJECTION]: Instructions in SKILL.md and across several reference files (e.g., reference/problems.md, reference/business-context.md) include explicit directives to disregard specific types of user or platform-level instructions. The instructions state that settings such as 'autopilot mode', 'bias to action', or 'decide, don't ask' do not waive the requirement for a mandatory 'Discovery Interview'. This is a direct attempt to override the operational behavior and autonomy constraints of the agent to enforce the skill's specific methodology.
[EXTERNAL_DOWNLOADS]: The reference/live.md file contains detailed instructions for the agent or user to download and install an external extension called 'srs-navigator' from the author's GitHub repository. It directs the installation of files into the ~/.copilot/extensions/ directory. These resources are vendor-owned (rafael-gorski) and part of the documented 'Live' feature.
[INDIRECT_PROMPT_INJECTION]: The skill features a 'Brownfield Discovery' mode in reference/problems.md that instructs the agent to read various repository artifacts—including source code, database schemas, issue trackers, and code comments—to 'harvest evidence' for identifying business problems. This ingestion of untrusted data from the project repository represents a surface for indirect prompt injection if those files contain adversarial text.
Ingestion points: Project repository files including source code, READMEs, database schemas, and issue trackers.
Boundary markers: The skill mandates 'Discovery Interview' phases where the agent must confirm its inferences with the user before proceeding.
Capability inventory: The skill utilizes file system operations (read/write) and platform tools (such as open_canvas) to manage artifacts and visualize the SRS graph.
Sanitization: No explicit sanitization or filtering is described for the data harvested from the repository artifacts.

problem-based-srs