The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to use the Bash tool to perform system-level tasks such as running project scripts (./scripts/gan-harness.sh) and managing development servers (npm run dev). This is a necessary capability for an autonomous coding agent but carries inherent risks if the agent's logic is subverted.
[PROMPT_INJECTION]: The workflow relies on an iterative feedback loop where agents read and interpret content from files generated in previous steps, such as spec.md and feedback-NNN.md. This architecture creates a surface for indirect prompt injection if malicious instructions are embedded in the generated project data.
Ingestion points: The Generator and Evaluator agents read spec.md and feedback files generated during the build process (SKILL.md).
Boundary markers: There are no specific instructions or delimiters mentioned to ensure the agent treats the content of these files as data rather than executable instructions.
Capability inventory: The skill utilizes Bash, Write, Edit, and Playwright MCP to interact with and modify the local environment (SKILL.md).
Sanitization: No explicit sanitization or validation of the generated specifications or feedback files is described in the harness instructions.
[SAFE]: The skill references several external resources from well-known technology organizations, including Anthropic, OpenAI, and technical engineering blogs. These references are used for architectural context and do not pose a security risk.

gan-style-harness