The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a multi-agent feedback loop where the Generator agent reads input from specification and feedback files produced by other agents. This creates a surface for indirect prompt injection if those intermediate files were to contain malicious instructions designed to influence the Generator's tool use.
Ingestion points: The Generator role is instructed to read spec.md and feedback-NNN.md files during the implementation phase in SKILL.md.
Boundary markers: The skill relies on natural language role definitions (e.g., "You are a Generator"), but lacks explicit delimiters or instructions to ignore embedded commands within the ingested data files.
Capability inventory: The agents in this harness are granted access to powerful tools including Bash, Write, Edit, and Playwright, which allow for broad file system and network operations.
Sanitization: No sanitization or validation logic is specified for the data files passed between the Planner, Generator, and Evaluator agents.

gan-style-harness