The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user requests and existing codebase content to generate and execute "deterministic validation commands." This creates a surface for indirect prompt injection where malicious content in the codebase or request could influence the generated commands.
Ingestion points: The agent explores the existing codebase (Step 2) and processes user-supplied build requests to formulate plans.
Boundary markers: The protocol utilizes "holdout scenarios" to prevent implementation gaming, but it does not specify the use of delimiters or instructions to ignore embedded prompts within the processed codebase files.
Capability inventory: The skill writes persistent state to the filesystem (IMPLEMENTATION_PLAN.md) and executes shell commands for task validation (Step 5).
Sanitization: There are no documented mechanisms for sanitizing or escaping content retrieved from the codebase before it is used to generate implementation plans or validation scripts.
[COMMAND_EXECUTION]: The workflow requires the agent to perform command execution at runtime to run "deterministic validation commands" (Step 5) as a gate for task completion.
[DYNAMIC_EXECUTION]: The skill generates implementation steps and corresponding validation logic based on its interpretation of the project requirements, which are then executed within the shell environment.
[EXTERNAL_DOWNLOADS]: The documentation references external resources and the official protocol repository at github.com/agent-frontier/wgm.

wgm