The Agent Skills Directory

[SAFE]: The skill provides a high-level coordination framework for AI agents and does not include any executable binary files, remote script downloads, or requests for sensitive user credentials.
[COMMAND_EXECUTION]: The workflow specifies the use of Git commands (such as git worktree, branch, and commit) to maintain isolated development environments for different tasks. The instructions include explicit safeguards, such as forbidding the use of git stash and other state-altering commands in shared worktrees to prevent accidental data corruption or leakage between builders.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external content, including user-defined goals, constraints, and reference artifacts.
Ingestion points: User-provided goals, qualitative/quantitative 'bars' for measurement, and reference artifacts (visuals, code, or text) are ingested into the context (SKILL.md, Section 1).
Boundary markers: The process implements strict isolation by launching 'fresh critics' for each round, who are provided only with the objective data and withheld the builder's rationale or history, preventing cross-contamination of instructions (SKILL.md, Section 4).
Capability inventory: The skill utilizes Git for repository management and basic filesystem operations for creating checkpoints and storing screenshots (SKILL.md, Sections 3 and 5).
Sanitization: While no explicit string sanitization is mentioned, the structural separation of roles and contexts acts as a functional barrier against prompt-based influence between iterations.

gauntlet-loop