Skill UI

The Agent Harness provides a robust, verifiable loop structure for driving complex goals across specialized domains. It converts a high-level objective into a verifiable task plan, executes tasks sequentially using domain-specific tools, and rigorously verifies every step with machine-run checks. The system includes mechanisms for retries, budget management, and mandatory human escalation, ensuring no goal is closed until all progress is verifiably complete or explicitly waived.

This skill is designed to mitigate false-positive safety triggers from strict AI classifiers (e.g., cyber, bio/chem, reasoning). It operates by making minimal, surgical edits to sensitive phrases, abstracting away specific domains or actions while preserving the original benign intent and structure. It ensures the prompt remains policy-compliant without losing its core purpose, making it ideal for academic or defensive inquiries.

This skill helps developers create highly focused, folder-specific documentation (`CLAUDE.md` and `AGENTS.md`). It systematically analyzes all files in a target directory to extract project purpose, operational constraints, naming conventions, and critical decisions. This deep context ensures that future AI agents (like Claude or Codex) operate with localized knowledge, preventing them from relying on global, generalized instructions and improving their accuracy in that specific domain.