methodical-hermes-skill-tool-builder

To create reusable agent skills and Tool SOPs safely, you must define explicit runtime contracts, guardrails for destructive actions, and verification coverage proportional to risk. This framework replaces brittle routing with evidence-backed stop rules and expected outputs.

Progressive disclosure in skill authoring keeps the SKILL.md router lean while mapping deeper procedural details into support files. This ensures the agent loads only the context it needs for a specific runtime contract, reducing token overhead and routing failures.

To add guardrails to tool workflows with credentials and destructive actions, you define Tool SOP artifacts with explicit stop/ask rules and approval thresholds. This enforces action boundaries and requires verification steps before executing sensitive procedures.

Convert a repeated agent behavior into a runtime skill when it requires autonomous execution with eval coverage. If the procedure needs manual oversight or lacks eval-driven hardening criteria, it should become a human playbook or SOP instead.

The limitation of a lean SKILL.md router for complex tool workflows is that insufficient support-file mapping causes verification gaps. If progressive disclosure is poorly mapped, the agent may lose context for credential-backed procedures and fail runtime contracts.