Agent Skills
Procedural guidance for the agent, with execution authority kept in the broker.
Semwright supports the Agent Skills package shape as an agent-side knowledge layer. A Skill can explain when and how to compose capabilities. It cannot grant authority or execute as an independent Semwright principal.
Keep four concepts distinct
Section titled “Keep four concepts distinct”| Surface | Responsibility |
|---|---|
| Skill | Procedural guidance loaded by an agent |
| Capability | A typed executable operation |
| Recipe | A declarative program over capabilities |
| Learned workflow | An explicitly recorded procedure, verified and promoted |
There is no skill.execute capability. A Skill’s tool metadata is not a Broker grant.
Discover progressively
Section titled “Discover progressively”A package contains SKILL.md with its name and description, with optional references, scripts and assets. Load the relevant guidance instead of flooding an agent with every integration document.
The repository includes focused guidance for Figma, Blender, Godot, audio/video production, artifact handoff, project continuity and workflow distillation. These packages explain operating loops; they do not certify every application environment.
Execute through the same boundary
Section titled “Execute through the same boundary”Skill text, application content and provider metadata are non-authoritative. Actual requests still enter schema validation and broker policy with real arguments and targets. A policy preview is informational and is rechecked at execution.
Use Recipes for explicit multi-step programs and workflow distillation for recording and promoting reusable procedures.
Source: Agent Skills interoperability and source Skill packages, source references checked 7 October 2026.