Documentation / guides
Overview
Add reusable, on-demand instructions to a Harness Loop without placing every procedure in its system prompt.
A skill is a named Markdown procedure that a model can load when the current task needs it. The reusable implementation lives in github.com/looprig/tools/skill; Harness supplies the tool execution, gate, Loop, and Session boundaries around it. This keeps a focused procedure out of every model request while still making its name and purpose discoverable.
How skills fit
%%{init: {"theme":"dark"}}%%
flowchart LR
C[Skill catalog] --> P[Runtime context]
P --> M[Model chooses a skill]
M --> T[Skill tool]
T --> G{Access gate}
G -->|approved| B[SKILL.md body]
B --> MThe catalog advertises metadata, not bodies. The model calls Skill with { "name": "weather-briefing" }; the tool prepares an access request and returns the authorized body as a tool result. Register the Skill tool with Loop tools just like any other tool.
Choose a source
| Source | Location | Trust model | Typical use |
|---|---|---|---|
| Embedded | skills/<name>/SKILL.md in the application binary | Curated by the application and restricted by a per-agent allow-set | Stable product procedures |
| Workspace | .skills/<name>/SKILL.md under a workspace root | Project-controlled and untrusted | Repository-specific conventions |
Embedded names win if both sources contain the same name. Workspace bodies are snapshotted during call preparation so the approved bytes cannot be replaced before execution.
Continue with SKILL.md format, then follow the complete Weather Assistant to see a model load a skill and call a domain tool in one turn.