Skills
Use Context Space skills as reusable operational guidance for agents.
Skills are reusable instruction packages discovered from SKILL.md files.
They are different from source documents. A source gives knowledge. A skill gives behavior.
Source vs skill
| Item | Purpose | Example |
|---|---|---|
| Source document | Ground the answer in content. | istanbul-guide.md, refund-policy.md, runbook.pdf |
| Skill | Tell the agent how to handle a class of tasks. | Trip planning rules, support escalation rules, incident review style |
In the travel guide sample, source documents contain city guide material. The trip planning skill contains instructions for how to build practical itineraries from that material.
Register a skill source
The discovery service looks for files named SKILL.md recursively under the skill source path.
Skill file format
The current parser expects YAML front matter followed by non-empty Markdown instructions.
How skills are used
When an agent uses a Context Space, Runiq discovers skills from attached skill sources. The active skills are added to the model instructions under Runiq Skills.
Skills are also emitted as stream events:
| Event | Meaning |
|---|---|
skill_loaded | Skill instructions were loaded for the run. |
context_provided | Context Space metadata, skills, and sources were attached. |
When to create a skill
Use a skill when the behavior should be reusable across agents or across many runs.
Good skill examples:
| Skill | Use |
|---|---|
| Travel planning | How to structure realistic itineraries. |
| Support escalation | When to escalate and how to phrase customer-safe answers. |
| Incident review | How to summarize timeline, impact, root cause, and follow-ups. |
| Policy review | How to compare a request against policy documents. |
| Sales discovery | How to ask structured qualification questions. |
Do not use a skill as a document dump. Put knowledge in sources. Put behavioral guidance in skills.
Dashboard preview
The skills view shows discovered SKILL.md packages, including their version, tags, source path, and the instruction content loaded for agent runs.