Skills
Skills are reusable instructions you can attach to agents when they need specific operating knowledge or behavior.
Skills
Skills are reusable instructions you can attach to agents when they need specific operating knowledge or behavior.
Reusable behavior, not source content
Use Skills for how an agent should handle a task. Use Sources for reference knowledge. Use Tools or MCP Servers when the agent needs to fetch live data or take action.
Where Skills Live
Skills have two layers:
| Page | Scope | Use it for |
|---|---|---|
| Skills | Organization | Create, search, install, edit, view, and archive skills |
| Agent Skills | Agent | Attach skills to one agent, reorder them, and enable them |
Both pages require an admin or superadmin role.
Create A Custom Skill
Go to Skills, then choose Create Skill.
The form includes:
- name
- slug
- short description
- tags
- instructions
The name can generate a slug automatically until you edit the slug yourself. Keep the short description precise because it is part of the compact runtime catalog shown to the model.
Simple example
Imagine a skill named Warranty Escalation.
- short description: "Handle warranty claims that need human review"
- tags:
warranty,escalation,support-policy - instructions: tell the agent when to collect basic claim details, when to stop guessing, and when to hand off to a human
Install From The Skills Library
The Skills Library tab contains platform-managed recipes. Installing a recipe copies it into My Skills as a normal organization skill.
If a recipe name or slug already exists, Kravos.ai creates a unique copy instead of overwriting the existing skill.
Attach Skills To An Agent
Open Agent Skills from the agent dashboard.
From there you can:
- attach active organization skills
- create a new organization skill and return to attachment
- enable or disable a skill for this agent
- reorder attached skills
- detach a skill from the agent
Use Save order after reordering. The prompt preview shows the compact catalog that will be available to the agent at runtime.
Create, Attach, Test, Archive
- create or install the skill in the organization workspace
- attach it in Agent Skills
- test the behavior in Playground or Voice Lab
- disable or detach it if the behavior is not ready
- archive it when it should no longer be used at runtime
How Runtime Fetching Works
Kravos.ai does not inject every full skill instruction into every prompt. Instead, the prompt receives a compact catalog with each enabled skill's name, short description, and tags.
That compact catalog is just the short list the model can scan to decide whether it needs a skill. The full private instructions stay outside the base prompt until they are needed.
When the agent needs the full instructions, it uses the reserved internal tool fetch_skill.
End-user privacy
End users do not see fetch_skill tool calls or full skill instructions. Operators can still inspect
runtime behavior from dashboard views where tool details are available.
Archiving Skills
Archiving a skill makes it read-only and removes it from runtime use. Existing agent assignments remain visible, but the archived skill will not appear in the active runtime catalog.
Use archive when a skill should no longer run but you still need its history for audit or reference.
Voice And ElevenLabs
When an agent has enabled skills, voice sync can expose a managed ElevenLabs tool named fetch_skill. That lets a
voice agent fetch the same private skill instructions without placing all instructions in the initial voice prompt.
The managed key behind that tool is separate from normal user-created keys.
Best Practices
- write one skill for one repeatable behavior
- keep the short description clear enough for the model to choose the right skill
- use tags for routing hints, not long explanations
- attach only the skills one agent actually needs
- test skill behavior in Playground before relying on production traffic
- use Tools or MCP Servers for live actions instead of putting action steps in a skill
Related Docs
Sources
Add reference knowledge the agent retrieves from.
Tools
Create callable actions and live data fetches.
MCP Servers
Connect external systems through the Model Context Protocol.