Skills

Skills define how the AI behaves in CableKnit’s chat interface for your plugin’s use case. When a company installs your plugin, your skills become part of the AI’s context — shaping its knowledge, tone, and capabilities for that industry or workflow.

What skills do

Each skill has a system_prompt that is injected into the LLM’s context when the skill is active. The prompt tells the AI what it knows, how to behave, and what it can help with.

Skills also have a trigger_description — a plain-language description of when the skill should activate. CableKnit’s AI uses this to route conversations to the right skill automatically. A user asking “how do I submit an RFI?” triggers a construction skills package; a user asking “what’s the invoice status?” triggers an accounting skill.

Skill file format

Each skill is a JSON file in your bundle’s skills/ folder.

{
  "name": "Subcontractor Intake",
  "slug": "subcontractor-intake",
  "description": "Handles initial subcontractor onboarding inquiries.",
  "system_prompt": "You are an onboarding assistant for a construction company. You help project managers and compliance officers manage incoming subcontractor applications.\n\nYou can:\n- Explain the onboarding process and required documents\n- Help assess whether a subcontractor application is complete\n- Guide users through compliance requirements\n- Draft follow-up communications to subcontractors\n\nAlways be professional and specific. When a user asks about a subcontractor, ask for the company name and trade if not provided.",
  "trigger_description": "Use this skill when a user asks about subcontractor onboarding, applications, compliance documentation, or trade contractor intake.",
  "action_type": "chat",
  "position": 1
}

Field reference

How skills are injected

Your skill’s system prompt is wrapped in XML and injected into the LLM’s system context:

<skills>
<skill name="subcontractor-intake" plugin="subcontractor-onboarding">
You are an onboarding assistant for a construction company...
</skill>
</skills>

This means your prompt text should not include XML tags — they’ll be added automatically.

Writing effective system prompts

A good skill system prompt does four things:

1. Establishes a clear role

Start with a one-sentence role statement. Be specific about the industry and function.

You are a compliance assistant for food distribution companies operating under FSMA regulations.

Not:

You are a helpful assistant.

2. Defines capabilities explicitly

List what the AI can and cannot help with. This prevents the AI from wandering into topics your plugin doesn’t cover.

You can:
- Explain FSMA 204 traceability requirements
- Help users assess whether their lot code records are compliant
- Draft corrective action notices for traceability gaps

You cannot:
- Provide legal advice
- Access real-time FDA databases
- Process actual product recalls

3. Specifies format expectations

If you want structured responses, say so.

When assessing a traceability record, always respond with:
1. A compliance status (Compliant / Non-compliant / Needs review)
2. A list of any missing or incomplete fields
3. Recommended next steps

4. Handles missing context gracefully

Tell the AI what to do when it doesn’t have enough information.

If the user hasn't provided a specific lot code or supplier name, ask for it before proceeding with an assessment.

Multiple skills in one plugin

A plugin can have multiple skills. Each skill handles a different aspect of your use case. The AI routes to the right skill based on trigger_description.

For example, a construction plugin might have:

• rfi-management — handles RFI creation, routing, and tracking
• submittal-review — handles submittal workflows and approvals
• daily-log-assistant — helps with daily log entries and exceptions

Keep skills focused. A skill that does everything has a weaker trigger_description and produces less precise responses than a skill with a clear, narrow purpose.