Skills

Overview

Skills are reusable instruction blocks you can attach to sub-agents. They help you share behavioral rules, reasoning guidelines, and formatting standards across multiple sub-agents without duplicating prompt content.

Use skills to:

Standardize behavior across sub-agents (tone, formatting, guardrails)
Separate concerns by moving specialized instructions out of the main prompt
Load instructions on demand to reduce context size when not needed

Tip

For the Visual Builder approach, see Skills in the Visual Builder.

Defining Skills

Skills are defined as SKILL.md files with YAML frontmatter. Each skill lives in its own directory.

File Structure

skills/
├── weather-safety-guardrails/
│   └── SKILL.md
└── structured-itinerary-responses/
    └── SKILL.md

Note

The directory name must match the name field in the frontmatter.

SKILL.md Format

skills/weather-safety-guardrails/SKILL.md

---
name: "weather-safety-guardrails"
description: "Safety rules for weather-dependent activity recommendations"
metadata:
  author: acme-corp
  version: "1.0"
---

- Never recommend outdoor activities during severe weather alerts.
- Always mention indoor alternatives when precipitation exceeds 50%.
- Include safety disclaimers for extreme temperature conditions.

Frontmatter Fields

Field	Type	Required	Description
`name`	string	Yes	Skill identifier. Must match the parent directory name.
`description`	string	Yes	When to use this skill. Used by the agent to decide whether to load the skill.
`metadata`	object	No	Additional key-value pairs (author, version, tags, etc.).

The body content after the frontmatter contains the actual instructions in Markdown format.

Loading Skills

Use loadSkills() to load all skills from a directory:

import path from 'node:path';
import { loadSkills, project } from '@inkeep/agents-sdk';
import { myAgent } from './agents/my-agent';

export const myProject = project({
  id: 'my-project',
  name: 'My Project',
  agents: () => [myAgent],
  skills: () => loadSkills(path.join('my-project/skills')),
  models: {
    base: {
      model: 'openai/gpt-4o',
    },
  },
});

The loadSkills() function:

Scans the directory for */SKILL.md files
Parses each file's frontmatter and content
Validates that the name matches the directory name
Returns an array of SkillDefinition objects

Attaching Skills to Sub-Agents

Add skills to a sub-agent using the skills configuration option.

Reference by ID

Reference skills defined at the project level by their ID:

import { subAgent } from '@inkeep/agents-sdk';

const activitiesPlanner = subAgent({
  id: 'activities-planner',
  name: 'Activities Planner',
  description: 'Plans activities based on weather conditions',
  prompt: 'Help users plan activities based on current weather.',
  skills: () => [
    { id: 'weather-safety-guardrails', index: 0 },
    { id: 'structured-itinerary-responses', index: 1 },
  ],
});

Skill Reference Options

Field	Type	Required	Description
`id`	string	Yes	The skill identifier (matches the `name` in SKILL.md).
`index`	number	No	Controls ordering. Higher index = higher priority in the prompt.
`alwaysLoaded`	boolean	No	If `true`, skill content is always included. If `false` (default), loaded on demand.

Attaching Skills

Define attached skills to subagents:

const mySubAgent = subAgent({
  id: 'my-sub-agent',
  name: 'My Sub Agent',
  prompt: 'You are a helpful assistant.',
  skills: () => [
    {
      id: 'inline-skill',
      index: 0,
      alwaysLoaded: true,
    },
  ],
});

Always Loaded vs On-Demand Skills

Skills can be loaded in two ways:

Always Loaded

Set alwaysLoaded: true to include the skill content in every prompt:

skills: () => [
  { id: 'brand-guidelines', alwaysLoaded: true },
]

Use for skills that apply to every interaction (brand voice, compliance rules).

On-Demand (Default)

When alwaysLoaded is false or omitted, skills appear as an outline in the system prompt. The agent can load the full content when needed using the built-in load_skill tool.

skills: () => [
  { id: 'technical-troubleshooting' }, // Loaded on demand
]

Use for specialized instructions that only apply to certain requests.

Example

A complete example with skills directory structure and usage:

activities-planner-advanced/
├── index.ts
├── agents/
│   └── activities-planner-advanced.ts
└── skills/
    ├── weather-safety-guardrails/
    │   └── SKILL.md
    └── structured-itinerary-responses/
        └── SKILL.md

index.ts

import path from 'node:path';
import { loadSkills, project } from '@inkeep/agents-sdk';
import { activitiesPlannerAgent } from './agents/activities-planner-advanced';

export const activitiesPlannerAdvanced = project({
  id: 'activities-planner-advanced',
  name: 'Activities Planner Advanced',
  description: 'Plans activities based on weather forecasts',
  agents: () => [activitiesPlannerAgent],
  skills: () => loadSkills(path.join('activities-planner-advanced/skills')),
  models: {
    base: {
      model: 'openai/gpt-4o-mini',
    },
  },
});

agents/activities-planner-advanced.ts

import { subAgent, agent } from '@inkeep/agents-sdk';

const activitiesPlanner = subAgent({
  id: 'activities-planner',
  name: 'Activities Planner',
  description: 'Plans activities considering weather conditions',
  prompt: 'Help users find activities based on weather forecasts.',
  skills: () => [
    { id: 'weather-safety-guardrails', index: 0 },
    { id: 'structured-itinerary-responses', index: 1 },
  ],
});

export const activitiesPlannerAgent = agent({
  id: 'activities-planner-advanced',
  name: 'Activities Planner Advanced',
  defaultSubAgent: activitiesPlanner,
  subAgents: () => [activitiesPlanner],
});

skills/weather-safety-guardrails/SKILL.md

---
name: "weather-safety-guardrails"
description: "Safety rules for weather-dependent recommendations"
metadata:
  author: acme-corp
  version: "1.0"
---

- Never recommend outdoor activities during severe weather alerts.
- Always mention indoor alternatives when precipitation exceeds 50%.
- Include safety disclaimers for extreme temperature conditions.

skills/structured-itinerary-responses/SKILL.md

---
name: "structured-itinerary-responses"
description: "Present time-aware itineraries with clear actions and citations"
metadata:
  author: acme-corp
  version: "1.0"
---

- Start with a one-sentence summary tailored to the user's location and weather window.
- Provide 3–5 time-blocked suggestions with clear actions, durations, and required items.
- End with a concise checklist plus source references when web results influence recommendations.

On this page