Managing Skills

Skills extend what agents can do. This guide covers how to manage them — from understanding automatic resolution to writing your own.

How skill resolution works

When an agent plans a task, Exponential automatically detects which skills are needed and resolves them. You don't have to manually assign skills to tasks.

The resolution order

1. Certified skills — checked first. These are verified by the Exponential team and auto-resolve without confirmation.
2. Local skills — checked second. Skills you've created and stored on your machine.
3. Community skills — checked last. Shared by other users; require your confirmation before use.

The first match wins. If a certified skill covers the need, local and community skills for the same capability are skipped.

What resolution looks like

When an agent submits a plan, you'll see a Resolved Skills card showing which skills were detected:

Each skill shows:

• Name and description
• Trust tier — certified, local, or community
• Source — where it comes from

For community skills, you'll see a confirmation prompt before the agent can use them.

Installing community skills

Community skills are discovered through the skill browser:

1. Go to Settings → Skills.
2. Browse available community skills.
3. Click Install on the skill you want.
4. The skill is downloaded to your machine and available for future resolution.

Installed community skills still show a confirmation prompt the first time an agent uses them on a task — this is a trust check, not a bug.

Writing your own skills

Basic structure

Create a SKILL.md file:

---
name: Database Migration
description: Creates and runs database migrations safely
tools: [bash, read, write]
tags: [database, migration, sql]
---

## Overview
This skill handles database schema migrations using our project's migration tool.

## Steps

1. Read the current schema from `prisma/schema.prisma`
2. Understand the requested change
3. Create a new migration file using `npx prisma migrate dev --name <description>`
4. Verify the migration SQL is correct
5. Run the migration against the development database
6. Update any affected TypeScript types

## Rules

- Never modify existing migration files — always create new ones
- Always back up the database before destructive migrations (DROP, ALTER)
- Test migrations on a local database first
- Include both up and down migrations when possible

## Examples

### Adding a new column

When asked to add a field to a model:

1. Add the field to `schema.prisma`
2. Run `npx prisma migrate dev --name add_<field>_to_<model>`
3. Update any queries that should include the new field
4. Update TypeScript interfaces if they exist

Where to save it

Note: Each skill lives in its own directory. The directory name is the skill's identifier.

Frontmatter fields

Writing effective skill content

Be procedural. Skills are instructions, not descriptions. Write steps the agent should follow.

Be specific. "Run the migration tool" is vague. "Run npx prisma migrate dev --name <description>" is actionable.

Include rules. What should the agent always do? What should it never do? Rules prevent mistakes.

Include examples. Show concrete scenarios with expected inputs and outputs. Agents learn from examples.

Minimum length: Skill content should be at least 100 words. If your skill is shorter than that, it's probably not detailed enough to be useful.

MCP servers

For skills that need external capabilities beyond file operations, you can connect MCP (Model Context Protocol) servers. MCP servers give agents access to external tools and APIs.

Common MCP server use cases:

• GitHub API — create PRs, read issues, manage repositories
• Database access — query production data, run migrations
• External services — Slack, Linear, Notion, etc.

MCP server configuration is managed at the system level. See the MCP Tools reference for details on the built-in tools available to all agents.

Tips

1. Start simple — write a skill for one specific workflow, not a general-purpose Swiss Army knife.
2. Test with real tasks — create a task that would need the skill and see if the agent resolves and uses it correctly.
3. Iterate on rules — if agents misuse the skill, add more specific rules and constraints.
4. Keep examples current — outdated examples lead to outdated output.