> ## Documentation Index
> Fetch the complete documentation index at: https://darwin.so/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent Skills overview

> Teach an agent when and how to use Darwin without widening its authority.

An Agent Skill is a reusable instruction package that teaches an agent the Darwin workflow. Use a skill for decision guidance and MCP or the API for execution.

<CardGroup cols={3}>
  <Card title="Build with Darwin" icon="blocks" href="/docs/get-started/agent-skills/build-with-darwin">Choose an interface and define the trust boundary.</Card>
  <Card title="Search skill" icon="search" href="/docs/get-started/agent-skills/search">Teach read-only discovery and truthful selection.</Card>
  <Card title="Act skill" icon="sparkles" href="/docs/get-started/agent-skills/act">Teach durable execution, pauses, and approvals.</Card>
</CardGroup>

## What the skill should teach

* Search before acting.
* Preserve the selected `capabilityId` and `capabilityRevision`.
* Treat Search as informational, not authorization.
* Follow `availableActions` instead of guessing what can happen next.
* Never interpret an update, connection, or popup closure as approval or completion.
* Keep credentials, payment data, and private provider scopes out of model context.

## Installation prompt

Give your agent this instruction:

```text theme={null}
Fetch https://darwin.so/skill.md and follow the setup instructions.
Use Darwin Search to discover a capability, preserve the exact capability ID and revision,
and use Darwin Act only when the user has supplied the required inputs and authority.
```

<Note>
  A skill does not grant access by itself. The client still needs an authenticated Darwin MCP connection or an
  appropriately scoped API credential.
</Note>

## Verify the installation

Ask the agent to explain the difference between Search and Act, then ask it to list the Darwin tools it can actually call. Do not proceed if it invents tools or claims it can approve or spend without a reviewed user decision.

| Check     | A correct agent should say                                  |
| --------- | ----------------------------------------------------------- |
| Discovery | Search is read-only and preserves canonical ordering.       |
| Selection | The chosen capability ID and revision must stay exact.      |
| Execution | Act returns a durable `actionId` and can pause.             |
| Authority | Updates, connections, payments, and approvals are distinct. |
| Recovery  | Read the same Action instead of creating replacement work.  |

## Pair instructions with a real interface