# Quick Start

Once `@maestria/prime-agent` is installed, start Prime Agent from the repository or project you want it to work in, then run `/reload` so the skills and extension metadata are rediscovered.

## Load the orchestrator

Use the orchestrator skill when a task is unfamiliar, cross-file, or risky:

```text
/skill:orchestrator I need to add rate limiting to the API. Plan and implement it.
```

It is methodology guidance, not a replacement for Prime's execution boundary. "Delegate to a specialist" means loading the relevant skill and applying its methodology; the extension does not spawn child agents.

## Call a specialist

Any specialist skill can be invoked the same way; see [Browse the Agents](/core/agents/) for the seven specialist roles.

```text
/skill:adventurer Trace the authentication flow and report the relevant files.
```

## Choose a workflow mode

Load a mode by description matching or explicitly as `/skill:fein`, `/skill:sonar`, or `/skill:blitz`. With the extension loaded, `/fein`, `/sonar`, and `/blitz` also work as slash commands and persist for the session:

| Mode    | Route                                                             |
| ------- | ----------------------------------------------------------------- |
| `fein`  | Full pipeline: reconnaissance, design, implementation, and review |
| `sonar` | Research only; stops before implementation                        |
| `blitz` | Fast route for small or well-understood work                      |

For example:

```text
/fein Add rate limiting to the API.
```

While a mode is active, the extension appends the mode's prompt to the system prompt on every agent turn. `/maestria-status` shows the current mode and the verified/deferred subset; `/mode-clear` returns to neutral routing.

## Review the result independently

After the builder completes, invoke the reviewer in a separate turn:

```text
/skill:reviewer Review the current work for correctness, security, and missing tests.
```

The reviewer skill is advisory. Prime Agent does not provide a skill-level tool-denial mechanism, so read-only roles state their intent without claiming a runtime boundary; use Prime's own execution controls where a hard boundary is required.