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

# AI specialists

> The role the model plays, the model it runs on, and what belongs in a specialist prompt.

The **specialist** is who the AI is for this job: a role prompt plus the model that runs it.

## What it contains

| Part   | Meaning                                                                    |
| ------ | -------------------------------------------------------------------------- |
| Alias  | A short display name, such as *Product content writer*.                    |
| Model  | The deployment model the generation call runs on.                          |
| Prompt | The role prompt, which goes first and verbatim into the generation prompt. |

## Deployment models

A specialist references a **deployment model**, not a model name. The same underlying model can exist as several deployments with different settings, so the deployment is what you pick. Available deployments are managed in tenant settings.

You can override inference settings for one configuration without touching the shared deployment, which makes a configuration the safe place to experiment with them.

<Info>
  Quality rules that use AI evaluation carry their **own** model, separately from the specialist. The model that writes the content does not have to be the model that checks it.
</Info>

## What belongs in the prompt

The specialist prompt describes role and behaviour:

* Who the AI is and what expertise it brings
* A short numbered list of behavioural rules — use only the data provided, adapt to the brand's voice, and so on
* How to respond — deliver the content, not explanations of it

## What does not belong in it

<Warning>
  Three things are commonly written into the prompt and should not be:

  * **Output field names and limits** — the JSON Schema rule owns those, and unlike prose it is machine-enforced
  * **Variant-specific guidance** — the variant owns that, and only the variant reaches the quality checks
  * **Quality criteria** — the rules own those, and they are already injected into the prompt automatically
</Warning>

Repeating any of them creates two sources of truth that drift apart. When migrating an older configuration, a fifty-line prompt usually shrinks to ten or fifteen lines.

## Why it stays short

Every rule you write is already shown to the model before generation, so it knows what it will be judged on. The prompt does not need to restate them — it needs to establish who is writing. [Where each requirement belongs](/speedtrain/authoring/placement)