> ## 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.

# Where each requirement belongs

> Every constraint lives in exactly one place. This page says which place, for every kind of requirement.

The single most common way a task configuration goes wrong is **duplication**: the same limit stated in the specialist prompt, again in special instructions, and again in a quality rule. They drift apart, and only the rule is enforced.

<Warning>
  Every constraint lives in exactly **one** place.
</Warning>

## The placement table

| You have…                                                                      | It goes in…                                           | Why                                                                             |
| ------------------------------------------------------------------------------ | ----------------------------------------------------- | ------------------------------------------------------------------------------- |
| The AI's role, expertise, and approach                                         | The specialist prompt                                 | The only free-form slot, and it comes first in the generation prompt.           |
| Output field names and types                                                   | A JSON Schema rule                                    | Injected as guidance **and** machine-enforced.                                  |
| Length limits, required fields, enums, formats                                 | A JSON Schema rule                                    | Deterministic. Models cannot count characters.                                  |
| Exact banned words and phrases                                                 | A forbidden-terms rule                                | Deterministic, with no misses.                                                  |
| Context-dependent judgment — tone, naturalness, a word used only as a buzzword | An AI evaluation rule                                 | Needs understanding, not matching.                                              |
| Brand voice, terminology, audience traits                                      | A variant option's fields                             | Reaches the generator **and** the checks, and can be scoped.                    |
| How the AI should use a dimension                                              | That variant's instruction                            | Per-dimension direction. The defaults usually suffice.                          |
| Reference data to look up at runtime                                           | A capability                                          | Live data beats pasted data.                                                    |
| Pictures the model must actually see                                           | The images capability, plus `$images` on the document | A vision model reads the picture; a text model reads a URL and guesses.         |
| Whether the generated text matches the pictures                                | An AI rule that opts into the images capability       | Verification is a separate call, and can use a different model than generation. |
| Generation-only context — campaign notes, edge cases                           | Special instructions                                  | Checks cannot see it, so only put things here that never need checking.         |
| Which data the job runs over                                                   | The session's input binding, not the configuration    | One configuration, many feeds.                                                  |

## The three-question test

For any new requirement, ask in this order:

<Steps>
  <Step title="Must it be enforced?">
    Then it is a quality rule — deterministic if it can be expressed as a schema constraint or a term list, an AI rule if it genuinely needs judgment.
  </Step>

  <Step title="Must it vary by language, brand, or audience?">
    Then it is a variant field, and possibly a scoped rule alongside it.
  </Step>

  <Step title="Is it just context for generation?">
    Then it is a special instruction — and that is a last resort, not a default.
  </Step>
</Steps>

## Why special instructions are last

<Warning>
  Quality checks never see special instructions. Anything written there is invisible to every check, so it can never be verified.
</Warning>

That makes them right for genuinely unverifiable context — a campaign's mood, a one-off note about an unusual batch — and wrong for anything you would be unhappy to see violated.

## Why variants beat prompt text

Variant fields reach the generator **and** every AI check's context. Prompt text reaches only the generator.

So if a brand's tone rules live in the specialist prompt, the brand-voice check has no idea what the tone was supposed to be. Put them in the variant and the checker is working from the same brief as the writer.