> ## 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.

# Use cases

> The kinds of work Speedtrain is built for, and the ones it is a poor fit for.

Speedtrain suits any job where the same content operation has to be applied to many records, consistently, with the result verifiable.

## Common jobs

<AccordionGroup>
  <Accordion title="Product content enrichment" icon="tag">
    Generate descriptions, bullet points, and marketing copy from structured product attributes.

    Typical setup: a **Brand** variant carrying voice and terminology, a **Language** variant per market, a JSON Schema rule enforcing field lengths, and a forbidden-terms rule blocking claims your legal team will not sign off on.
  </Accordion>

  <Accordion title="Feed compliance" icon="check-double">
    Rewrite catalogue content so it satisfies a channel's content policy — character limits, banned phrasing, required attributes, no HTML.

    Most of this job is deterministic. A JSON Schema rule with `maxLength` and `pattern` constraints does the heavy lifting; AI checks cover only the judgment calls.
  </Accordion>

  <Accordion title="Classification and attribute extraction" icon="tags">
    Assign category codes, extract structured attributes from free text, or normalise inconsistent values across suppliers.

    Constrain the output with an `enum` in the schema rule so the model cannot invent a category that does not exist in your taxonomy.
  </Accordion>

  <Accordion title="Translation and localisation" icon="language">
    Produce per-market content from one source, with per-language rules — formality, terminology, forbidden contractions.

    Use a **Language** variant and scope the rules that genuinely differ per language rather than writing near-identical global rules.
  </Accordion>

  <Accordion title="Image-grounded content" icon="image">
    Let a vision model see the product photos and describe what is actually there — colour, pattern, cut — instead of guessing from a URL.

    The generator and the checker are separate calls, so one model can write the description while a different model independently verifies it against the picture. [Capabilities](/speedtrain/key-concepts/capabilities)
  </Accordion>

  <Accordion title="Chained pipelines" icon="link">
    Feed one task's approved output into the next: extract attributes in stage one, write copy from those attributes in stage two.

    Stage two reads the first task's stage output rather than the raw index. [Content sources](/speedtrain/key-concepts/content-sources)
  </Accordion>
</AccordionGroup>

## When Speedtrain is the wrong tool

<Warning>
  Speedtrain is a poor fit for one-off or highly bespoke content. The platform's value is in repetition and verification — a job run once cannot amortise the cost of writing the quality rules that make it trustworthy.
</Warning>

It is also the wrong tool when the criteria for a good output cannot be written down. If you cannot express what "correct" means as a schema, a term list, or a question a reviewer could answer consistently, no amount of prompt work will make the results dependable.

## Sizing a job

Three numbers decide whether a job is worth automating:

| Number               | Why it matters                                                                                                                        |
| -------------------- | ------------------------------------------------------------------------------------------------------------------------------------- |
| Records in scope     | Below a few hundred, manual work may be cheaper than authoring and tuning the configuration.                                          |
| Variant combinations | Each combination produces its own output. Two languages and two audiences means four generations per record, and four times the cost. |
| Approval rate        | The share of items the gate passes. This decides how much human time the job actually consumes.                                       |

<Tip>
  Measure the approval rate on a preview session before committing to a volume. A configuration that passes 90% of items is a different proposition from one that passes 40%.
</Tip>