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

​
Common jobs

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.
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.
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.
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.
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
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

​
When Speedtrain is the wrong tool

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.
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:
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%.