How to brief an AI tool so the work comes back right

This site has two jobs that run through a model: translating content into ten languages, and generating the hero images for these guides. Both work now, and neither worked at first. What is left is five practical rules.

One job, and a defined output shape

A request framed as a single task with a clear response structure comes back correct far more often than one that explains context and hopes. In the translation, for example, the request is for an object with exactly the same keys, not "translate this article".

Do not show it what it can break

This is the rule that saved us the most. Content here is stored as structure, with block types, keys and links. Hand a model that structure and ask for a translation, and it returns translated text with a damaged structure: a key gone, a link altered, an order reversed.

What you do instead is extract only the sentences into a flat list, translate that, and write the result back into the original structure. The model never sees the technical fields, so it cannot touch them. Link addresses and service names stay outside too.

Check the answer with the same validation as everything else

Text coming back from the model goes through the same validation any content in the system goes through, before it is stored. If it does not pass, that language fails and is logged, and the rest carry on. A model's answer is unvalidated input, exactly like a form somebody filled in.

Make re-running cheap

The translation here keeps a signature of the source text. If the text has not changed, that language is skipped and costs nothing. This sounds like an operational detail and it changes behaviour: when re-running is free, you check and correct without thinking twice.

And what stays with a person

Deciding whether the result is any good. With the images, four of fourteen were rejected and re-rolled. There is no automatic check that says "this looks like a photograph of a poster", so somebody looked. That is the part that does not get handed over.

We wrote about the failure mode itself in where AI fails.

Automation built with validation rather than hope