Improve LLM prompt by providing clearer instructions and additional context

[flodolo]

I'm currently experimenting with a variant of the existing prompt.

Besides a general rewrite, for example to try and solve a problem with missing closing periods, I think there's value in passing the comment and string ID to the LLM to provide additional context.

Other things that might be worth investigating in the future, but not as part of this iteration:

• Test gpt5-mini instead of gpt4.1, across a few languages.
• Store more context, e.g. some info about the project (additional field in the model), and some structured info about files. That could be interesting for Firefox, where specific files (e.g. asrouter.ftl) are more marketing oriented, and might benefit from a looser translation.

[flodolo]

Another interesting thought would be passing terminology matches.

[mathjazz]

There's a lot of good ideas in this issue, but also quite a bit of scope mixed together:

• general rewrite
• prompt changes
• LLM model changes
• adding contextual data (comments, string ID, terminology)
• potentially extending the data model with project/resource-level context

Since much of this overlaps with the Explore additional LLM integrations OKR, it might make sense to step back and define a clearer approach (e.g. a short spec or phased plan), and then break this into smaller, focused issues.

That would help us evaluate impact more systematically, and keep individual changes easier to implement and review.

[flodolo]

I started using this issue to collect ideas, which is not particular useful. The intent is not to tackle all of them here and now. I'll move those to our backlog document.

My goal with this issue is to slightly improve the prompt, and include information we already have (comments, IDs, terminology matches). That is self-contained, it's a small change, and resulting are consistently better than the prompt we currently have, at least for Italian.

I think there is value in making the change now, and use the data we collect to decide if that improved usage or not when we start working expand LLMs a few months from now.

Does that work?