Many translators are tempted to translate an entire project inside one long AI dialogue.
By “AI dialogue,” I mean an ongoing conversation/session with the AI. A related concept is the “context window,” which is the material inside that dialogue that the AI must consider when producing the next translation: instructions, glossary terms, corrections, examples, previous source text, previous target text, and so on.

In chatbots, as the AI dialogue gets longer or older, older material is forgotten and drops out of the context window. In other words, the AI loses earlier instructions or terminology. That means you can just keep on going forever inside a single AI dialogue, but you’ll have to keep feeding it relevant instructions as you go. The AI loses focus and you can’t effectively iterate for optimal output as a professional.

CotranslatorAI behaves differently: it keeps the entire AI dialogue available as you stay inside the same context window. That is useful because the AI does not forget earlier material in that dialogue, but it also means that the active context keeps growing. You can see this happening by watching the numbers rise in the token counter on the status bar at the bottom of the main window.

The problem is not only cost. Token costs are usually manageable with the right model. The bigger issue is focus. As the AI dialogue becomes longer and longer, the AI has more and more information to weigh. If the dialogue contains old instructions, changed terminology, superseded glossary entries, corrections, and ad hoc notes, the AI may become less predictable. The output can gradually become less consistent or start reflecting outdated instructions, even though the AI has access to the entire dialogue.

That is why I would not recommend using one very long AI dialogue for an entire large project. Instead, I would use a strong anchor prompt and start a new AI dialogue when the batch, document type, or terminology set changes.

An anchor prompt is the stable instruction set for a specific document, batch, or project phase. It should contain the relevant client instructions, the relevant terminology, any batch-specific preferences, and ideally a small set of carefully reviewed source-and-target examples from the current document. Those examples act as a “gold standard” and help the AI reproduce the right style and terminology. But by focusing an anchor prompt only on relevant resources, we avoid bogging the AI down and are able to sharpen its output.

So rather than repeatedly feeding terms into the ad hoc field during a long dialogue, I suggest that you update the anchor prompt for the current batch and then start a new AI dialogue from the Segment field. That gives the AI a clean, focused context while still keeping the output consistent. And every time there’s a change (for example, your client updates instructions), just update your anchor prompt and start a new AI dialogue after that.

This is the basic template for the anchor prompt that I use:

You are an expert [topic] translator of [source] [document types] to [target]. I need your help translating [document title] to [target].

I will provide you with [source] text to translate to [target]. Correctly translate it, matching the translation as closely as possible to the following requirements.

Here are some very important general instructions:
– Keep each translated segment on its own line in order to match the source segments. Do not combine translations from multiple segments into sentences or paragraphs or add empty lines between translated segments.
– Do not capitalize each word on a line; use sentence case. In other words, capitalize the first word and put a period at the end of every sentence.
– Provide no explanations or labels; just translate.
– Place tags in the right places (ex. {MQ}, [2], etc.)

Here is a glossary of key terms to follow.
[Add only terms relevant to the current batch]

Here is some related [source] content to refer to for context:
[Include relevant text from the PDF itself as context, especially if you’re working in a CAT tool.]

Here are some previous translated segments to refer to:
[Add your “gold standard” source and target translations]

Translate this:

You can see that it includes spots for key terms, document context from your PDF (especially if you’re working inside a CAT tool), and your gold standard translation.
If different batches need different terminology or instructions, you can maintain several anchor prompts for the same overall project and use the relevant one each time. That way you are not trapped in one enormous dialogue, and you do not have to keep patching the AI’s behavior through ad hoc notes. You update the anchor prompt, start a clean AI dialogue, and let the AI work from a focused, curated instruction set.