AI Context

AI context is free-form guidance you write for LLM-based machine translation. It is injected into the translation prompt so the model knows your product, your voice, and the specifics of each string — instead of guessing.

Note

AI context is read by LLM providers only (OpenAI, Anthropic). Classical neural providers (Google, DeepL, Amazon, Azure) do not accept instructions — see the provider matrix.

Three levels

Context lives at three levels, from broadest to most specific. Each level is optional — set only what earns its keep.

Level	Where to edit	Limit	Use it for
Project	Project Settings → General	8,000 chars	Brand voice, audience, tone, product terminology — everything true for every string in every language
Language	Languages → AI context on a target language	2,000 chars	Per-language rules: formality register (du/Sie, tu/vous), regional conventions, script preferences
Key	Key editor → AI context	2,000 chars	Per-string nuance: where the string appears in the UI, what a placeholder refers to, gender/number constraints

All three are combined into the translation prompt:

• Project and language context are stable for an entire translation run, so they ride the cacheable part of the prompt — generous project context does not multiply your token cost per key.
• Key context is sent per key, only for the keys that have it.

Precedence

More specific wins. When guidance conflicts, key context overrides language context, which overrides project context. The model is told explicitly that project- and language-level guidelines are defaults and per-key context takes priority.

key context  >  language context  >  project context

What to write at each level

Project — the briefing you would give a new human translator:

Product: Comvi, a translation management system for developers.
Audience: professional software developers; technically fluent.
Tone: concise, direct, friendly but not playful. No exclamation marks.
Never translate: product names (Comvi, Comvi i18n), code identifiers,
CLI command names.

Language — only what differs for that target language:

German: use informal "du" throughout. Compound nouns are fine, but
prefer shorter synonyms in buttons and labels.

Key — only what the model cannot know from the string itself:

Button label on the checkout page. "Order" is a verb here, not a noun.

Tip

Don’t repeat yourself across levels, and don’t duplicate what other features already handle: glossary terms are injected automatically, and translation memory supplies prior human translations. Context is for what the model can’t otherwise know.

Key description vs. key AI context

A translation key has both a description and an AI context field. They serve different readers:

• Description — internal notes for your team. Never sent to the model.
• AI context — instructions for the model. Sent with every LLM translation of that key.

If you previously kept translation hints in key descriptions, move the model-facing parts into the key’s AI context field — descriptions no longer influence machine translation.

Editing context

1. Project level

   Go to Project Settings → General and fill in the AI context field.

2. Language level

   Open the Languages page, choose a target language, and edit its AI context in the language’s dialog.

3. Key level

   Open any key in the editor and set its AI context alongside the description and character limit.

Changes apply to the next machine-translation request — existing translations are not re-translated automatically.