AI context

View .md

AI context is guidance you write for AI machine translation, injected into the prompt. Only the AI providers (OpenAI, Anthropic Claude) read it; traditional MT providers do not accept instructions. See the provider matrix.

Levels

Level	Where to edit	Limit	Use it for
Project	Settings → General, AI translation context (sent to MT)	8,000 chars	Brand voice, audience, tone, product terminology: everything true for every string
Language	Languages, the target language’s dialog	2,000 chars	Formality register (du/Sie, tu/vous), regional conventions, script preferences
Key	The key editor, next to the description and character limit	2,000 chars	Where the string appears, what a placeholder refers to, gender and number constraints

Every level is optional. Changes apply to the next machine-translation request; existing translations are not re-translated.

Precedence

key context  >  language context  >  project context

More specific wins. The system prompt tells the model that project and language guidelines are defaults and per-key context takes priority.

What to write

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.

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

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

Write the project context as the briefing you would give a new human translator, the language context as only what differs for that language, and the key context as only what the model cannot know from the string. Do not repeat yourself across levels or duplicate what other features handle: glossary terms are injected automatically and translation memory supplies prior human translations.

Description and key context

A key has both Internal notes (not sent to MT) (the Description field in the Create new key dialog) and AI translation context (sent to MT). Notes are for your team and never reach the model; the AI context is sent with every AI translation of that key. Put translation hints in the AI context field.

Collected UI context

The in-context editor plugin reports where each key is rendered in your running app, and Comvi keeps a sanitized profile: the key’s UI role, a length hint, and the text of nearby elements. Rendered values are never stored, only references to other keys in the catalog. The profile is sent with each AI translation of the key in a ## UI context (collected) block, and the key’s context panel in the editor shows the same thing the model sees. Comvi keeps the five most recent observations per key per screen group; nothing needs configuring.

Cost of a long project context

Project and language context sit in the prompt’s system message, which is identical for every key in one run, and key context is sent per key. Both AI providers cache that stable prefix, but only Anthropic’s usage reporting excludes cache reads, so only there does a long project context stop costing full price on every key. On OpenAI the reported prompt tokens include the cached ones, and Comvi meters what the provider reports.

Generate a draft project context

Click Generate description next to the project AI context box on Settings → General. Comvi samples the project’s key names and source texts and asks an AI model for a draft covering domain, audience, tone and terminology.

The project needs at least 30 keys; below that the button is disabled.
The draft is not saved for you. Review it, edit it, and save the field.
Limited to 5 generations per hour per project.
Needs project edit permission and is a dashboard action only; API keys cannot call it.