Custom prompts

Reusable instruction templates that steer the AI's tone, register, and formatting — shared with your organization or kept private.

A custom prompt is a saved instruction the AI receives alongside your source text. Where the glossary pins down terminology, a custom prompt shapes style: tone, register, and formatting rules.

Examples that work well:

"Translate in a formal, business register. Use the polite second person (Sie) in German, vykání in Czech."
"This is product copy for a children's app. Keep sentences short and use vocabulary appropriate for ages 6–10."
"Preserve all Markdown formatting (**bold**, [links](...)) exactly. Do not translate text inside code blocks or placeholders like {{name}}."

A custom prompt has

Field	What it does
Name	Shown in the prompt picker (e.g. "Formal Czech").
Prompt	The instruction text the AI receives.
Private	Off = every member of your organization can see and use it. On = only you can. The dashboard notes: "Private prompts cannot be seen or used by other organization members."

Creating and managing prompts

Prompts are created from the Manual translation page: open the custom prompt dropdown and choose to create a new one. The dialog asks for the name, the prompt text, and the private flag. You can edit or delete prompts from the same picker.

That page doubles as the prompt-development loop: select the prompt, paste representative text, and watch the output change as you refine the wording — the translation re-runs whenever you switch prompts.

Where prompts apply

Custom prompts are applied on the manual translation flow — the dashboard page and the underlying POST /translations/translate API endpoint (via its custom prompt parameter).

Batch auto-translate and the single-cell sparkle suggestions currently use the built-in translation policy plus your glossary; they don't take a custom prompt. If a batch needs consistent terminology, put it in the glossary — that is applied everywhere.

Writing prompts that work

Keep them short. A few sentences. The AI also receives a fixed instruction set (preserve meaning, tone, formatting); your prompt is for what's specific to you.
Be concrete about what to preserve. "Don't translate placeholders like {{name}}" is more reliable than "preserve formatting".
One prompt per use case. Formal support replies, playful marketing, legal text — three prompts, not one mega-prompt trying to cover all three. Pick per translation.
Terminology belongs in the glossary, not in prompts. A prompt saying "always translate Dashboard as Übersicht" works only where that prompt is applied; a glossary entry applies to every AI translation in the organization.

Behaviour and edge cases

Privacy is enforced at use time. A private prompt can be seen, used, edited, and deleted only by its creator; others in the organization get a permission error even with a direct reference.
Deleting a prompt doesn't touch any translation created with it — translations stand on their own once saved.
Editing a prompt's text affects future translations only; nothing is re-run.
Creation, updates, and deletion are recorded in the audit log (custom_prompt.created / custom_prompt.updated / custom_prompt.deleted).

Manual translation — where prompts are picked and developed.
Glossaries — the terminology counterpart, applied to every AI call.