What does the feather button next to the prompt do?

The feather opens LuviTransLex, LUVI's prompt check: it lists what the selected model is known to mishandle in your prompt and fixes it in one click, and can optionally rewrite the whole prompt in that model's dialect — free of charge, and never without your say-so.

Last updated:

Where it is

In the header of the Prompt section, next to the colour-chip and prompt-history buttons. It lights up for models that have a prompting guide and pulses until you open it once.

Two things it does

Corrections. The guide's hard rules run on your text as you type: a forbidden word for this model, a wrong reference syntax, a length over the cap, a missing element the model needs. Each issue is listed with what would change. Apply n corrections applies them all; nothing changes until you press it.

Rewrite. Behind an experimental switch, Rewrite with Haiku restates your whole prompt in the model's dialect — camera first for one family, timecoded shots for another, audio channels named the way that model reads them. You see the proposed text side by side with yours and choose Apply to prompt or discard. Your original is never replaced silently.

Limits

  • Prompts up to 4,000 characters.
  • Up to 30 rewrites a minute per account.
  • The rewrite keeps your intent, entities and quoted lines; it does not invent content.

Models without a guide

The feather still opens, but with nothing to check the panel says so. Those models simply have no known traps recorded yet; the model page shows which families have a guide.

The same rules elsewhere

Claude and ChatGPT apply the same checks through the connector before generating, and workflow prompt nodes are linted with them before a run.

Frequently asked

Why is the feather pulsing?

The model you selected has a prompting guide and the panel is closed. Open it to see the checks.

Does the rewrite cost credits?

No. Only generations are charged.

I wrote the prompt in Turkish. What comes out?

The rewrite is in English, because that is what the models respond to best; names and quoted lines are kept as you wrote them, and a spoken line stays in its language.

More questions