AI agent? Open the Agent desk: one form in, finished prompt out. Or call window.PromptBench.
Prompt Bench Studio
Start here

From the client's words

Paste what they asked for. It sets the master, cast size, lanes and dials the words can justify, keeps what you already chose, and turns the rest into questions.

Explicit choices always win. Undo puts everything back.
Master

What kind of picture

One master per prompt. It is the role the model plays, and it adapts its wording to the size of your cast.

Brief

In your words

Free text, wrapped and placed by the packs. Keep it to what is true; the modules below handle the look.

Printed first

How it is drawn

These lanes go to the top of the prompt, before anybody is described, so a face is drawn in the same hand as the room.

This image

What is in it

Stack as many as you like. Anything that needs more characters than your cast is held back and says so.

Dials

Degrees, not numbers

Each position prints a sentence, shown under the slider exactly as the model will read it.

Guardrails

How the model behaves

The switches that decide whether it asks, how much it may invent and how carefully it counts.

Variation

Controlled chance

A seed picks from the variation pools. Same seed, same prompt, every time.

Agent desk

One form in, finished prompt out. Built for AI agents working from a client's words, and for anyone in a hurry. Same engine as the Studio: what you compose here is what the Studio shows.

How to use it

  1. Paste the client's request, untouched, into Client's words.
  2. Add what you know for certain: the cast, one per line. Leave anything you do not know empty.
  3. Press Compose. Empty lanes are filled only where the words justify it.
  4. If the status says not ready, ask the client the questions listed, add the answers, and compose again.
  5. Give Prompt to the image model with any reference photos. Give Shareable to anyone else.

Scripted? window.PromptBench.compose({ brief, cast, recipe }) returns the same result as JSON. window.PromptBench.manifest() describes every input.

The job

Vocabulary: every master and module installed, with its number and needs