Quick Start

This page has one job: help you complete one real workflow with the shortest possible path.

!!! warning Prompt Optimizer does not include a ready-to-use model by default. Before your first run, configure at least one text model in Model Management, or analysis, optimization, testing, and evaluation will not work.

First-time users: only do these 4 steps

1. Configure one text model in Model Management
2. Open the workspace that matches your input structure
3. Run one Optimize or Analyze action on the left
4. Run Test / Result Evaluation / Compare Evaluation on the right

If you are not sure where to go next:

• model setup: see Model Management
• workspace choice: see Choose Workspace
• analysis vs evaluation: see Testing & Evaluation

Step 1: configure the minimum model setup

• text workspaces: at least 1 text model
• image workspaces: at least 1 text model + 1 image model
• multi-image generation: also make sure the right-side image model supports multiple input images and prepare at least two images

For most first-time users, the simplest path is:

1. configure one text model
2. run left-side analysis or optimization
3. run right-side testing
4. run one evaluation

Once that works, add a second text model only if you actually want side-by-side comparison.

See Model Management and Model Testing Strategy for more detail.

Step 2: choose the right workspace

Step 3: understand left vs right

This is the most important product boundary in the app:

• left side edits prompts
• right side runs real outputs
• right-side evaluation judges real execution evidence

For the full boundary, see Testing & Evaluation.

Step 4: run one minimum workflow

In any text workspace, you can start with this sequence:

1. enter the original content on the left
2. click Optimize or Analyze on the left
3. move the result into the lower workspace area
4. choose original / workspace / vN on the right
5. run one or more tests
6. start with Result Evaluation
7. then use Compare Evaluation if you have multiple columns
8. apply valuable suggestions back to the left workspace

If model connection fails, then check deployment and connection environment

!!! warning The web version is not a built-in proxy layer. The browser still sends requests directly to the model service you configure.

Three small examples

Example A: System Prompt Workspace

Left side:

You are a technical support assistant. Answer user questions clearly.

Right-side test message:

My app crashes on Windows at launch. What should I check first?

Example B: User Prompt Workspace

Left side:

Write a poem about autumn.

No extra test text is needed on the right in this workspace. The user prompt itself is the tested object.

Example C: Text-to-Image Workspace

Left side:

An orange cat standing on a neon street in the rain, cinematic mood, rich detail, low saturation.

Then:

1. choose a text model for left-side prompt work
2. choose an image model on the right
3. compare prompt versions through real generated images

Example D: Multi-Image Workspace

Open /#/image/multiimage.

Upload at least two images, for example:

• image 1: the character reference
• image 2: the outfit or style reference

Left side:

Use image 1 as the main character identity. Preserve the outfit style and material feel from image 2. Generate one cohesive new image instead of mechanically stitching the references together.

Then:

1. drag the cards to confirm the meaning of image 1 / image 2
2. remove the wrong image with the top-right X
3. click Optimize on the left
4. choose multi-image-capable image models on the right
5. keep one model fixed and compare original / workspace / vN
6. then keep the prompt version fixed and compare image models

Common mistakes

“Web deployment” means there is a proxy layer

No. The browser still connects directly to your configured model service.

Why can’t I connect to my local model?

The most common reason is browser-side CORS or mixed-content restrictions. In those cases, the desktop app is usually a better fit.

Will clearing browser data erase my setup?

Yes. Browser-side data is stored locally in the browser. Export from data management before clearing it.