Set up once, then forget it.
Two minutes to a working setup, whether you are on a cloud key or a model on your own machine.
Install
Re-Prompter needs Chrome 116 or newer, or any Chromium browser of the same vintage — Edge, Brave, Opera, Arc.
Add it to your browser
Install from the Chrome Web Store, or load the unpacked build from chrome://extensions with Developer mode on and Load unpacked pointed at the dist folder.
Open the panel
Click the toolbar icon and choose Open the panel, or press Alt Shift R. The welcome screen appears the first time.
Pin it
The puzzle-piece menu, then the pin. The side panel is the workspace; the popup is only a launcher.
Connect a cloud key
Pick a provider card on the welcome screen, paste the key, press Test and continue. Chrome asks for permission to reach that provider's host at that moment, and not before.
OpenAI — a key from platform.openai.com. Text, vision, image generation and masked editing from one key. The simplest single-provider setup.
OpenRouter — one key, hundreds of models, and the only provider that publishes prices, so cost estimates are real rather than typed in.
Anthropic Claude — the best brief writer here. It does not generate images, so pair it with OpenAI or an image-capable Gemini model.
Google Gemini — strong vision, and image output on its image-capable models. A key from Google AI Studio.
Press Refresh models after saving a key. The model list, its capability badges and any published prices come from the provider itself.
Connect Ollama
Nothing leaves your machine, and the whole text half of Re-Prompter costs nothing to run.
# start the server ollama serve # a text model for briefs, titles and edit interpretation ollama pull llama3.2 # a vision model for reference and style analysis ollama pull llava:13b
Then press Detect on the Ollama card. Chrome asks for permission to reach localhost the first time. Re-Prompter reads the installed list and asks each model's own manifest whether it can see images, so a text-only model is never handed a picture.
Any OpenAI-compatible server works the same way: add it under Providers, give it a base URL such as http://localhost:1234/v1 for LM Studio, and tick what it can do.
Route the tasks
Settings, then Routing. Eight tasks, each with a preferred model and an optional fallback.
Press Route automatically for a sensible default: the strongest writer gets the briefs, a vision model gets the analysis, an image model gets the artwork, and anything local gets the titles. Then change whatever you disagree with.
Automatic provider fallback is off. Leave it off unless you are happy for a failed request to be re-sent to your second provider without asking.
Activate Pro
Settings, then Licence. Paste the key in the shape REPR-LIFE-XXXX-XXXXXXXXXX and press Activate.
One key covers three browsers. Free a seat from the manage page when you move to a new machine — no email needed. Re-Prompter re-checks quietly about once a week and keeps working offline for a fortnight in between.
Your first thumbnail
Say what the video is
One sentence in the Idea box. Add a headline if you have one.
Press Generate
The panel shows each stage as it runs: writing the brief, merging the style, applying persona constraints, building the image prompt, generating, typesetting, saving.
Read the brief
Press Inspect brief. Change any field. Press Generate again and the artwork follows your edit, not the model's first idea. Press Regenerate to throw the brief away and start over.
The headline
Leave Headline rendering on the deterministic overlay whenever the wording matters, which is almost always.
The artwork is generated with no lettering and the space kept clean, then your exact string is typeset on canvas. Headline typography opens the controls: font, size, weight, alignment, position, maximum width, line height, letter spacing, stroke, shadow and a background plate, previewed live over the real image. You can load a custom font file for the current project; it stays in your browser.
Editing
Open any generation and type the change.
Paint a mask first when the change belongs to one object among several. Undo and Redo walk the history tree; Compare shows before and after; Restore jumps to any earlier version. Nothing is overwritten, so branching costs nothing.
If your image model has no edit endpoint, Re-Prompter re-describes the image with the change applied and rebuilds it, and says so. A painted mask cannot survive that path, and it tells you rather than ignoring it silently.
Personas and styles
Both are built once and reused everywhere.
A persona needs three to ten photographs and a confirmation that you have permission to use them. A style needs three to ten references sharing a look, or just a written description. Set a strength, switch off any attribute you do not want, and pick them from the Create screen.
Export
PNG, JPEG or WebP, at the generated size or any preset.
The filename pattern accepts {project}, {date}, {version} and {title}. The Projects screen exports everything as JSON, with or without image data. Your API keys are never in it.
Troubleshooting
"The selected model doesn't accept images." You have routed a vision task to a text-only model. Settings, Routing, Vision analysis — pick one with a vision badge.
An Ollama model answers with nothing. It has no vision projector. Pull a vision model and route the vision tasks to it; Re-Prompter tells you this rather than leaving you guessing.
"The browser could not reach…" The host permission was declined, or the provider blocks browser requests. Re-enable the provider and accept Chrome's prompt.
Nothing answered at localhost:11434. Ollama is not running, or you declined the localhost permission. Run ollama serve and press Detect again.
The image request came back with no image. The prompt was probably refused. Reword it; named real people and violent or explicit detail are the usual triggers.
The panel will not open from the keyboard. Chrome needs a user gesture for the side panel. Click the toolbar icon instead, or check the shortcut at chrome://extensions/shortcuts.