Take a photo of a page, a whiteboard, a screenshot or handwriting, and Quick Draft reads the text into your draft as Markdown.
Ways to start
- Scan in the capture bar. It opens the camera, then the crop screen. See Capture actions.
- Refine on a photo or image that’s already in the draft. It opens the same crop screen for that image. Videos don’t have Refine.
- The Scan widget. See Widgets.
The crop screen
- Drag the corners to frame the part you want.
- Use the buttons at the top to flip or rotate the image 90°.
- Choose what to do:
| Button | What it does |
|---|---|
| Cancel | Closes the screen without changes. |
| Read text | Recognizes the text in the cropped area and inserts it at the cursor. |
| Crop | Keeps the cropped image without reading text. |
What happens to the image:
| Started from | Read text | Crop |
|---|---|---|
| Scan | Only the text is added. The photo isn’t kept. | The cropped photo is added to the draft. |
| Refine | The text is added and the image stays as it was. | The cropped image replaces the original attachment. |
The result
- The text is inserted as Markdown at the cursor, like a paste. Headings, lists and tables are kept where the model recognizes them.
- If there’s no text in the image, you’ll see No readable text.
- If the request fails (AI usage limit, a problem with your own key, and so on), a cross appears and the crop screen stays open so you can try again or cancel. See Troubleshooting.
Several pages
To scan several pages, scan them one at a time. Each Read text adds its text at the cursor.
If you add several photos at once (for example with Photo or Image), they’re added as attachments. Tap Refine on each one to read its text.
Better results
- Pick the text recognition model in Settings → AI → Models → Text recognition. Mistral OCR models are built for documents and keep tables and layout. See Choosing AI models.
- Change what the model is asked to do with the Text recognition prompt, for example to also translate the text into English. By default it keeps Markdown formatting and the original line breaks. Mistral OCR models don’t use a prompt. See AI prompts.