Docs The Studio
Work with the agent
What to ask for, what it can do, and how to stay in control.
The agent is an AI assistant that edits the design with the same tools you have. Type in the box at the bottom of its panel and press Enter (Shift+Enter for a new line).
What to ask for
| You want | Say something like |
|---|---|
| A whole carousel | "Make a 5-slide carousel about 5 mistakes when setting up a company in Dubai" |
| A single post | "A bold quote post that says: done is better than perfect" |
| A change to one thing | Select the layer first, then: "Make this shorter and bolder" |
| More slides | "Add a closing slide with a call to action" |
| It on brand | "Apply the brand to every page" |
| Other sizes | "Make a Story and a square version" |
| Captions | "Write captions for Instagram, LinkedIn and X" |
| A picture | "Find a stock photo of a team meeting for the cover" or "Generate an image of…" |
| A check | "Check this design for problems" |
| A schedule | "Schedule this for Tuesday at 9am" |
Getting better results
- Give it the substance. "A carousel about our new pricing" gets generic slides. "A carousel about our new pricing: Solo is now $29, Pro adds approvals, annual saves two months" gets a real one.
- Select first when you mean one thing. The chip above the box shows what the agent can see as selected.
- Ask for one change at a time when refining. Each request is one undo step, so small requests are easy to take back.
- Say the number of slides and the platform if you care about them.
While it works
- The canvas is locked ("Working… the canvas is locked") so you and the agent are never editing two different versions.
- Each step appears as a card: "Created 5 slides", "Wrote captions". A cursor labelled Soda agent shows where it is working.
- Stop ends the request. What it had already done stays, and you can undo it.
Taking things back
- ⌘Z undoes the agent's last request as one step.
- Each step card has Undo this step and Show changes (which highlights the layers it touched).
- If a step cannot be undone on its own because later edits depend on it, the card says to use ⌘Z to step back instead.
What it can do
- Build pages and whole carousels (2 to 20 slides) from 26 built-in layouts, in your brand colours and fonts.
- Add, change, move, align, restack and delete layers.
- Rewrite text, and shrink it to fit its box.
- Apply the brand kit to a page or the whole design.
- Check a design for readability and brand problems.
- Create versions in other formats.
- Write a separate caption for each platform, within that platform's limits and in your brand voice.
- Search stock photos, generate images and remove a background (where image tools are switched on).
- Look at a rendered picture of the page to judge its own work.
- Propose a schedule for you to confirm.
What it will not do
- Schedule or publish by itself. It prepares a card with the time and accounts. Nothing happens until you choose Confirm schedule (or Confirm and publish). Not now drops it.
- Guess a time. If you do not say when, it asks.
- Send a post for approval or delete a post. Do those yourself.
- Use words on your banned list. Copy containing one is rejected and rewritten.
- Claim results. It has no crystal ball; any forecast is labelled an estimate.
- Upscale, extend or recolour an image. Those are not available yet.
Limits and cost
- A design costs 2, 8 or 15 credits depending on the quality chosen in the corner of the message box (Standard, Pro or Max; the Free plan has Standard) and a quick edit or captions 1, however many steps it takes; the chat shows what each reply used. Generated images (4) and background removal (2) are extra. See AI credits.
- A request can take up to 12 steps. A very large job may stop with "Stopped after the maximum number of steps for one request": just ask it to continue.
- A message can be up to 4,000 characters.
- The design must be saved before a request is sent. If you see 'Wait for "All changes saved" and send again', do exactly that.
Two preferences
Under Settings → Agent you can turn on Follow the agent (the canvas moves to where it is working) and Notify me when the agent finishes (a browser notification if you have moved to another tab). Both are saved in your browser.