Work with the agent
The agent is how anything in your project gets built. You describe what you want in plain language, the agent does the work, and the preview updates so you can judge the result before your customers see it. This page covers the conversation itself. For the loop from request to live app, see How Spark works.
Open the chat
Click the Edit button on any app or internal tool and the Spark panel docks on the right. It stays with you as you move around the project — same conversation, same scroll position, same unsent draft. (For where everything else sits, see the platform tour.)
The chat knows where you’re standing. A pill above the message box names the surface you have open, and when a preview is showing, the page it’s on rides along too — so “make the heading bigger” lands on the right heading without you naming the app or the page. Each sent message keeps a small record of that snapshot (“Marketing site · /pricing”). Take the pill off to have a request treated project-wide.
Ask for outcomes
Describe the result you want and why. Leave the implementation to the agent.
- Vague: “Improve the homepage.”
- Useful: “On the homepage, move the opening hours above the menu — customers keep asking when we’re open.”
The second version gives the agent a goal it can verify and a reason that shapes a hundred small decisions along the way.
Watch it work
While a turn runs, the panel shows a timeline of steps in plain verbs: Editing, Updating the preview, Searching the web, Generating an image, Reading the logs, Looking at the page. Finished steps settle into past tense, so the one still in its running form is where the agent is right now. Between steps, the running row shows the agent’s latest thought in its own words.
When the turn ends, check the result in the preview — and when you’re happy, publish.
One turn at a time
The agent handles one request per turn. Send another message while it works and yours waits its turn, marked Queued with a Cancel beside it; it goes out as soon as the current turn settles.
Stop ends a turn early. The agent finishes the step it’s on, then stands down. Edits already made stay in place, so look at the preview and steer from there.
Answer its questions
Some decisions the agent won’t make for you. When it needs one, it pauses and asks with a card in the chat — a short set of questions, or a change to review. While it waits, that app shows it’s waiting on you, in its sidebar row and its status bar, and nothing on it moves until you answer. Reply in the card and the turn picks back up.
Attach files
Drop a file or image into the message, or use the attach button in the message box’s action row. Attachments are context for the conversation: the agent can read a document, study a screenshot, match a reference. They’re temporary, and they belong to the chat. A file that should live in the app — a photo on a page, a document in the knowledge library — has to be added to the app itself; ask, and the agent copies it in.
Choose a model
The model picker sits under the input, beside the send button. It sets which model builds for you:
| Model | Best for |
|---|---|
| Claude Opus 5 | Most capable — for complex, tricky changes |
| Claude Sonnet 5 | Fast and capable — for everyday edits |
| Claude Haiku 4.5 | Fastest — for small, quick tweaks |
| Claude Opus 4.8 | Previous flagship — for complex, tricky changes |
| Kimi K3 | Moonshot AI’s flagship open-weights model |
Claude Sonnet 5 suits most everyday work. Switch to Claude Opus 5 when a change is intricate or keeps coming back wrong.
What stays yours
Three things the agent leaves to you, by design:
- Publishing. The agent updates the preview and stops there. Taking a version live happens with the Publish button, and only you press it — see Preview and publish.
- Pricing. Prices, subscriptions, and offers are written by the Payments page’s own agent, a separate chat opened from that page. This agent can pass a request along, and the change happens there — see Get paid.
- Project settings. Your business profile and brand, your domain, and the rest of Project settings are yours to edit directly. The agent reads them and builds accordingly.