Chatting with the designer
How to talk to the designer agent and what kinds of requests work well.
The chat panel on the left of the editor is a conversation with a designer agent. It has full context about the design (output type, brand assets, snapshot history) and a set of tools for writing code in a sandbox and publishing snapshots.
You don't have to talk to it like a developer. Talk to it like a designer.
What works well
The designer agent is best at:
- Concrete asks — "Make the headline larger", "Swap the hero for something more abstract", "Add a fourth slide on pricing"
- Bigger structural moves — "Reorganize as a three-column layout", "Convert this page into a 5-slide deck"
- Style adjustments — "More minimal", "Tighter spacing", "Use the brand orange as an accent only, not a fill"
- Replacements — "Replace the placeholder image with something darker", "Use Inter instead of system font"
- Constrained explorations — "Show me three header variants" (it will iterate within one design rather than producing three separate snapshots)
It's less good at:
- Pixel-precision — "Move that 4 pixels to the left" — easier to describe the layout intent
- Specifying exact CSS values without context — "Set padding to 24px" works, but "Make it feel less cramped" works better
- Multi-page navigation within a single design — Page output is one page; for app flows, use Prototype
How a turn looks
When you send a message, here's the flow:
- The status indicator switches to Generating with a spinner
- A shimmer overlays the preview while the agent works
- The agent's tool calls appear inline in the chat as small pills (e.g.
get_brand_context,write_sandbox_file,publish_design_snapshot) - The assistant message streams in, explaining what it changed
- Once the new snapshot is published, the iframe reloads with the updated design
- Status switches to Idle with a green dot
The whole loop usually takes 10–60 seconds depending on the size of the change.
The brand-assets protocol
The first time you ask for anything beyond a wireframe, the agent calls get_brand_context. If your design has no brand assets, it will refuse to proceed beyond a sketch and ask you for:
- A logo (URL or uploaded file)
- A palette (hex codes or a brand-style preset)
- Fonts (font family names or links)
You can either provide them inline ("Use #f54e00 as the primary, #1a1a1a as the dark, Inter for headings, IBM Plex Sans for body") or skip the protocol by saying something like "no brand assets yet — just sketch a wireframe".
See Brand assets for the full picture.
Cold-start time
The first message in a design takes longer than later ones — typically 2–3 minutes. That's because the sandbox needs to install dependencies before the agent can run. Subsequent messages in the same session take 5–10 seconds plus generation time.
You'll see a Generating indicator the whole time. Don't worry if it takes a while on the first turn.
Re-prompting and undoing
If a snapshot didn't go the way you wanted, you have two options:
- Tell the agent what to change — "Actually, undo the headline change and keep the previous version"
- Wait for Pin version (phase 4) — bookmarks a known-good snapshot you can roll back to
In v1.0, there's no rollback button. The agent's history of snapshots is preserved, but the active snapshot is always the most recent one. Re-prompting is the cleanest path back to a previous look.
Tools the agent uses
The agent has a small toolkit. You don't call these directly — they fire automatically based on what you ask for:
- get_brand_context — Reads the design's brand assets (logo, palette, fonts, references)
- write_sandbox_file — Writes or updates a file in the design's sandbox
- publish_design_snapshot — Builds the design and uploads the static output to storage
You'll see them as pills in the chat as the agent works. They're useful as a sanity check — if you don't see publish_design_snapshot after a turn, the snapshot didn't update.
Cancelling a turn
There's no cancel button in v1.0. If a turn is taking too long or going the wrong direction, wait for it to finish, then send a follow-up that course-corrects ("undo that and try X instead"). Hard cancellation is on the roadmap.