Iterating

After the first generation, most of your time goes into iteration: nudging dimensions, swapping materials, fixing things the agent did not quite get right. The cheapest path is usually not a chat message.

The iteration hierarchy

In order of cost (free → expensive):

1. Parameter sliders — 0 credits, instant.
2. Undo / Redo — 0 credits, instant.
3. Checkpoint restore — 0 credits, instant.
4. Iteration prompt — 15 credits, 10–30 s.
5. Full re-generation — 30 credits, 20–60 s.

Try them in that order.

1. Sliders first

If the change is "wider," "taller," "more drawers," "different finish," "open the door," or "show the closed state" — there is almost certainly a slider already. Drag it. See Parameters for the full reference.

If the change you want does not have a slider, ask for one:

Add a slider for the drawer opening from 0 to 15 cm.

The agent adds the parameter in one iteration (15 credits) and from then on the change is free.

2. Undo / Redo

Top-left of the viewport: the curved arrows. Single-step undo of the last geometry edit. Use it when:

• The agent's last iteration made it worse and you want it back to one second ago.
• You toggled an enum and want to compare.
• You want to A/B between two versions before committing further changes.

Undo is local — it does not refund the credits already spent on the iteration. It just reverts the state.

3. Checkpoints

The History icon in the header opens the Checkpoint Panel. Every successful build creates an auto-checkpoint. You can:

• Restore — replace the current model with the checkpoint's geometry.
• Pin / Name — keep this checkpoint past the auto-retention limit (20 newest).
• Diff — see what changed between checkpoints.

Use checkpoints when:

• You want to try a wild experiment but keep the current version safe.
• You want to compare two distinctly different directions side-by-side.
• The agent has drifted over many iterations and you want to go back to a known-good baseline.

Pin before risky changes

Before asking the agent for a major restructure ("rebuild this as three modules instead of one"), pin the current state as a named checkpoint. If the rebuild lands you in a worse place, restore is one click.

4. Iteration prompts

When sliders and undo cannot get you there, ask. A few principles:

Name the part

Use the names the agent used. Look at the chat reply or the parameter panel to find them:

• ✅ "Move the left door hinge to the other side."
• ✅ "Make the top shelf 2 cm thicker."
• ❌ "Make that thing a bit different."

One change per message

Easier to evaluate, easier for the agent to apply cleanly, easier to undo if it goes wrong.

• ✅ "Make the toe-kick 5 cm taller."
• ❌ "Make the toe-kick taller, change the wood to walnut, add a back panel, and move the handles."

Use numbers when you can

Vague modifiers are easy to misread:

• ✅ "Make the desk 2 cm shorter (so 73 cm)."
• ❌ "A tiny bit shorter."

Use the screenshot-and-fix loop

For changes you can see but not describe (proportions, alignment, balance), ask the agent to take a screenshot from the relevant angle and then describe what you want changed:

Take a screenshot from the front. The handles look too low — move them up so they sit ~5 cm from the top of the drawer face.

The agent will take a screenshot, look at the result, and then make the edit. This loop is great for getting alignment right.

5. Full regeneration

Sometimes the model has drifted too far and the cleanest fix is to start fresh. Two ways:

• Same project, new build — type a fresh prompt as if the project were empty. The agent will write a new geometry block from scratch. Costs 30 credits.
• New project — Projects → New Project. Cleanest slate. Also useful if you want to keep the original around for reference.

If you find yourself doing this frequently, the upstream cause is usually that the initial prompt was under-specified. Lock in dimensions and style in the first prompt and you will iterate less.

What an "expensive" iteration looks like

Some changes are deceptively expensive. Watch for these:

• Asking the agent to do math you could express as a slider. Adding a slider once, then dragging it many times, is dramatically cheaper than asking the agent to nudge dimensions in chat.
• Long conversation history. After 30+ messages on the same project, every iteration pays more in input tokens. If you are deep in a session and a single iteration feels slow, consider starting a fresh project from the current checkpoint.
• Repeated near-misses. If two iterations in a row landed close but not quite, the third is usually a worse spend than restoring the previous checkpoint and iterating from there.

Stopping a runaway iteration

If the agent is in the middle of a turn that you can tell is going wrong, hit the Stop button (the square icon next to Send). The chat turn aborts immediately. You are charged only for what was generated up to that point — there is no refund for partial output, but you are also not charged the full 15.

Next: Materials & style — how the agent uses the material library, and how to override it.