Chat-driven design

Most CAD tools start with a blank canvas. Prompt2CAD starts with a conversation. You describe the piece, the agent asks what it needs to know, you approve a plan, and the agent writes geometry code that the viewport renders live.

This page is the mental model. If you understand the four stages, every weirdness in the app starts to make sense.

The four stages

1. PROMPT  →  2. PLAN  →  3. BUILD  →  4. REFINE

1. Prompt

You type what you want. The agent reads your message in the context of the entire conversation so far — previous prompts, previous models, previous decisions all carry forward. There is no "reset on every message" model; the conversation is the design.

2. Plan

Before writing any geometry, the agent will typically ask one to three clarifying questions and then show a plan:

• Overall dimensions (width × height × depth).
• A bubble diagram or sketch of the major elements.
• Style direction (Scandinavian, mid-century, shaker, industrial, …).
• Structural notes (back panel, plinth, hardware visibility).

The plan appears as an inline panel in the chat with an Approve button. Approve to proceed, or paste a follow-up to change direction. You can also skip the plan step for trivial follow-ups; the agent uses its judgement.

Why ask before building?

Building geometry usually accounts for most of a new model's token usage. The viewport estimates about 2 minutes to first geometry, although complex designs and retries can take longer. Reviewing the plan can catch a misread before the agent spends more tokens building the wrong geometry.

3. Build

Once the plan is approved (or skipped), the agent assembles the model — laying out the structure, picking materials, and declaring the sliders you will use to adjust it. The viewport builds the scene live as the agent works.

You will see one Write geometry card in the chat covering the build. Click it to expand the details if you're curious; otherwise just watch the model appear.

4. Refine

This is the loop you spend most of your time in. Two ways to refine:

• Sliders — every adjustable knob the agent declared appears in the Parameters panel. Dragging a slider re-renders the model locally with the new value. No credit cost, instant.
• Follow-up prompts — anything sliders cannot cover. The agent either makes a targeted edit to the existing model or rewrites the relevant section from scratch, depending on what your prompt asks for. Each turn is charged from actual token usage rather than at a fixed iteration price.

Iterate freely. The agent retains the design state, the project history, and any past screenshots, so you can say "go back to the version with three drawers" and it will know what you mean.

What the agent can do

The agent has the basics you would expect of a model-building assistant: it can shape the model, manage sliders, change materials, snap the camera to a view, take screenshots from any angle, and check measurements and clearances. It can also search the public web when a design depends on an external fact such as a current manufacturer datasheet, component mounting dimensions, a standard, or a material property. Research is automatic when useful; there is no separate search button. Sources are cited when the agent discusses researched facts, but a search result is reference evidence rather than a dimensional guarantee, so confirm safety-critical and production-critical measurements against the original document.

It can also find groups of generated or imported parts whose bounding volumes might be near each other before focusing a screenshot or running a more precise check. A bounding-volume match is conservative candidate evidence, not an exact surface-distance measurement or by itself proof that two mesh surfaces touch or collide. There is no separate control or gate for web research or scene inspection; both happen inside the generation or iteration turn you already requested. Search-provider fees are not charged as a separate line item, while the model tokens used to process research remain part of the turn's normal usage-based credit cost. Each visible action surfaces in the chat as its own labelled card so you can see what happened.

You almost never need to name an action — describing what you want in plain language is enough. The agent picks the right approach based on your prompt.

Why the agent sometimes "thinks out loud"

The agent occasionally streams a short thinking pass before an action, surfaced as a faded chunk in the chat. You can ignore it — it exists so you can see what the agent is reasoning about if a turn surprises you. It is not billed separately.

What the agent cannot do

A few things are deliberately out of scope:

• It cannot remember between accounts. Each Google account has its own conversation history.
• It cannot access files you have not supplied. It can read images and supported CAD/3D files that you explicitly open or attach to a project, but it cannot browse the rest of your computer.
• It cannot access private or signed-in websites. Web research is limited to public information, and it does not download or execute remote CAD files or code.
• It cannot edit the geometry code while you have an unapproved plan open. Approve or dismiss first.
• It does not produce production drawings. The agent can show a geometry-derived parts list / preliminary BOM with stable IDs, dimensions, quantities, materials, and CSV export, but it does not add tolerance callouts, hardware schedules, edging, grain direction, nesting, or toolpaths.

Practical takeaways

• Be concrete in the prompt. Dimensions, material, style. Vague prompts get clarifying questions; specific prompts skip straight to the plan.
• Approve plans quickly or push back hard. If the plan looks wrong, say so before the more token-intensive geometry build begins.
• Use sliders first, prompts second. Width, height, drawer count, finish enums — try the slider before asking the agent to "make it a bit wider."
• Trust the conversation. You do not need to repeat context every message. The agent already knows what you are building.

Next: Parameters — how sliders actually work, and why most of your iteration should happen there.