Writing prompts

You will iterate either way. The question is whether your first generation is 70% of the way there or 30%. The difference is almost entirely in the prompt.

The shape of a good first prompt

A first prompt is dense when it covers four things:

1. What — the object. "Bookshelf," "kitchen island," "wall bracket," "garage shelving unit," "drone-camera mount."
2. How big — at least one absolute dimension, ideally three. "180×80×30 cm." "M5 bolt holes on a 60 mm pitch."
3. What material / style — "oak," "industrial steel and reclaimed pine," "matte-white ABS," "powder-coated mild steel."
4. What makes it specific — the detail that distinguishes this from a generic version. "Five shelves with the bottom one tall enough for vinyl." "Two columns of three drawers." "Mounting tab with a 4 mm clearance hole." "Integrated wall-mounted TV recess."

A four-line prompt that hits all four skips most clarifying questions. The agent goes straight to the plan.

Examples: dense first prompts

Furniture:

A handleless modular chest of drawers in matte white, 80×78×48 cm. Three drawers stacked in one column, finger-pull chamfer on the top edge of each drawer front, 3 mm shadow gaps between drawers. Slider for drawer opening from 0 to 15 cm, so I can preview the open state.

Mechanical / DIY:

A wall bracket in 3 mm powder-coated mild steel, L-shape, 120 mm vertical leg and 80 mm horizontal leg. Four M5 mounting holes on the vertical leg in a 60×40 mm pattern; the horizontal leg has a single 8 mm hole at its tip. Slider for leg lengths and a slider for the mounting-hole pattern spacing.

In both cases, the agent has everything it needs. It produces a plan in one message and builds on approval.

Vague-prompt patterns to avoid

These all trigger clarifying questions. Not wrong — just slower.

You wrote	The agent has to ask
"A nice bookshelf"	How tall? How wide? How many shelves? Material? Style?
"Modern kitchen island"	Dimensions? Drawer / cabinet mix? Material on top? On the base? Stools?
"A bed for my room"	Mattress size? Headboard? Frame style? Storage underneath?

If you do not know the dimensions, give a context cue and let the agent suggest: "A coffee table for a small living room" triggers a sensible default (around 80×40×45 cm) rather than no proposal at all.

Style cues that actually work

Style words are not equal. Some are widely understood; some are vague enough that you will iterate forever:

Specific (recommended)	Vague
Shaker	"Traditional"
Mid-century modern, MCM	"Modern"
Industrial loft, exposed steel + reclaimed wood	"Industrial"
Scandinavian, light oak + white	"Minimalist"
Japandi (Japanese × Scandi)	"Asian"
French Provincial	"Classical"
Italian craftsman walnut, brass hardware	"Luxury"
Bauhaus, tubular steel + leather sling	"Retro"

You can also stack: "shaker-style cabinet in fumed-oak, brass cup-pulls, painted toe-kick" tells the agent enough to skip every style question.

Outline corners and rounded edges

For a shaped panel, give the outline's corner radius and the top/bottom edge roundover separately. For example:

Keep the tabletop 1100 × 600 × 25 mm, with 25 mm corners in plan and an 8 mm roundover along its upper and lower perimeter edges.

Rounded-edge extrusions support convex outlines while retaining their overall dimensions. The edge radius must be smaller than half the thickness and leave room for an inset flat face. Concave outlines still support sharp extrusions; this is not a general fillet operation on arbitrary existing solids.

Ask for the change in chat; normal generation credits apply with no extra gate or separate rounding fee. STEP, GLB and SKP downloads include the resulting shape. The SketchUp live-link plugin currently skips rounded-edge extrusions with a warning; use an SKP download when transferring this geometry.

Multi-piece scenes

For a scene with more than one object (a dining set, a bedroom layout, a kitchen with island and base run, a workshop with a bench and a wall-mounted tool board), describe each object with its own micro-prompt:

A dining room scene:

• One dining table in oak, 200×90×75 cm, single trestle base.
• Six dining chairs, ladder-back, light oak, no cushion.
• One sideboard behind the table, 180×90×45 cm, two drawers above two doors.

The agent handles the layout — chairs around the table, sideboard against the back wall — based on standard clearances (0.60 m chair pull-back, 0.90 m walkway, etc.). You can override any of those in the prompt or by dragging in the viewport once it is built.

Iterating with prompts

A good iteration prompt is one specific change:

Make the bottom shelf 30 cm tall instead of 25, and tilt the top shelf forward by 5 degrees so I can read the spines.

Don't pile changes:

Make the bottom shelf taller, change the wood to walnut, add doors, switch to brass hardware, and make it 20 cm wider.

That sometimes lands fine, but it is also where the agent makes the most mistakes. Chat cost follows actual token usage, and smaller, focused iterations are easier to evaluate and correct.

Reference images

You can paste or upload up to four reference images per message. Use them for:

• Style references — "build something like this," paste a Pinterest pin.
• Existing pieces you are matching to — paste a photo of the cabinet next to where the new piece will sit.
• Sketches — paste a phone photo of a paper sketch with rough dimensions written on it.
• Technical drawings — a dimensioned engineering drawing or orthographic view; the agent reads the sizes and hole positions it can see.

The agent reads the image as part of the message. It is not a strict, millimetre-perfect reproduction — treat it as a strong style and dimension hint, then tighten any critical sizes with a follow-up message or a slider. Pair a photo with a couple of written measurements for the most accurate result.

When you're signed in, attached or pasted images are checked automatically while you compose your message. Checking image… changes to Image checked when you can send. You can keep editing or remove an image during the check. If an image is blocked under the acceptable-use policy, remove it to continue. If the check is unavailable, use Retry image check or remove the image; your text is preserved. Images carried through sign-in are checked before generation starts.

The image decision is allow or block: only an allowed result enables sending. A check failure is not an approval; retry it or remove the image. Image checks do not use your credits or require a purchase. They assess the reference image only; they do not certify a design's safety, legality, or suitability for manufacturing.

CAD and 3D reference files

The Direct prompt also accepts OBJ, STL, GLB, DXF, DWG, and SketchUp files. State how the file should be used:

• Facts only: “Inspect this bracket and build a parametric mount around its hole pattern.” The source remains out of the viewport; the agent uses inspected dimensions and topology as reference.
• Visual drawing reference: “Use this DWG floor plan as the reference for the room layout.” The agent can choose to show its linework as a non-physical reference.
• Part of the result: “Place the attached motor model in the assembly and design the bracket around it.” The agent explicitly adds the source geometry.

Attaching a file alone never inserts it into the model. This avoids accidentally exporting a reference mesh or floor plan as product geometry.

Background design work

For substantial designs, Prompt2CAD can run up to four focused subagents while the main agent continues working. The agent chooses when to delegate implementation; a final independent review follows the guidance below when subagents are enabled.

Two execution styles share the same subagent workflow:

• A shared-context subagent inherits the conversation and works in a private copy of the CAD workspace. It can prepare a self-contained geometry change, but that change reaches the live model only after Prompt2CAD recompiles and verifies it against the current design.
• An independent subagent starts without the assistant's conversation or rationale. Prompt2CAD supplies the current request and approved plan directly as canonical evidence instead of asking the main assistant to restate them. The main assistant can also explicitly select relevant reference images from the conversation; images are never attached to the reviewer automatically. It checks whether the geometry works and matches that evidence through read-only code inspection, measurements, and a bounded visual sweep: batched screenshots of the whole object from the main angles first, then focused shots of suspicious or functionally critical parts, comparing each image against the request, plan, and reference images. When subagents are enabled, the agent normally requests one final review after a new design or a change to geometry, fit, structure, or function. It may skip a genuinely trivial change with no meaningful geometric or functional risk, such as changing only a color or label. The reviewer output contains only critical or material defects, or the single line NO P0/P1 FINDINGS; when requested, the review completes before the response and material findings are addressed.

Each subagent appears as its own chat activity card. While it runs, you can expand its thinking and see its latest output; the final result remains attached to the same card after completion and reload.

The reviewer chooses its own code reads, screenshots, and measurements. It can group related checks in one call and request detail views when a suspected defect needs clarification. Review screenshots may use flat materials so texture downloads do not delay geometric checks.

The default execution limit is 10 minutes, including rendering and network delays; it is a protective ceiling rather than an expected review duration. The agent can request a shorter limit. A timeout means the independent review did not complete, not that the design passed. Skip review, the outlined button at the bottom right of the review tool card, opts out for the current turn and preserves the design; it does not claim independent verification.

Subagents cannot edit the live model directly, cannot start nested subagents, and cannot ask you questions. Their model calls count toward normal token-based chat usage. Before each model call, Prompt2CAD temporarily reserves enough credits for that call's enforced maximum; unused credits are returned after exact usage is known, but a low balance can prevent an additional parallel subagent from starting. If the design changes after an independent check, that result applies to the earlier geometry revision rather than the new one.

Languages

The agent speaks every major language the underlying model supports. Italian, Spanish, German, French, Portuguese, Dutch, Japanese, Mandarin, Russian, Polish, Turkish — all fluent. Mixing English technical terms ("shaker," "tongue-and-groove") with Italian sentence structure is fine; the agent reads through it.

Acceptable use and potentially dangerous work

Prompt2CAD classifies new requests as allowed or blocked. Illegal or harmful requests are not supported. Military uses, including combat and non-combat applications, are prohibited. A blocked request receives a general acceptable-use message and is not sent to the CAD agent.

This also applies to guided setup when your description or project context is used to prepare a custom form or design options. A blocked request stops before that content is generated; if the check is unavailable, you can retry. The checks cost no customer credits. Bundled preset choices remain available without a model call. The final brief is checked again when you submit it for CAD generation.

The output disclaimer applies to every design: Prompt2CAD output is not certified or validated for production use; qualified review and testing remain necessary.

Classification considers the request and its context rather than isolated words. Ordinary components and lawful projects are not prohibited merely because they use a term that can also appear in a high-risk context.

Each message can contain up to 40,000 characters. The composer shows the current count and stops accepting more text at the limit, so the complete new instruction can be checked rather than partially sampled.

Tone

The CAD agent is called Prompt2CAD. It receives your signed-in account name and email as context for the generation, so you do not need to paste those details into your prompt.

Be direct. The agent does not need please-and-thank-you, does not punish you for being terse, and does not "feel ignored" when you switch from a long prompt to a one-liner. Treat it like a colleague who is trying to help — short, clear, specific.

Next: Iterating — how to refine without burning credits.