Chat Basics

How to talk to SphereHarness so it does what you meant — and how to tell, quickly, whether it did. Read this once before the task pages; most of the difference between a smooth session and a frustrating one is in these habits.

Where you type. SphereHarness lives in the desktop agent beside your CAD window, not inside it: click the floating balloon, press Ctrl+Alt+Space, or use the Open Desktop Agent button on your CAD's SphereHarness ribbon.

Just describe the work

There is no command syntax. Describe what you want the way you would tell a colleague who is sitting at the keyboard. SphereHarness picks the CAD operations, runs them in your open document, and reports each step as it completes.

Create a 40 x 40 x 10 mm block, then add a 6 mm hole through the centre

You can then follow up in the same conversation, and it edits the model it just built rather than starting again:

Make that hole 8 mm instead
A worked conversation showing a prompt and SphereHarness's step-by-step reply
Each step is reported as it finishes, with the values that were actually used.

Units and coordinates

Every number you type is millimetres unless you say otherwise; “20” and “20 mm” mean the same thing. Angles are degrees. Forces are newtons. Say “inch” explicitly if you need imperial.

When a position matters, say which frame you mean. “World” coordinates are the model's global X/Y/Z; sketch coordinates are measured in the sketch's own plane. Being explicit costs four words and saves a rebuild.

Draw the rectangle at world (-115, -20, 11), 40 x 40 mm

Six habits that make SphereHarness reliable

  1. One feature per message on anything complicated. A long plan — twenty features in one prompt — can run out of room part-way through and stop in the middle. Ask for a feature, check it, ask for the next.
  2. Say what you expect the result to be. An expected volume, an expected dimension, an expected number of holes. SphereHarness cannot fake a measured value, so stating the target turns a hopeful instruction into a checkable one.
    Cut the Ø28 hole through both prongs. I expect the volume to drop by about 22,170 mm³ — measure before and after and tell me the actual change.
  3. Ask for the measurement, not the summary. The prose is written after the work; the measurement comes from the CAD system. See Verifying results.
    What does the mass properties tool report as the volume, in cubic millimetres?
  4. Name geometry by where it is, not by what it is called. “The face whose normal is +Z” and “the edge near (−90, 13, 44)” are unambiguous. “Face3” and “the top edge” are not — face numbers belong to one particular solid, and “top” is not the same direction in every CAD system.
  5. Take a checkpoint before anything risky.
    Take a checkpoint of this document before you start

    If the result is wrong, ask to go back to it rather than unpicking the changes by hand.

  6. When something is wrong, report the measurement. “That is wrong” gives SphereHarness nothing to work with; a number tells it exactly what to look for.
    The volume reads 9,000 mm³ but a 40 × 30 × 10 block should be 12,000. Find what is missing.

Watching it work

Steps stream into the chat as they complete, each with the values actually used — not the values you asked for. Read them. A hole reported at Z = 0 when you asked for Z = 20 is visible in the transcript long before it is visible in the model.

For features placed away from the origin, the reply includes placement evidence: the world coordinates the geometry actually landed at. If those are not the numbers you asked for, stop and fix the placement before building anything on top of it.

Screenshots

Take a screenshot of the current model

The image is posted in the chat. SphereHarness can also look at its own screenshot and act on what it sees, which is how it catches drawing layout problems.

Set an isometric view, take a screenshot, and tell me whether the ribs are on the correct side

A full-screen capture including dialogs, ribbon and panels is available too — useful when you are asking for help with something on screen.

Cut the slot, then show me the result from the top

The assistant can ask for a picture of the model right after a step, from the view you name, and look at it before it carries on. It does this when a number cannot tell whether the shape is right, not after every step. Every step that changes the shape also comes back with a measured check (bounding box and volume change). A step that reports success but changes nothing is flagged, so the assistant looks again before going on.

A plan picture before building

Show me an image of what you plan to build, with the main dimensions, and wait for my approval

The assistant draws a plan with top, front, right and isometric views and the overall dimensions, then stops so you can approve or change it. Nothing is built in the CAD until you say so. The picture is shown in the chat and sent as an image on Telegram, Teams and the local API.

What the chat can show

The desktop agent's chat draws more than text. You do not ask for a format — the assistant is told what the chat can display and picks one:

  • Tables for comparisons (materials, options, parts side by side).
  • Charts — bar, line, pie and scatter — for numbers, drawn on your PC with no internet access.
  • Diagrams for a process or a sequence of steps (flowcharts and sequence diagrams).
  • Option buttons when it needs you to pick: click one (or Tab to it and press Enter) and it is sent as your answer. Once you have answered — or typed something else instead — the buttons are disabled.
  • Collapsed sections for long logs and output — click the title to open one.
  • Pictures inline: screenshots and images the assistant saved.
Chart the volumes of the bracket, the plate and the pin

A chart or diagram that cannot be drawn is shown as its source text with one line saying why. Right-click a chart or diagram to copy its source.

Replies that go to Telegram, Teams, WhatsApp or the portal's chat page are plain text (Teams also shows bold, lists and links): the assistant is told which one it is writing for, so it does not send tables or charts there. A screenshot is sent as a photo on Telegram and shown on the chat page; Teams and WhatsApp get a description instead.

Undo, checkpoints and rollback

Undo the last operation
Take a checkpoint called "before the shell"
Revert to the "before the shell" checkpoint

A checkpoint is a restorable copy of the document, so it survives things that undo does not. Take one before a long or destructive sequence.

Refusals are a feature

SphereHarness will sometimes tell you it did not do something. That is deliberate. Operations that used to report success while quietly doing nothing now stop instead, and the refusal names the measurable reason — the nearest real edge and how far away it was, the gap that was too large, the solver status of a mate that would not solve, a cut that would have removed more material than the body contains.

Read the refusal text: it usually contains exactly the number you need to correct your request.

If a reply reads like a terse mechanical list of operations rather than an explanation, that is a safety net: when a model finishes a long tool run without writing a summary, SphereHarness builds one from the actual operation records instead of leaving you a blank bubble. The work did happen. Ask for a proper summary if you want one.

Asking about the model

SphereHarness is not only for changing geometry.

What features does the active part contain, with their parameters?
What is the volume, mass and centre of gravity of this part?
Give me the exact bounding box
How many faces does this body have, and how many are cylindrical?
What is the minimum distance between the boss and the rib?
What document is active, and where is it saved?

Selecting and looking at things

SphereHarness can highlight geometry in the viewport and change your view, which is often the quickest way to confirm you and it are talking about the same thing.

Select and highlight all the cylindrical faces so I can see which ones you mean
Set an isometric view and fit the model to the window
What is currently selected?
You do not need to select anything before asking for a fillet, chamfer, shell, draft, hole, pattern or boolean — those operations do their own selection. Selecting first is for when you want to see something.

When there is no operation for it

Occasionally you need a CAD command that has no direct equivalent in the chat. SphereHarness can open the command's dialog, read it, fill it in and confirm it.

Search the CAD commands for anything about "wrap", then open that command's dialog

See Native Macros & Automation for the full picture, including the caveat that a waiting dialog blocks the conversation until it is answered.

Doing the same thing many times

Create five plates 100 x 60 mm with thicknesses 5, 6, 8, 10 and 12 mm, named PLATE-01 to PLATE-05, saved in C:\parts

SphereHarness runs the sequence once per item and reports each result separately, so you can see which ones worked.

Five ways to make SphereHarness yours

  • Skills — standing instructions. “All drawings are A3, third angle, dimensions to one decimal.”
  • Macros — a recorded sequence you replay on demand, with no AI model in the run.
  • Variations — one master document in a grid of named sizes.
  • Knowledge base — your standards and catalogues, searchable and quotable.
  • Memory — durable facts about how you work.
    Remember that our default material is 6061-T6 and our default tolerance is ±0.1 mm

Exporting a conversation

Conversations can be exported from the chat window, so you can keep a record of what was asked and what was done, attach it to a support request, or hand it to a colleague picking the job up.

Presentation mode (screen sharing)

The chat card's menu has Presentation mode. While it is on, the card shows "Signed in" instead of your name, e-mail, organisation and plan, the administrator-mode banner turns grey ("UI control: elevated"), usage and cost are hidden, and the chat masks e-mail addresses, your name, your organisation and C:\Users\<name>\ paths in what it shows. It is display only — the conversation, its export and what the AI reads are unchanged — and only a person can switch it. AutoCAD's Start tab shows your Autodesk account name: open a drawing tab before sharing, or ask the assistant to switch to one.

Language

SphereHarness answers in the language you set in Settings, and the change applies immediately. See Settings & Languages.

What SphereHarness cannot do for you

It cannot take responsibility for the result. Everything it produces is a draft for your review — that is not a disclaimer bolted on the end, it is how the tool is designed to be used. Verifying results explains what to check and how.

See also