Selection-aware edits and revision guards
Carry geometry references and design intent into an edit without guessing which entity you meant.
“Make this hole larger” is a clear instruction when two engineers are looking at the same feature. It becomes less clear when the model changes, the selection is lost or a face number is reused. A useful conversational CAD workflow must preserve what was selected, which model revision it belonged to and which requirements should survive the edit.
Selection provides geometric context
Intact lets a user select parts, faces or edges in the viewer, attach references or sketches and describe a question or edit. Before work starts, Ask AI saves the description, native selection, source revision and artifact, references and an optional viewport image. This creates a concrete request that can be read again instead of relying only on a transient highlight.
A practical request might be: “Increase the selected mounting hole to 8 mm. Preserve the other holes and the outside dimensions. Ask before reducing the required clearance.” The selection identifies the intended target; the text states the change and the invariants. A reference image can explain intent, but it should not silently override measurable native geometry.
Understand where Ask AI actually runs
When a supported ChatGPT host exposes the component messaging bridge, Intact can send a follow-up carrying the saved request identity and description. That capability is detected at runtime. It is conditional, rather than a guarantee that every browser or desktop host can insert a request into the current conversation.
In the localhost viewer, Ask AI otherwise starts a separate model-scoped task through the signed-in OpenAI local client automatically. Its progress, answer and clarification replies appear in Intact. That task does not join or impersonate the currently open ChatGPT conversation. The workflow does not require a separate model-service API key.
The Ask AI handoff documentation describes these paths. If a bridge attempt fails, the error remains retryable rather than silently starting a duplicate local task. The release tests distinguish mocked bridge behavior from real local-client execution; live current-chat delivery in Windows ChatGPT was not verified in the 0.16 release report.
Revision guards protect the intended starting point
An edit request belongs to a model state. If another operation changes that state before an edit is applied, the original selection may no longer mean the same thing. Revision guards make that dependency explicit. The task should read the current model, compare the expected revision and reconcile changed context before applying a mutation.
This matters even for simple parts. Adding a fillet can split an edge or replace a face. A numeric face identifier that once described the mounting surface can become stale after rebuilding. Treating the old number as a permanent semantic name would hide ambiguity instead of solving it.
Bind references, then verify them
Intact supports stable labeled references to inspected entities. Same-artifact references retain identity; other matches use geometric signatures, while hole anchors can follow a named feature and center index. These mechanisms have limits. Reordering hole centers changes the index convention, and multiple geometric candidates must not be resolved by guessing.
When a referenced feature disappears, unrelated source changes invalidate its interpretation or several candidates match, reselect and bind the intended entity. Replies preserve conversation context and current references, while stale numeric references are dropped after geometry changes. Explicit ambiguity is a useful result: it tells the user what decision remains unresolved.
Preserve editable intent and check the outcome
For structured projects, the task should retain supported sketches, parameters and feature relationships. Updating an operation should preserve unrelated steps. A staged feature graph is not yet rebuilt geometry, and a failed rebuild should preserve the last valid geometry. Imported STEP and arbitrary Python programs do not automatically become complete parametric feature histories.
After an edit, inspect the resulting native geometry and verify the actual export against the updated requirements. Ask whether the selected hole has the requested diameter, whether other parts remain and whether clearance still passes. A confident answer in the conversation is not a substitute for those checks.
Give questions and analysis the right scope
A read-only question should not mutate CAD. Historical previews remain read-only, and simulation requests cannot change the model. Mechanical Ask AI can run an actual analysis once material, supports and loads are sufficient; missing physical inputs should lead to clarification. Preparation without solver evidence is not a successful engineering result.
See the structured modeling contracts for reference and verification limits. The website demo presents prebuilt browser interactions; real AI and CAD tasks run through the native Windows local setup. Read the system architecture before starting an editing session.
Use this guide in ChatGPT
Enable the installed Intact Simulation plugin in a new conversation. State the design goal and the checks you expect, then ask the assistant to report the native measurements and failed or unverified criteria. Use Review for a measured overview and ask for a parameter preview before committing an alternative. Keep the accepted revision available while you compare results.
Follow How to Use for the practical controls and Architecture for the audit and verification chain. Plugin setup explains the local runtime and your own ChatGPT login.
Install once.
Design through conversation.
Download Intact v0.20.0 for Windows. Install privately in Codex with your own account; CAD runs on your computer.
Follow the installation procedure →