ChatGPT for architects: a practical DWG workflow

Choose the deliverable before uploading

An architect receiving a consultant’s DWG usually needs a concrete answer: which room labels differ from the schedule, which references are missing, or which layers need remapping. “Review this drawing” leaves both the evidence and the deliverable undefined. A better first task is a numbered issue register for one floor, with a sheet location and an action for each finding.

Use ChatGPT to organize a review, compare exported information or draft a small automation routine. A DWG attachment being accepted does not establish that its objects were parsed. Ask which representation the assistant used: rendered pixels, extracted text, an object inventory or a connected CAD tool.

OpenAI’s PDF visual retrieval FAQ describes an Enterprise-only feature for PDFs attached in conversations; GPT Knowledge and Project Files use text-only retrieval. If visual retrieval is unavailable, attach separate readable image crops of the plan.

Prepare one drawing package

Keep the issued source unchanged. Make a review copy and record the drawing number, revision, issue date and authoring application. Resolve the external references needed for the chosen floor; identify any intentionally omitted files. Export the relevant layout to a readable PDF and include enlarged crops for dense details.

For comparisons, add a structured table exported from CAD: room number, room name and area, or block name, handle and selected attributes. Preserve leading zeros in identifiers. State whether dimensions are millimetres, metres or feet and inches, and whether areas are square metres or square feet. A title-block scale describes a plotted view; it does not establish model-space units.

Include only client information permitted in your organization’s approved AI environment. Removing a client’s name does not remove sensitive layouts or security information.

A prompt for a useful first review

Review only Level 02 in drawing A-102 revision C. Inputs are the exported plan PDF and room_schedule_C.csv. Dimensions are millimetres; scheduled areas are square metres. Compare room numbers and names between the two inputs. Return a table with room number, PDF location, schedule value, visible plan value, discrepancy and suggested action. Preserve leading zeros. Do not infer hidden rooms or measure areas from pixels. List unreadable labels separately. Do not change files or make a compliance assessment.

First inspect two rooms you know well. If ChatGPT cannot locate them reliably, improve the input before extending the review. Ask it to distinguish a confirmed mismatch from a label it could not read. Otherwise an apparently clean issue list may simply reflect missing evidence.

Keep the review tied to revision C. If a revised schedule arrives, restart the comparison with an explicit new pair of inputs rather than allowing old and new values to mix in the conversation.

From a proposed script to a controlled CAD operation

For repetitive work, request a read-only AutoLISP or other supported CAD routine before requesting edits. Give the exact AutoCAD release, operating system, available runtime and target object types. Ask for an inventory of affected objects and exclusions. Review the code before loading it in your approved CAD environment.

A sensible first automation lists text objects on specified layers and writes their handles and values to a table. A later operation might rename layers using an approved mapping. Specify collision handling, locked layers and reference-dependent layers; a program must report exceptions instead of silently skipping them.

Run the reviewed routine on a disposable copy. Compare object counts, a known dimension and a plotted layout before and after. Reopen the saved output. A successful script message is weaker evidence than the file you can open and inspect.

ChatGPT can also act through configured tools: OpenAI documents custom MCP connections. That is a separate integration requiring the appropriate server, permissions and CAD operations. A chat upload does not create that connection, and a product-help connection does not imply access to your drawing.

Keep PDF conversion separate from drawing review

If the source is only a PDF, establish whether it contains vectors or a scan. Autodesk’s PDF import options distinguish vector objects from attached raster images. Importing a scanned sheet does not automatically produce editable wall geometry.

Calibrate imported geometry against known dimensions and inspect text, layers and curves. Review findings against the source sheet; do not treat an imported outline as an exact recovery of the original CAD model.

For PDF-to-DWG reconstruction in an ARAI environment with the reconstruction skill enabled, provide the source sheet, known dimensions, target units and acceptable deviation. Open the result in CAD and compare it with the PDF.

Check your drawing in ARAI

Attach the DWG, a matching PDF and the external references needed for the selected floor. Add the room schedule, selected room scope, length and area units, and agreed boundary rules for measuring each room. Ask ARAI to compare the schedule with areas from those drawing contours. Request object identifiers and sheet locations so you can review each finding in CAD.

Check room areas in ARAI. Attach your files to the draft before sending. Check the reported discrepancies against the drawing and schedule before making changes. See the Revit export guide.