AI for
design.

Start with a real design task

Introduce the fictional CBS drawing package: a project brief, a written specification, a sample CAD file and a few known design rules. Ask the room which decisions require professional judgement and which repetitive checks could be assisted by AI. Frame the goal as a clearer, faster design review with traceable evidence, not a claim that AI can certify a drawing.

Show: One brief requirement, the related drawing element and the expected review note.

Meet the models and effort settings

A model is the engine; reasoning effort controls how much work it spends on the request. Compare the same drawing task at different settings: extracting dimensions is easier than reconciling a specification, several drawing sheets and conflicting notes. More effort can improve difficult analysis, but it costs more time and does not replace checking the source.

Choose the lowest setting that produces an adequate, checkable result. Availability and behaviour vary by model and workspace. OpenAI model guidance ↗

Discuss: Which task needs a fast pass, and which deserves deeper review?

What the harness needs to handle CAD

The harness gives a model instructions, project context, tools, permissions and checkpoints. For design work it must know which specification and rule set apply, which drawing revision is current, and whether it may only review or may create a revised copy. A PDF export can support visual review; reliable geometry, layers and native DWG or DXF editing require a compatible CAD parser, application or connected tool.

Show: Model → approved standards library → CAD-capable tool → issue report → designer review.

Free, Pro and business setups

Compare plans against the actual design workflow: file limits, available models and tools, usage, administration, data controls and cost. A free account is suitable for learning with fictional material; a paid individual plan can support regular solo work; a managed business workspace better fits controlled team access to real project drawings. Separate subscription costs from API and CAD integration costs, and check the chosen service's data terms before uploading confidential files.

Decide: Which experiments use sample drawings, and which need managed CBS access?

Design principles: brief to drawing

Work through the specification as a designer would: purpose, site and user needs, dimensions, clearances, accessibility, buildability, materials, annotations, drawing scale, layers and revision control. Ask AI to convert written requirements into a checkable drawing checklist, flag missing or contradictory information and draft a first-pass drawing plan. Demonstrate that each item needs a source, an applicability decision and a responsible designer; a plausible-looking plan is not evidence that it meets the project brief or relevant standards.

Exercise: Turn three brief clauses into drawing checks with measurable pass criteria.

Build a reusable standards and rules library

Create one approved source of design constraints that the assistant can call on for every relevant project: the project specification, company drafting rules, approved details, applicable code provisions and licensed standards. Each rule records its source, clause or reference, edition, jurisdiction, project scope, priority, owner and review date. A Codex skill can describe how to apply the rules; a connected, access-controlled library can supply current source material. Resolve conflicting rules with the designer instead of letting the model guess.

Build: Five sample rules with a source citation, applicability and a clear drawing check.

Review a CAD file with AI and bots

Send a sample drawing and its revision details through the rules workflow. In Codex, a connected CAD tool can extract drawing data; the model compares it with applicable rules and produces an issue register with sheet, layer or element, source clause, observed evidence, proposed change and unresolved questions. A visual PDF review is useful for discussion but may miss geometry or metadata. Compare the same scoped review with Grok or another bot, checking its file handling, citations, permissions and any proposed action rather than accepting an unsupported compliance claim.

Review: Which findings are measured, which are inferred and which need a designer?

Make a controlled CAD revision workflow

Map the workflow: receive the approved drawing and brief, identify the current revision, retrieve applicable rules, run geometry and document checks, propose edits, approve the changes, then create a new drawing copy through a CAD-capable integration or script. Re-run deterministic checks, compare revisions and keep a change log with source references. A plugin can expose read and write CAD actions while a skill governs their use. Native CAD editing is only possible when the chosen tool supports the format and operations; the designer still reviews and signs off the revised drawing.

Exercise: Define the tool permissions, approval point and acceptance checks for one drawing change.

Choose a small CBS pilot

Select one anonymised drawing and its specification, nominate five approved rules and agree on the expected output: a cited issue list first, then a revised copy if a suitable CAD connection is available. Record how the team will measure useful findings, false positives and time saved. Name the qualified designer who checks the final drawing and the person who owns updates to the rules library.

Leave with: One drawing, five rules, one review owner and one success measure.