ChatGPT for Building Design: Briefs, Options and Code Checks

ChatGPT suits the writing around a design: turning a client's enquiry into a structured brief, setting out options with trade-offs, and listing the regulation topics a project might touch. It should never be the source for what a regulation says or for a calculation. Treat each code point it raises as a question and answer it from the current document.

The split that works is by how an error gets caught. A mistake in a brief or an option list is caught by reading, because the client or a colleague says "that isn't what I meant". A wrong regulation answer or a wrong calculation is not caught by reading: the text looks equally fluent either way, and only a lookup or a calculation exposes it. ChatGPT will write clause-style statements whether or not the clause exists.

Follow me on Instagram@sagnikteaches

The examples below follow one made-up three-person practice (two designers and an administrator) through an enquiry from a café owner who wants to extend the back of her building, add seats and keep trading during the works. Every enquiry, output and figure is illustrative, written by hand to show the pattern, and the code topics are prompts for you to check, never statements of what any code requires.

Connect on LinkedInSagnik Bhattacharya

Two practical choices come before any prompt: which ChatGPT plan you use, and what you paste into it. Settle both once for the whole practice rather than person by person, because a setting each designer chose alone is one nobody can vouch for when a client asks where their drawings and emails went.

Subscribe on YouTube@sagnikteaches

Three zones: draft freely, draft and verify, never delegate

Sort every job in the practice into one of three zones before opening ChatGPT. The zone decides who checks what.

OpenAI's own usage policies draw a similar line. They rule out tailored advice that requires a licence without a licensed professional involved, and the revision that takes effect on 20 November 2026 adds that you may not state or imply that the AI itself holds professional credentials. In a design practice the professional who answers for the drawing is you. The tool drafts; you answer.

Set up a practice Project and decide what client data goes in

Give the practice a ChatGPT Project rather than a pile of loose chats. OpenAI's help page describes a project as a place that keeps related chats, files and instructions together, with the instructions set under Project settings in the three-dot menu. Custom GPTs are being retired (OpenAI's announced stop date is 11 December 2026), so a Project is the safer home for standing instructions. On the plan limits, a project holds 5 files on Free, 25 on Go and Plus, and 40 on Business and Pro.

Paste the standing instructions once and every chat inside the project inherits them. These are the ones I'd start with:

You assist a small building design practice. You write and organise; you do not decide.
Rules:
1. Never state a regulation clause number, figure, dimension or limit as fact. If a topic needs a rule, name the topic and the type of document that holds the answer.
2. Never perform or check calculations (structural, thermal, daylight, escape widths, drainage, ventilation, areas, costs).
3. Never say a design "complies". Write "to be confirmed against [named document]".
4. If the input does not say something, write UNKNOWN. Do not assume.
5. Use British English, plain words and short sentences. No marketing language.
6. Keep client names, addresses and contact details out of every output. Use "the client" and "the site".

On the data side, check the plan before pasting anything about a client. OpenAI's Data controls page says that on personal plans (it names Free, Plus and Pro) the setting Improve the model for everyone is on by default. You turn it off under Settings, then Data controls, and you keep your chat history. Turning it off does not delete anything already saved. Project chats are covered as well: on personal accounts, shared project data can be used for training only when the owner and every contributor have the setting on. Business, Enterprise and Edu workspaces are not used for training by default.

Two smaller details matter in an office. Temporary Chat keeps a conversation out of your history and out of training, although OpenAI may retain it for up to 30 days for safety. And the thumbs-up and thumbs-down buttons are an exception to your opt-out: OpenAI says that if you send feedback, the whole conversation attached to it may be used for training. The practice rule is no feedback buttons on any chat that touched a client's project.

Whatever the plan, strip identifiers on the way in. Replace the client's name with "the client", the street with "the site", and delete phone numbers, emails and anything covered by a confidentiality agreement. A brief written from "a café owner with a leased corner building" loses nothing, and you add the real names back in your own document.

From a rambling enquiry to a one-page brief

Enquiries arrive as voice notes, long emails and half-remembered phone calls. This is the job ChatGPT does best, because the source text sits in front of it and a human reader will spot a misreading within seconds. Here is an illustrative enquiry, already anonymised, as the café owner might leave it on voicemail and the administrator typed it up.

"We want to extend out the back so we can fit about twelve more seats and a bigger prep kitchen. We rent the building and the flat above is someone else's. We can't close for more than a couple of weeks, summer is our busy time. Budget's around a hundred and forty thousand, I think, but I haven't talked to the landlord about contributing. We need a toilet that wheelchair users can use, our current one isn't. I'd love some outside seating as well if it's possible. Oh, and there's a bin store out the back that we need to keep."

The prompt does three things: it fixes the headings, it bans assumptions, and it forces the client's own numbers through unchanged.

Turn the enquiry below into a one-page design brief with these headings: Project in one sentence; Client goals; Must-haves; Nice-to-haves; Constraints stated; Budget and timing; People who must agree; Questions we must ask before any design work.
Use only what the enquiry says. Anything not stated goes under "Questions", never in the brief. Quote the client's own numbers exactly as given.
ENQUIRY:
[paste the anonymised enquiry]

A typical first draft comes back looking like this (illustrative, shortened):

Project in one sentence: Extend the rear of a rented café to add about 12 seats, a larger prep kitchen and an accessible toilet.

Must-haves: An accessible toilet; the existing bin store retained; closure of no more than "a couple of weeks".

Nice-to-haves: Outside seating.

Budget and timing: "around a hundred and forty thousand, I think"; summer is the busy period.

People who must agree: The landlord; the owner of the flat above (inferred, not stated).

Questions: Does the budget include fees and kitchen equipment? What does the lease allow? Is the landlord contributing? How many seats are there now? Which weeks count as summer?

Read it against the enquiry and fix what ChatGPT bent. In a first draft I'd expect at least two of these:

• It wrote "$140,000 including fees". The owner never said that, so the line goes back to her words and "fees included?" moves to Questions.
• It turned "we can't close for more than a couple of weeks" into "phased construction". That is a design decision the client never made. Delete it.
• It converted "about twelve more seats" into "capacity of 12 additional covers". Keep "about twelve" until someone counts the floor area.
• It added a person to "People who must agree" by inference. Keep the inference only where you can point to the sentence that supports it, and label it as inferred, as above.

Once the brief is right, the same approach feeds the paperwork around it. A tidy brief is the raw material for a fee proposal, and how architects use AI for fee proposals, briefs and paperwork covers that stage. If you collect enquiries through a form, building and summarising a client questionnaire with AI covers that route.

Reading a client's hand sketch

Plenty of owners send a photo of a sketch drawn on the back of an envelope. ChatGPT can read an uploaded photo, but treat the result as transcription, never measurement. Ask for a list of rooms and every dimension written on the sketch, copied exactly, with anything illegible marked UNREADABLE and no dimension inferred. Then lay the list beside the sketch and tick each line. The typical failure is a handwritten 1 read as a 7, or a width taken from the wrong arrow, so the tick-through is the step that matters.

Layout options with trade-offs and failure points

With a corrected brief, ask for options. The useful version of this prompt does not ask "what's the best layout". It asks for several, each with the condition that would make it fail, because the failure conditions show you which facts about the site you haven't told it yet.

Using the brief below, set out four layout options for the project. Include one option that does less than the client asked for.
For each option give: the idea in two sentences; what it gives the client; what it costs in space, disruption or complexity (words only, no figures); what would make this option fail; the regulation topics it would trigger (topics only, no clause numbers); and the one question that would decide whether to pursue it.
Do not invent site facts. If the choice depends on something the brief does not say, list it as a question.
BRIEF:
[paste the corrected brief]

The answer is long, so here is the shape of it for the café brief (illustrative, condensed to one line per cell):

Treat the list as prompts for your own thinking, not as design. ChatGPT hasn't seen the site, so a typical run includes one option that can't work (here, a version that put the new kitchen where the bin store sits) and one you hadn't listed. The impossible one is still useful, because the "fails if" line tells you which constraint to feed back in. Re-run with "assume the bin store stays where it is" and the list tightens. Then pick two options, sketch them yourself, and move on. For why a generated plan breaks against a real brief, AI floor plan generation for architects walks through the failures.

A regulation-check workflow in four steps

This is the part that keeps the practice safe, so it has a fixed shape. ChatGPT supplies a list of topics. People supply the answers. A log shows which is which.

Step 1: ask for topics, never answers

For a single-storey rear extension to an existing café with a flat above, list the regulation topics a designer would normally need to check, as a table with three columns: Topic; The question the designer must answer; The type of document that holds the answer (for example "the adopted building code", "the approving authority's published guidance", "the fire strategy").
Do not quote clause numbers, figures, dimensions or limits. Do not say whether the design complies. If you are unsure whether a topic applies, say so in the Topic column.

A usable answer has eight to twelve rows. Typical ones (illustrative): separation between the café and the flat above; means of escape from the enlarged seating area; accessible toilet and the route to it; toilet provision for the new seat count; kitchen extraction and ventilation; drainage and grease management; energy performance of the new envelope; structural work to the existing wall; land-use approval beyond the building regulations; and the lease. The last two aren't regulations in the strict sense, but they gate the project, and ChatGPT is good at remembering gates.

Step 2: turn each topic into a named document

The project owner, not ChatGPT, decides where each answer lives. For every row, name the document and the section you will open: the building code or regulations adopted for the site and the edition in force, any local amendments, and the approving authority's own guidance for this building type. Then search the live document for the topic's words. Never search for a clause number someone gave you, because the number is the part ChatGPT is most likely to get wrong.

Step 3: log each check

The log is a table in the project folder. Each line shows who opened what and what happened. Here is the café log part-way through (illustrative):

Step 4: stop at the signature

Until a line has a document, a section and a date, the topic is open and nobody writes "complies" beside it. The wording on drawings and in emails stays "to be confirmed against" followed by the named document. The person who signs reads the log before signing, and that person, not the software, carries the responsibility.

A builder's version of this routine is in can AI answer building code questions: a builder's safe-use routine.

An illustrative miss: the clause that read perfectly

Suppose a junior asks ChatGPT how wide the accessible route to the new toilet must be. The answer comes back with a width in millimetres, a clause number and a sentence that reads like a quotation. Nothing about it looks wrong. When the junior opens the adopted code, the clause number belongs to a different section, and the width matches an older edition that a later edition replaced. The answer was fluent, specific and stale.

What caught it was the log: the line had no document or section, so the review stopped there. Three habits follow from that case. A topic isn't closed until the log has a document and a section. A clause number from ChatGPT never goes onto a drawing note. And asking "what is your source?" proves nothing, because you'll get a citation-shaped answer back, which is the same fluent text again.

Calculations stay out, and what ChatGPT can still do around them

Calculations fail differently from prose. A wrong beam size, an area out by a factor of ten, millimetres confused with metres: each looks perfectly plausible on the page, and none is caught by reading. ChatGPT has no access to your drawings or analysis software, and a confident figure from it is a guess with good manners. The same boundary applied to heating and cooling loads is the subject of can AI do an HVAC load calculation, and it holds across every discipline in the practice.

There is still plenty of useful work around the numbers, all of it text:

• List the input data an engineer will need before starting (loads, spans, ground information, finishes), then draft the request email.
• Write the assumptions page that sits at the front of a calculation pack, from the engineer's own notes.
• Turn an engineer's result paragraph into plain English for the client. Paste the engineer's figures exactly and tell ChatGPT it must not change, round or add any number.
• Keep a calculation register (which calculations exist, who owns each, status) in a table you can paste in and update.

When a client arrives with ChatGPT's answer

Expect this to become routine. A homeowner writes: "ChatGPT says a four-metre extension needs no approval, so can we skip the drawings?" Don't argue with the tool. Ask for the chat, then use ChatGPT yourself to turn its claims into a checklist (illustrative prompt):

Below is an answer a client received from ChatGPT about their project. List every factual claim in it as a numbered list, one claim per line, and beside each one name the type of document that would confirm or refute it. Do not say whether any claim is right. Remove any names or addresses.
ANSWER:
[paste the client's chat]

The claims list becomes the agenda for your reply, and the log from the previous section is where the answers get recorded. A reply that works (illustrative):

Thank you for sending that over. I've split the answer into six statements and checked each against the documents that apply to your site. Four hold up, one depends on a condition that applies to your property and one is out of date. I've attached the list with the source for each, so you can see how we reached it. This is why we confirm these points before drawing anything: the answer for your site depends on the edition in force and on local rules that a general chatbot can't see.

That reply is courteous, shows the work and moves the client from "the AI said" to "the documents say". It takes about twenty minutes, and clients tend to value the structure more than the verdict.

Checks before anything leaves the practice

Pin this beside the screen. Every item is a yes or no.

• Is every client name, address and contact detail absent from what was pasted in and from the draft?
• Does every number in the draft appear in the source text, or has a person recalculated it?
• Is every regulation statement either deleted or backed by a log line with document, section and date?
• Has every UNKNOWN become a question to the client or a note to a consultant?
• Does the draft avoid the words "complies", "approved" and "meets" unless a named person has confirmed them?
• Does it claim nothing about what the client wants beyond what the client said?
• Do the initials of the checker appear on the log line or the file?

Cost and time for a three-person practice

Plan prices are list prices in USD and change, so confirm them on the pricing page. At the time of writing, Plus is $20 a month per person. ChatGPT Business is $25 per user a month billed monthly or $20 billed annually, with a minimum of two seats. For three people that comes to about $60 a month either way (three Plus plans, or three Business seats on annual billing), or $75 on monthly Business billing. The difference is control: Business is not used for model training by default and adds workspace controls and group sharing of projects, while three personal accounts mean three sets of settings to police.

The time case is modest and honest. Take a practice that handles eight new enquiries a month. Writing a brief by hand takes about 45 minutes. With ChatGPT it takes about five minutes to paste and prompt and ten to read and correct, so 15 minutes, a saving of 30 minutes each, or four hours a month. The regulation list saves little, because the verification takes as long as it ever did. The gain there is completeness: the list reminds you of the gate you'd have remembered on day three rather than day one.

Try it on one live enquiry this week. Anonymise it, run the brief prompt, correct the draft against the source, and time each step. If you'd rather set up the Project, the log and the rules with someone, that's what my AI implementation consultation covers.

Questions a design practice asks before adopting ChatGPT

Do I need to tell clients that the practice uses ChatGPT?

Read your appointment terms and your professional indemnity policy first, because both can set requirements that a tutorial cannot see. If ChatGPT only drafts your own wording and no client data goes in, there may be little to disclose, but ask your insurer to confirm. If client material does go in, say so plainly in your terms and name the plan and settings you use.

Can ChatGPT draft the design justification for an approval application?

It can turn your own notes into readable prose, which saves time. It cannot know which policies the approving authority applies this year, so every policy reference it adds must be removed or checked against the authority's current documents. Give it your reasoning and ask it to write only that, with UNKNOWN wherever your notes are silent.

Should staff be allowed to use their own personal ChatGPT accounts?

A practice-owned workspace is easier to control. Personal accounts have their own data settings, and on personal plans the model-training switch starts on, so each person would need to turn it off and keep it off. Whatever you choose, write down which accounts are allowed, which settings apply and what must never be pasted in, then have everyone sign it.

How often should the regulation log be re-checked?

Re-open it at every stage gate and whenever the approving authority announces a new edition or amendment. Each line carries the document, section and date opened, so a re-check means reading the same section again and noting any change. A line older than the current edition of its document is treated as unchecked.

Further reads

• Is ChatGPT Plus Worth It for an Architect? A Practice Cost Check — Price the plan options for a practice before buying seats.
• Claude vs ChatGPT vs Gemini for Architects: Which to Pay For — See how the three assistants compare on design-office work.
• Copyright and Client Data Risks of AI in Architecture Practices — Cover copyright and client-data risks in your written policy.
• Can AI Write an Architect's Snagging List From Site Photos? — Another drafting job where a person must check every line.
• Can AI Help a Small Architecture Practice? Where It Saves Time — Where AI saves a small practice time across the whole week.
• AI Automation for Architects: Enquiry to Fee Invoice, Step by Step — Automate the path from first enquiry to fee invoice.
• ChatGPT for Interior Designers: Ten Prompts for Briefs and Schedules — Ten more prompts, written for briefs and schedules on the interiors side.
• ChatGPT for Landscape Architects: Planting Notes, Reports and Emails — A checking routine for using ChatGPT on planting notes, reports and client emails: it drafts the words, you choose and verify every plant.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help pages (Data controls in ChatGPT; Projects in ChatGPT; ChatGPT Business overview), OpenAI usage policies and ChatGPT pricing pages, all read on 1 November 2026.