ChatGPT & Claude for specs, reports & comms
The least glamorous AI in the studio is the one that quietly hands you back two hours a day. Learn which language model to reach for, and exactly where it will lie to you with a straight face.
Two architects, same deadline. One wrote the spec by hand till midnight. One had a checked draft by 8pm.
Picture a Friday in a small Pune practice. A 14-page outline specification is due Monday, the client wants a progress report by email, and there are three sets of meeting minutes still in a notebook. The first architect opens a blank document and starts typing the way she always has. The second pastes her rough site notes into Claude, asks for a structured spec draft, then spends her evening doing the part only she can do — checking every clause, every dimension, every product against reality. Both ship on Monday. One of them got her weekend back. The catch is that the model also quietly invented an IS code that does not exist, and she caught it because she knew to look.
Same family, different reflexes — pick the model by the job
Claude for the long document, ChatGPT for the table
Both ChatGPT (OpenAI, GPT-5 era as of 2026) and Claude (Anthropic) are large language models — plausibility machines for text. They do not look anything up; they predict the most plausible next words. That single fact decides what you trust them with.
Where they diverge in practice is shape of work. Claude carries about 200,000 tokens of context — roughly a full project brief, a long tender, a 40-to-50-page spec — and holds coherence across the whole thing without losing the thread. So when the job is a long, structured document — a full specification, a proposal, a studio SOP, a 30-page client report — reach for Claude. Nice tie-in: Studio Matrx's own platform is built on Claude.
ChatGPT is the stronger tabular thinker. When you want a clean BOQ skeleton, a trade-by-trade schedule, a comparison matrix of three flooring options with columns for cost, durability and maintenance — ChatGPT's table output tends to be cleaner and its plugin ecosystem broader. Two more you'll meet in an Indian office: Microsoft Copilot, living inside Outlook, Word and Teams for firm documents, and Google Gemini, with long context tied into Google Workspace.
Drafts, summaries, reshaping — the boring 60% of studio writing
Map this to the green list from Module 0 and the use-cases are obvious. Language AI earns its keep wherever a knowledgeable human reviews the output before it matters:
Design briefs — turn a messy client conversation into a structured brief. Technical specs — a first draft of an outline specification you then correct clause by clause. Reports — a stage report or a condition survey written up from your bullet notes. Client emails — a polite, clear reply to an awkward variation request, in your tone. Meeting minutes — paste rough notes, get clean action-tracked minutes. Research synthesis — digest a 90-page tender or a stack of product PDFs into the three things that matter.
The time saved is real and it compounds. Indian firms report meaningful time savings on exactly this descriptive, divergent work. But notice what unites every item: a human who knows the project reads it, corrects it, and owns it before it leaves the studio.
The model writes the first 60% in minutes. The last 40% — the judgement, the checking, the voice — is still yours, and that is where the value is.
Every one of them hallucinates codes, IS numbers, products and prices
Here is the line that keeps you out of trouble: ALL of these models hallucinate on building regulations and codes. Setbacks, FAR, fire norms, NBC clauses, IS code numbers, prescriptive dimensions — a model will state a wrong one in the same confident prose as a right one. There is no red underline under a fabricated clause.
They do the same with products, prices, vendors and citations — inventing a tile range, a faucet model number, a supplier or a reference that looks completely real. The danger is never that the model is sometimes wrong; it is that it is wrong confidently, with no error message, inside otherwise-excellent prose.
So the workflow is fixed: diverge with the model, converge with the gazetted source. Let it draft the spec; you verify every code reference, every product, every number against the actual NBC, the actual bye-law, the real product catalogue, the real price list. The model is your brilliant intern who is never the architect of record — and regulatory verification stays manual, always.
Build a small library of saved prompts for your repeat documents — outline spec, stage report, RERA-style progress note, variation email. Feed the model your firm's house style by pasting a past (anonymised) example and asking it to match the tone and structure. Then route everything it touches that is code-, dimension- or structure-related through your normal professional check. A practical rule: nothing the model wrote enters a tender, a sanction set or a contract until a registered professional has re-derived every load-bearing number. The draft persuades; your verification builds.
Your highest-return uses are client-facing: design narratives, mood-board write-ups, scope-of-work sections, and the diplomatic emails that keep a fit-out on track. Claude will happily turn your concept notes into a warm two-paragraph design story a client will actually read. But it invents FF&E — sofas, fabrics, stone, light fittings — that look real and aren't for sale at that price. Use it to write the _direction_ and the _language_; specify the actual products from a real, sourced, costed schedule. Never let a generated product name reach a client as if you can buy it next week.
As a one-person studio this is the single highest-ROI AI you will adopt — it gives you the back-office of a firm three times your size. Free tiers of ChatGPT, Claude and Gemini handle most of it; paid tiers (roughly the price of two coffees a week as of 2026) unlock longer documents and better reliability. The discipline you cannot skip: you are the only checker. Keep a 'never trust without verifying' list taped to your monitor — codes, IS numbers, dimensions, products, prices, citations — and confirm each one from a real source before it ships.
Claude (Anthropic)
LLM — long documents
About 200k-token context; best for full specs, briefs, proposals, SOPs and 40-to-50-page reports where coherence matters. Studio Matrx is built on Claude. Still hallucinates codes and prices — verify them.
ChatGPT (OpenAI, GPT-5 era)
LLM — tables & structure
Strongest tabular output: BOQ skeletons, trade schedules, unit-price matrices, comparison tables; broad plugins. Will confidently invent product specs and IS numbers, so check every cell that carries a fact.
Microsoft Copilot
LLM inside Office/Outlook/Teams
Drafts and summarises right where your firm's documents already live — Word, Outlook, Teams. Convenient for emails and minutes; same hallucination caveat on any number, code or product it produces.
Google Gemini
LLM — long context + Workspace
Long context tied into Google Docs, Sheets and Gmail; good for teams living in Workspace. As with all of them, treat any cited regulation or price as a draft to verify, never the ruling.
“Claude has a huge context window and reads my whole brief, so when it cites an IS code or a setback it must be pulling it from a reliable source.”
Context window is about how much text the model can hold in view at once — not about accuracy. A 200k-token context lets Claude stay coherent across a long document; it does not give it a fact-checker or a live link to the NBC. It still generates the most plausible-sounding clause, which can be confidently, fluently wrong. Long memory and true facts are two different things — verify every code reference against the gazetted source regardless of how much it 'read'.
Workshop — draft an outline spec, then red-team it
You'll turn a few rough site notes into a structured outline specification using a reusable prompt template, then hunt down everything the model invented. Forty-five minutes, free tier of any chat AI, one real (or imagined) project.
Free: Claude or ChatGPT (Claude preferred for the long draft). Bring rough notes for one room or one trade from a real project.
SPEC-DRAFTING PROMPT TEMPLATE (paste, then fill the [brackets]): You are helping me draft an OUTLINE specification, not a final one. I am the architect of record and will verify every code, dimension, product and price myself. Project: [type, e.g. 3BHK apartment interior fit-out], in [CITY]. Trade/section: [e.g. internal wall finishes]. My rough notes: [paste your bullet notes]. Write the specification section with these rules: 1. Use clear sub-headings and numbered clauses. 2. Where a CODE, IS NUMBER, DIMENSION, PRODUCT or PRICE would go, insert a placeholder in CAPS like [VERIFY: IS code for ...] instead of guessing. 3. Flag anything you are unsure about. 4. Match a calm, professional Indian-practice tone.
- 1Paste the template, fill the brackets with your real notes, and run it. Read the draft once for structure and tone.
- 2Now red-team it: go clause by clause and tag every number, code, IS reference, product name and price as a thing to check. If the model ignored the placeholder rule and invented a code anyway, circle it — that is the lesson, live.
- 3Verify the two most load-bearing items against a real source: the actual IS code or NBC clause, the real product catalogue, the real rate. Mark each one right, close, or invented.
- 4Re-run the same notes through the other model (ChatGPT if you used Claude) and compare which gave the cleaner structure and which invented more. Keep the better starting point.
- 5Ask the model to reshape the verified draft into a one-paragraph plain-English summary for the client email. Notice how fast the comms layer falls out once the spec exists.
- 6Save your filled-in prompt as a reusable template, and write your keep-line: one sentence on what you will let the model own for specs going forward, and what you will always verify yourself.
You’ll walk away with
A structured outline-spec section drafted in minutes, a red-teamed list of everything you had to verify, and a reusable spec-drafting prompt template tuned to your studio's voice — plus a personal rule for how far to trust the model on technical writing.
Two quick experiments, five minutes each.
- 01Paste rough meeting notes and ask for 'clean minutes with an action table: owner, action, due date'. Check that every action is real and nothing was invented to fill the table.
- 02Ask the same model to write the same variation-request email in three tones — firm, warm, neutral — and notice how much faster picking from drafts is than starting from blank.
Language AI is the quiet, highest-ROI tool in the studio: Claude for long documents, ChatGPT for tables, both for the boring 60% of writing. They draft fast and they hallucinate codes, products and prices with total confidence. So diverge with the model, converge with the gazetted source — and keep your name, and your verification, on everything that ships.
Claude = long specs, briefs, reports (200k context). ChatGPT = BOQ tables, schedules, comparisons. Copilot and Gemini live in your office suite. All four hallucinate codes, IS numbers, products and prices — draft with them, verify every load-bearing fact against the real source yourself.
Should I use ChatGPT or Claude for architecture work?
Use both, by job. Reach for Claude when the task is a long, structured document — a full specification, a project brief, a 40-page report — because its larger context window holds coherence across the whole thing. Reach for ChatGPT when you want clean tables — a BOQ skeleton, a trade schedule, a comparison matrix. Either way, verify every code, dimension and price yourself; both invent facts with equal confidence.
Can I trust an LLM to write a building specification?
Trust it to write the first draft — the structure, the clauses, the tone — not the final, buildable one. It will fluently invent IS codes, product names, dimensions and prices that look real. Use it to skip the blank page, then go clause by clause verifying every technical reference against the actual gazetted code, the real product catalogue and a real price list. The draft is the model's; the responsibility stays yours.
Why does ChatGPT make up IS codes and product numbers?
Because it generates the most plausible-sounding text, not retrieved facts. A fabricated 'IS 4567' reads exactly as confidently as a real code, because the model is matching the pattern of how such references usually look — not consulting a database. It has no internal sense of true versus plausible, so it cannot warn you. That is why anything load-bearing — legally or structurally — must be checked outside the model.
If language AI can draft a spec, the obvious next question is whether it can help with the numbers and the rules around it — drafting BOQs, cross-referencing codes, and digesting tenders. That's exactly where we go next.