How to Use Claude AI in Architecture: Specs and Planning Reports

Upload the document to a Claude project, ask for a summary that cites clause numbers, then check every cited clause against the original. Never paste client personal details, unissued fee terms, security-sensitive layouts, confidential consultant reports or anything under a non-disclosure clause. Redact names and addresses first and use a business plan.

A summary is a reading aid, not a compliance check, and Claude can only summarise what it can read, so the first thing to learn is where its reading stops. Design uses of AI get most of the attention, and they have their own tutorial in AI concept options for small practices. This one covers the written documents that eat a project architect's week: specifications, planning papers and the letters that follow them.

Follow me on Instagram@sagnikteaches

What Claude can read, and where it goes blind

Anthropic's upload page (updated 23 July 2026) sets the boundaries. Claude accepts PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB and JSON files, plus XLSX if you have turned on code execution and file creation, and JPEG, PNG, GIF and WebP images. A chat takes up to 20 files, each up to 500 MB, and PDFs are limited to 1,000 pages. Project files are tighter: 30 MB per file, with the number unlimited but the total content having to fit within Claude's context window. For PDFs of 100 pages or fewer, Claude analyses both the text and the visual elements such as charts and graphics. From 101 to 1,000 pages it processes text only. Over 1,000 pages it will not upload at all.

Connect on LinkedInSagnik Bhattacharya

One more detail matters for big projects. Anthropic's Projects page says that on a paid plan, when project knowledge approaches the context window limit, Claude switches on retrieval mode (RAG) to expand capacity. That is useful for question-and-answer, but it means Claude may be pulling excerpts rather than reading every page. For a summary of a whole specification, upload the document to a chat instead of project knowledge, or split it, and ask Claude which sections it actually drew on.

Subscribe on YouTube@sagnikteaches

When you ask for page references, tell Claude to use the PDF page numbers shown in your viewer, not the numbers printed on the page, which is Anthropic's own advice. Otherwise a specification with a 12-page front section will have every citation off by 12.

What never goes into Claude, and what goes in redacted

Most practices have a confidentiality clause somewhere in their appointment terms and no written rule about what may be typed into an AI tool. This list is a starting point for that rule, in three tiers. Adjust it for each client, because their terms govern, not mine.

A redaction routine that takes ten minutes

Do it in Word, before saving a copy as a PDF, with find and replace. Keep a one-line key somewhere local so you can put the names back into Claude's output.

Then check the places find and replace misses: headers and footers, the file name, the document properties (author and company fields), and text inside images. If you are saving as a PDF, open the PDF and run a search for the client's surname before uploading.

Five questions before you press send

• Is the document under a confidentiality clause? If so, has someone with authority cleared this use?
• Are names, addresses and numbers gone from the body, headers, footers, file name and properties?
• Is this a business plan? Anthropic's privacy centre says that by default it does not use inputs or outputs from its commercial products, such as Claude for Work, to train its models. On consumer plans (Free, Pro, Max and Ultra), your chats are used to improve Claude if you allow it, or if a chat is flagged for safety review, and Incognito chats are not used for improvement even when you have allowed it.
• Is memory off for this chat? Anthropic's memory help page says memory is on by default for Free, Pro, Max and Ultra plans (on Team and Enterprise, owners decide whether it is available), and that each project has its own memory. You can turn memory off for a single chat from the "+" menu before sending the first message, or turn off "Use account memory" on a project's page.
• Will anyone press the thumbs up or down button? Anthropic keeps feedback data for up to five years on consumer and commercial plans alike, and its commercial privacy page says a conversation sent as feedback is stored in full and may be used to train its models. On Team and Enterprise, an owner can switch the feedback button off with the "Rate chats" setting in the organisation's Data and Privacy settings.

The wider checklist for drawings and client data is in an architecture firm AI policy checklist, and the copyright side is in copyright and client data risks of AI in architecture practices.

A worked specification example: four passes over 64 pages

Take a 64-page specification for a two-storey rear extension. Everything here is invented for illustration, including the clause numbers and the sample outputs. The architect has made a redacted copy and uploads it to a chat rather than to project knowledge, because the summary needs the whole document. At 64 pages it is under the 100-page line, so Claude can read both text and visuals. The work happens in four passes, each with a different prompt, because one prompt that asks for everything produces a summary that is vague about all of it.

Pass one: an obligations register

This is the specification for a two-storey rear extension (redacted). Build an obligations register for the contractor as a table with these columns: clause number, PDF page, obligation (quote the key words), timing or trigger, evidence the contractor must provide.
Rules: use the clause numbers as printed. Quote key words rather than paraphrasing. If timing or evidence is not stated, write "not stated". Do not add obligations that are not in the document. At the end, list any sections you could not read or only partly read.

The output (illustrative) starts like this:

Three things need fixing. Claude has merged clauses 3.4 and 3.5 into one row, so 3.5's separate certificate requirement has vanished. The PDF page column uses printed page numbers for the first rows, which are two pages out, because the front cover and contents page carry no printed numbers. And the "not stated" in the last row is only correct for that clause; the timing is in the preliminaries, so the architect adds a note rather than trusting the blank. None of this is a reason to abandon the register. It is a reason to check it, which the section on the three-clause test below turns into a routine.

Pass two: submissions and approvals

The second prompt asks for something narrower: every submission, sample, approval, test, certificate or warranty the contractor must provide, with the clause, who receives it, how long before the work it must be submitted, and what the document says happens if there is no reply. Claude returns 14 items. The architect then does something no prompt can do and runs a search of the PDF for "submit", "approval", "sample", "certificate", "test" and "warranty". The search finds 15 requirements. The missing one sits in a table in Appendix B, exactly the kind of place an assistant skims.

That count check is the cheapest completeness test there is, so run it every time. Once the list is complete, it earns its keep: six of the 15 have lead times of 10 working days or more, which means six dates the project architect and contractor must put into the programme before the work is ordered. That list goes straight into the first coordination meeting.

Pass three: do these two statements conflict?

For conflicts, do not hand over the whole document and ask Claude to find them. Give it two texts you have already picked, from the specification and from a drawing note, and ask a narrow question.

Here are two statements from the same project. Do they conflict? Quote the words that matter in each. If the answer depends on information that is not given, say exactly what is missing. Do not decide which one governs.

Specification clause 3.4: "Fire-stopping to all penetrations through compartment walls shall achieve the same fire resistance as the wall."
Drawing note 12: "Seal all service penetrations with acoustic sealant."

A good answer (illustrative) reads: "Possibly. Clause 3.4 requires fire resistance equal to the wall. Note 12 mentions only acoustic sealant and says nothing about fire resistance. Whether they conflict depends on whether the sealant specified is also tested at the wall's rating. Missing: the product data sheet and the wall's rating." The architect raises a query with the contractor and designer. The final instruction in the prompt matters: which document governs when two disagree is a question for the contract, and a confident wrong answer from an assistant is worse than none.

Pass four: the questions a contractor will ask

The last pass looks forward. Ask for the ten questions a contractor is most likely to raise about the specification, citing clause numbers, covering ambiguities, missing information and anything that conflicts with the way the work is normally done. Sample items (illustrative): "Clause 6.2 asks for a 'high-quality finish' with no standard or sample as a benchmark. What is the benchmark?" and "Clause 8.1 refers to 'the approved supplier'. Where is the list?"

In this illustration, four of the ten questions were specific and worth acting on, four were generic ("What is the programme?") and two were wrong because the answer was already in the document. That is a normal yield. Delete the generic ones, check the wrong ones, and fix the four real gaps in the specification before it goes out, which is cheaper than answering the same queries during tender.

The whole exercise, including the keyword search and a spot check of 12 clauses against the PDF, takes about 70 minutes in this illustration: ten for setup, 20 for reviewing the four outputs, 30 for the spot checks and 10 for the count search. Log your own first run against your usual method before believing any saving, because the checking is where the time goes, and it should.

Planning documents: conditions, consultant reports and what changed

Planning papers differ from specifications in one way that matters for prompts. They mix fixed requirements (the conditions attached to an approval) with arguments and opinions (the officer's reasoning, a consultant's recommendations). Treat them separately.

Conditions into a table you can act on

Paste the redacted conditions, one document, and ask for a table: condition number, the exact wording quoted in full, when it must be met (before work starts, before a stage, before occupation, ongoing), the evidence needed, and who provides it. Quoting the wording in full is the control: the table is only a navigation aid, and the wording is what binds. An illustrative row: "Condition 4: samples of external materials to be approved in writing before above-ground work begins. Trigger: before superstructure. Evidence: sample panel and materials schedule. Provider: architect and contractor." Counting is a second check: if the decision lists 11 conditions and the table has 10 rows, find the missing one.

The authority's own wording and process vary a great deal, and so does whether a condition needs formal discharge, so Claude's table tells you where to look, not what the law requires. For the wider question of how far to take AI with a planning application, read where to draw the line on planning applications alongside this tutorial; for a statement written for the application itself, can AI draft a design and access statement covers what the architect still checks.

A consultant report in one page

A 40-page drainage, ecology, noise or heritage report usually has six findings that change the design and thirty pages of method. Ask for the six. The prompt below works for any of those report types.

From this consultant report (redacted), list only the findings that could change the design or the drawings. Table columns: finding (quote the key words), page, assumption it rests on, recommendation, drawing or specification item affected, who must act.
Rules: do not restate figures unless you give the page and table number. State units exactly as printed. If a recommendation is conditional, quote the condition.

A drainage example (illustrative): "Finding: surface-water discharge must be restricted to the rate in Table 3. Page 12. Assumption: based on a single infiltration test in one pit. Recommendation: attenuation tank sized to Table 3 volume. Affects: external works drawing, specification section on drainage. Who: drainage consultant, then architect." The assumption column is the valuable one. It shows where a design decision hangs on thin evidence, which is the question to put to the consultant. Check every figure against the printed table, and check the units, because "litres per second" and "cubic metres per hour" are easy to confuse in a summary.

What changed between version one and version two

Revised reports arrive with a polite covering note and no list of changes. Upload both versions in one chat (the limit is 20 files) and ask for a table of every change to a finding, figure or recommendation, with page references in both versions. Check the three biggest changes yourself. Then ask the follow-up that earns the effort: which of these changes affect drawings that have already been issued?

Consultant letters and chasers that stay factual

Letters to consultants are where an assistant's helpfulness becomes a risk, because it apologises, agrees and promises on your behalf without being asked. Give it facts and a tone, and keep conclusions to yourself.

Draft a short letter to the structural engineer (use the role, no name) from these facts. Facts: the calculations referred to in their letter of [date] have not been received; we need them before the [stage] issue on [date]; the foundation drawing is waiting on them.
Rules: factual and polite. State what is missing, what it holds up and the date needed. Do not apologise, do not accept responsibility for delay, do not agree to any change in scope or fee, and do not refer to any conversation I have not described. Under 150 words.

Without those rules, the draft tends to look like the first version below. With them, you get the second.

• Unprompted draft: "We apologise for the delay in coordination and are happy to accept responsibility for any resulting changes. As discussed on the phone, we would appreciate your calculations at your earliest convenience."
• With the rules: "Your letter of [date] refers to calculations that we have not received. The foundation drawing for the [stage] issue on [date] depends on them. Please send them by [date], or tell us the date by which you can."

The first version admits responsibility, refers to a phone call nobody described and sets no date. Search every AI draft for the red-flag words: sorry, apologise, accept, agree, guarantee, confirm, "as discussed" and "at your earliest convenience". Anything about fees, delay claims or a dispute is a letter you write yourself or take to your adviser, not one you start from an assistant's draft. The architect signs and sends; Claude never does.

Setting up Claude projects so every job starts the same

Anthropic's Projects page says projects are available to all users, including Free accounts, which can create up to five. A project holds knowledge files and a set of instructions that apply to every chat inside it. With memory on, later chats in the same project also draw on that project's memory, so details can travel from one chat to the next. Sharing a project with colleagues is a Team or Enterprise feature, with "Can view" and "Can edit" permissions.

For a small practice, three projects by document type work better than one project per job: "Specification review", "Planning papers" and "Consultant letters". Start a fresh chat for each job inside the right project, and switch memory off for that chat from the "+" menu before the first message, so one job's figures cannot carry into the next. Keep project knowledge to your own template material, such as your house exclusions list and your clause checklist, and not client documents. Put the standing rules in the project instructions, so nobody has to retype them:

You help a small architecture practice read and draft project documents.
Always: cite clause numbers and PDF page numbers for every point; quote key words rather than paraphrasing; write "not stated" if the document does not say; list any part of a document you could not read.
Never: invent a clause number, standard, figure or date; say that a design or product complies with anything; decide which document governs; apologise, accept responsibility or agree to a change on the practice's behalf.
Documents are redacted. Refer to people by role. Use plain English with our usual spelling. Give tables when I ask for a table.

On a Team plan, share the three projects with everyone so the instructions are identical across the practice, and how to set up Claude Projects as a shared team assistant walks through that. When a job ends, delete its chats; on commercial plans, Anthropic says a deleted conversation disappears from your history immediately and from its back-end storage within 30 days.

The three-clause test and four other checks

A summary is only as safe as the check behind it, and the checks should take minutes, not hours. These five cover most of the ways a document summary fails.

1. The three-clause test. Ask for a clause number and PDF page against every point. Pick three at random, find each in the original and compare the words. If one is wrong, treat the whole summary as unchecked and read it against the document.
2. The count check. Search the PDF for the words that mark obligations ("submit", "approval", "certificate", "sample", "test") and compare the number of hits with the number of rows Claude gave you. The 14 against 15 in the worked example is the typical result.
3. The unread-parts check. End every prompt with "list any sections you could not read or only partly read". A blank answer is not proof, but a long one tells you where to look.
4. The two-run check. Run the same prompt in two separate chats and compare the tables. Where they differ, one of them is wrong, and that row deserves your attention first.
5. The units check. Compare every figure and unit with the printed table. A summary can turn "litres per second" into "litres per hour" without any sign of hesitation.

Record that the check happened. A one-line note on the project file is enough: "Summary prepared with AI assistance from a redacted copy. Checked by [name], [date]; three of three clauses verified." If a client or insurer ever asks how you used these tools, that line answers it. Never let a summary replace the reading of a clause that carries a decision: the summary tells you which clause to read.

Plans and cost for a small practice

These are Claude's list prices from its pricing page in November 2026, in USD. Check the page for your account, since vendors change plans often.

A quick sum for a four-person practice: four Team Standard seats cost 4 x $25 = $100 a month on monthly billing, or 4 x $20 = $80 a month on annual billing. A sole practitioner has a real choice. Team needs at least two seats, so $40 to $50 a month, with the business-plan data defaults. Pro costs $17 to $20, with consumer data rules: fine if the model-improvement setting is off, memory is managed, and every document is redacted. Neither plan removes the need for the never-paste list. Whether the Team plan earns its seat price is a sum you can run on your own headcount.

If you would rather set up the workflow with someone, a first project, the redaction routine and the checks, my AI implementation consultation covers it. Otherwise, begin with one specification you already know well: run the four passes, run the checks, and see how many of Claude's rows you would have written the same way.

Questions architects ask about Claude and documents

Can Claude read drawings as well as specifications?

It accepts images and PDFs, but it does not read a drawing set the way an architect does. Anthropic's upload page says Claude analyses text and visuals in PDFs of 100 pages or fewer, and text only from 101 to 1,000 pages. Use it for written documents, and treat anything it says about a drawing as a prompt to look, not a finding.

Is a free Claude account fine for client documents?

Consumer plans let you choose whether your chats are used to improve the model, so check that setting, and remember it can change. Anthropic says its commercial plans, Team and Enterprise, do not use inputs or outputs for training by default. For client documents prefer a business plan, and remove names and addresses either way.

Should every job get its own Claude project?

A project per document type (specifications, planning papers, consultant letters) usually works better than one per job. Free accounts allow five projects, and project files count against the context window. Per-job projects make sense on a Team plan if you need to share them; delete their chats when the job ends.

How do I check a summary quickly?

Run the three-clause test. Ask Claude to cite the clause number and PDF page for each point, pick three at random, and find them in the original yourself. If one is wrong, check the whole summary. Use PDF page numbers, not the printed ones, as Anthropic's own tip advises, so citations match.

Further reads

• Claude vs ChatGPT vs Gemini for Architects: Which to Pay For — Which assistant to pay for, by job.
• Claude (Formerly Cowork) for Architects: Five Admin Tasks to Try First — Five admin tasks to try with Claude first.
• AI for MEP Engineers: Specification First Drafts and Checks — The engineer's side of specification drafting.
• Is Claude Team Worth It for a Small Business? — Whether the Team plan earns its seat price.
• How Can AI Help Architects? Briefs, Planning Notes, Consultant Chasers — Briefs, planning notes and consultant chasers.
• AI Software for Architectural Drawings: Notes, Schedules and Sheet Checks — Checking drawing sets, which this tutorial leaves out.
• AI Programs for Architects: Room Data Sheets From a Client Brief — A worked example of turning a messy client brief into room data sheets with AI, and the review pass that catches what the draft invents or drops.
• AI Tools for Architects 2026: A Toolkit From Brief to Handover — Which AI tool to use at concept, design, tender, site and handover, what each costs on its own pricing page, and three stacks for a five-person practice.
• How to Use AI in an Architecture Firm: A First-Quarter Pilot Plan — Run three 13-week pilots for meeting minutes, fee proposals and site reports, with baselines, named KPIs, a weekly quality sample and a written go or stop rule.
• Top AI Tools for Architects: Eight Compared by Job and Price — Eight AI tools matched to five practice jobs, with USD entry prices, plan limits and a monthly cost for a five-person practice.
• AI Drawing Review: Check Architectural Sets Before You Issue Them — Where AI is reliable on a drawing set (index, notes, references), where it fails (dimensions, line types), and the checklist and sign-off that close the gap.
• Architecture Firms Adopting AI Design: What to Tell Your PI Insurer — A PI insurer disclosure checklist for architecture practices using AI: proposal forms, material changes, what to log and what to ask your broker.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Anthropic's file-upload help page, Projects help page, privacy centre articles on training and retention, and Claude's pricing page, all read in November 2026. The specification and sample outputs are invented illustrations.