ChatGPT Prompts for Civil and Billing Engineers (India): Tested, With Failures

ChatGPT is genuinely useful to a civil or billing engineer for four jobs: drafting documents that follow a known structure (RA bill covering letters, site instructions, method statements), explaining and cross-checking calculations you already understand, turning messy Excel into clean sheets through its Python data-analysis mode, and preparing checklists and formats. It is unreliable for three jobs that matter most on Indian sites: quoting IS code clauses and values, measuring anything from a drawing image, and pricing against the current DSR. The prompts below are built around that split, with real outputs and the failures we hit while testing them.

Every prompt library ranking for this query in 2026 is a list of fill-in-the-blank templates: no named engineer, no IS code, no DSR, no example of what the model actually returned. This one is different in exactly those four ways. Copy the prompts, but read the failure notes first.

What ChatGPT can and cannot do for engineers (2026)#

The exam evidence deserves one more sentence because it is the best independent test we have. Naser and colleagues gave GPT-4 the Fundamentals of Engineering civil exam in 2023 and it scored 70.9% on FE and 46.2% on PE structural questions, with errors that included computing a moment of inertia twelve times too large while laying out the correct method. Two years later, Oblitas, Adeeb and Cruz-Noguez ran six current models on the official 100-question FE Civil practice exam: strong on text, weak on anything visual, and weakest of all on structural engineering and statics. That pattern, fluent method with unreliable numbers and near-blindness to diagrams, is what you should expect on your own work. ChatGPT is also only one layer of the site stack — our AI tools for site engineers guide maps the rest: Hindi voice typing, offline translation, and the DPR tools worth a demo call.

Before you paste anything: the data rule#

On the Free, Go, Plus and Pro plans, OpenAI uses your conversations and uploaded files to train its models by default. Turn that off before you upload a client's drawing or a tender BOQ: Settings, then Data Controls, then switch off "Improve the model for everyone". Temporary Chats are also excluded from training. Business and Enterprise plans are not trained on by default. If your contract has a confidentiality clause, and most private-sector ones do, treat uploading unmasked client documents to a consumer AI account as a breach until your client says otherwise. Strip project names and client names from anything you paste; the prompts below work just as well on anonymised inputs.

Prompts for documents and correspondence (where it earns its keep)#

1. RA bill covering letter. "Draft a covering letter from a contractor to the Executive Engineer forwarding Running Account Bill No. 4 for [work name], agreement no. [x], for the period [dates]. Reference the measurement book pages, state the gross value of work done to date and the net amount claimed since the previous bill, mention that measurements were jointly recorded and signed, and request certification and release of payment within the contractual period. Formal Indian PWD register, under 200 words."

What we got: a usable letter on the first try, correctly structured around the Form 26 logic of up-to-date value minus previous bill. It invented a "Clause 7" for payment timing. Delete any clause number it writes and insert your own agreement's.

2. Site instruction or non-conformance note. "Write a site instruction to the subcontractor for [description of defect, e.g. honeycombing observed in column C-14 at first floor]. Reference the relevant IS 456 requirement in general terms without quoting a clause number, state the rectification method to be proposed by the subcontractor for the engineer's approval, set a 48-hour response, and note that the cost of rectification is to the subcontractor's account under the work order."

The "without quoting a clause number" instruction matters. Left free, the model cites clauses that may not say what it claims. Let it write the prose; you add the clause from the code on your shelf.

3. Method statement skeleton. "Produce a method statement outline for [activity, e.g. raft foundation pour of 380 cum M30 concrete using two transit-mixer fleets and a boom placer]. Sections: scope, references, resources, sequence, quality checks with hold points, safety, and contingency for pump breakdown and rain. Mark every numerical value that must be confirmed from the approved drawings or mix design with [CONFIRM]."

The [CONFIRM] tag is the single most useful habit in this list. It turns the model's tendency to fill gaps into a visible checklist instead of a hidden error.

4. Joint measurement request and JMR format. "Draft a letter requesting joint measurement of [items] on [date] and attach a joint measurement sheet format with columns for item, location, length, breadth, depth in that order, unit, quantity, and signature blocks for contractor and client engineer." It reproduces the L, B, D order that IS 1200 Part 1 prescribes for booking dimensions, and it will produce a clean table. Compare against our joint measurement sheet format before adopting.

Prompts for calculations (use as a second calculator, never the first)#

5. Steel weight and BBS cross-check. "For the following bar list, compute the cut length per bar and total weight using nominal mass per metre from IS 1786 Table 1: [paste rows: bar mark, diameter, shape, dimensions, number]. Show the mass per metre you used for each diameter and state where hook and bend allowances are assumed. Present as a table."

Tested with 12, 16 and 20 mm bars: it used 0.888, 1.58 and 2.47 kg/m, which match IS 1786 Table 1, and offered the d²/162 shortcut as a check. It applied a 9d hook allowance without being told the bend type, which is the IS 2502 value for a particular hook but not necessarily yours. Our bar bending schedule guide covers the allowances it guesses at.

6. Concrete material take for a nominal mix. "For 25 cum of M20 nominal mix concrete, estimate cement bags, sand and coarse aggregate by the conventional dry-volume method with a 1.54 factor, and separately state what IS 456 Table 9 specifies per 50 kg of cement for M20. Flag that M25 and above require design mix under IS 456 clause 9."

This is where the model misled us. It produced the volumetric 1:1.5:3 figures correctly, then attributed those ratios to IS 456 Table 9. Table 9 is by mass: 250 kg total dry aggregate and 30 litres maximum water per 50 kg cement for M20, and it contains no M25 row at all. The volumetric ratios are trade convention. The answer looked authoritative and was wrong in its citation. Use our concrete calculator for the numbers and the IS 10262 mix design guide for what the code actually says.

7. Brickwork quantity. "Calculate the number of modular bricks (190 × 90 × 90 mm, 10 mm joints) and mortar volume for [wall dimensions and openings], deducting openings as per IS 1200 Part 3 rules, and show every step." It gets the arithmetic right if you give it the deduction rule. If you do not, it applies a made-up threshold. The IS 1200 guide lists the actual deduction thresholds by part, which differ for masonry, formwork and finishing.

8. Unit and rate sanity check. "Here is a BOQ extract [paste item, unit, quantity, rate]. For each line, check whether the unit matches the standard IS 1200 unit for that item, flag any rate that is outside a plausible range for [city] in 2026, and explain each flag in one line. Do not change any values." It is good at catching a slab item measured in sqm that should be cum, or a painting item in cum. Its "plausible range" is a guess, not the DSR; treat the flags as questions to ask, not findings.

Prompts for Excel and data (its strongest engineering use)#

9. Measurement sheet to abstract. Upload the sheet, then: "This is a measurement sheet with columns item, description, nos, length, breadth, depth, quantity. Using data analysis, verify every quantity equals nos × L × B × D, list rows where it does not, then produce an abstract sheet grouped by item with total quantity and unit. Do not round intermediate values." In our test it found two rows where a formula had been overwritten with a typed number, which a visual check had missed. This is the Python data-analysis mode doing real work; OpenAI's own guidance is that it extracts reliably from spreadsheets and unreliably from scanned or image-based tables.

10. Steel reconciliation. "Upload attached: steel received (date, dia, weight) and steel consumed per BBS (element, dia, weight). Reconcile by diameter: received minus consumed minus declared scrap, show variance in kg and percent, and list diameters where variance exceeds 3%." Pair with the material reconciliation format for the IS 1786 rolling-margin tolerances it will not know.

11. DPR from a voice note transcript. "Convert this site update into a daily progress report with sections: manpower by trade, equipment, work done today by location with quantities, materials received, hindrances, tomorrow's plan, and safety observations. Keep every number exactly as given; where a quantity is missing write [NOT REPORTED]." This is the single best productivity prompt for a site engineer, and our DPR format guide has the template it should land in.

The teardown: three prompts that failed and why#

Quoting IS clauses. We asked for the IS 456 clause governing minimum cover for footings. It gave a plausible clause number and a value, delivered with total confidence, and the clause number was wrong. This is structural, not a bug. Copyright in every Indian Standard vests in the Bureau of Indian Standards under section 10(5) of the BIS Act 2016, and reproducing any part requires BIS authorisation. The model therefore has no licensed text of the codes; it reconstructs from forum posts, coaching notes and student blogs, and the reconstruction is fluent. There is no peer-reviewed benchmark of IS clause accuracy that we could find, but the international evidence points the same way: a 2026 study of regulatory hallucination documents models transposing clauses across frameworks and drifting within them. Rule: the model may tell you which code to open; only the code tells you what it says.

Measuring from a drawing image. We uploaded a clean PDF floor plan and asked for wall lengths. On Free and Plus, PDFs get text-only retrieval: the model extracts digital text and discards the images, so it read the title block and nothing else. Uploaded as a PNG, it produced lengths that were confidently wrong, because OpenAI's image FAQ lists exactly the limits a drawing hits: weak precise spatial localisation, confusion between solid, dashed and dotted lines, misread rotated text, approximate counting, and images resized before analysis, which destroys scale. We covered this in depth when we tested whether ChatGPT can prepare a BOQ. If you need quantities from drawings, that is a measured-takeoff job, not a chat job; see what an AI BOQ generator actually does.

Pricing at DSR. Asked to price ten items at CPWD DSR rates, it produced numbers with two decimal places and no edition. Some resembled DSR 2021 figures, none matched DSR 2023 with correction slips. The CPWD DSR guide explains the correction-slip cycle that makes any memorised rate stale. Give the model your rate table as a spreadsheet and it will apply it accurately; ask it to remember rates and it will invent them.

Prompt-writing rules that survived testing#

1. Give it the rule, not the job of knowing the rule. Paste the deduction threshold, the mass per metre, the rate. It applies rules well and recalls them badly.
2. Tag every unverified number. "[CONFIRM]" and "[NOT REPORTED]" instructions convert silent gaps into visible ones.
3. Forbid clause numbers unless you supply them. Ask for "the requirement in general terms".
4. Prefer spreadsheets over images for anything numeric. Text-based inputs scored 87% in the FE study; images scored 42%.
5. Ask for the working, not just the answer. Arithmetic slips inside a correct method are the documented failure mode; you can only catch them if the steps are visible.
6. Use one-shot examples. On FE Civil questions, providing a single worked example recovered 30 of 51 wrong answers in a 2025 study, more than any other prompting technique tested.
7. Never let it certify. The RA bill is Form CPWA 26, signed by named engineers on the strength of the measurement book. The model can draft the covering letter; it cannot stand behind a quantity. Our post on whether AI will replace quantity surveyors explains why that chain stays human.

Where SiteSetu fits#

SiteSetu is not a chatbot and does not compete with one. Where these prompts stop, at measured quantities from drawings and DSR-priced BOQs an engineer can defend, our drawing-to-BOQ service starts: AI takeoff with every quantity traced to its L × B × D against IS 1200, verified by a civil engineer before you see it. Bring your own ChatGPT subscription for the drafting; send us the drawing for the numbers. If you are learning the fundamentals the prompts assume, the free QS course covers measurement, BOQ, rate analysis and RA billing the way sites do them.

FAQs#

Can ChatGPT read AutoCAD DWG files?#

No. OpenAI's supported upload types are XLSX, XLS, CSV, TSV, DOCX, PPTX, PDF and TXT, plus PNG, JPEG and GIF images; DWG and DXF are not supported. Exporting a DWG to PDF gives text-only retrieval on consumer plans, and exporting to PNG gives an image the model cannot measure reliably. For quantities you need a takeoff tool or an engineer, not a chat model.

Is ChatGPT accurate for civil engineering calculations?#

Partially. On Fundamentals of Engineering civil exam questions, current models score above 70% overall and around 87% on text-only questions, but only 42% on diagram-based questions, and structural and statics are the weakest subjects. The typical failure is a correct method with an arithmetic or unit slip, so always ask for the working and check it.

Can ChatGPT quote IS codes correctly?#

Not reliably. Indian Standards are BIS copyright and the model has no licensed text of them, so it reconstructs clause numbers and values from secondary sources and fills gaps confidently. Use it to identify which code and part to open, then read the clause yourself.

Is it safe to upload client drawings to ChatGPT?#

Not by default on Free, Go, Plus or Pro plans, where OpenAI uses conversations and uploaded files for training unless you opt out under Settings, Data Controls. Business and Enterprise plans are not trained on by default. Check your contract's confidentiality clause and anonymise documents before uploading.

Which ChatGPT plan does a site engineer need?#

The free plan covers the drafting prompts, with a limit of three file uploads a day. The Plus plan at ₹1,999 a month as listed on OpenAI's India pricing page adds expanded uploads, the stronger reasoning models and projects, which matter if you use the Excel data-analysis prompts daily. Neither plan changes the drawing-measurement or IS-code limitations.

Can ChatGPT prepare a BOQ from a drawing?#

No. It cannot measure from images accurately, cannot read DWG, and does not know current DSR rates or IS 1200 deduction rules unless you supply them. It can format a BOQ from quantities you give it. We tested this specifically in our post on ChatGPT and BOQ preparation. For the full landscape beyond chatbots — adoption data, prices, and what actually works on Indian sites — see AI in the construction industry in India.