The short version: ChatGPT is good at structuring engineering problems, generating hypotheses, and doing the boring documentation around the work, but it is unreliable at the maths itself, so your prompts need to separate “help me think” from “check my numbers”. Give it the constraints, the failure mode you are worried about, and a required format for the answer, and it will save you hours. Skip any of those three things and you get plausible sounding nonsense.
Worth reading next: What Is Prompt Engineering, and Does a Marketing Team Need to Learn It.
Why engineering prompts fail in a different way than marketing ones
I write a lot about prompting for marketing, content, and admin work. Those prompts fail quietly. You get a bland blog post or a generic email and you just rewrite it. Engineering prompts fail loudly, and sometimes they fail in a way that looks completely correct until someone builds the thing.
I learned this through a client, not a textbook. A man called Dave runs a small mechanical engineering firm outside Birmingham, eleven staff, they design conveyor and handling systems for warehouses. He came to me for help with his marketing, not his engineering, but three months into working together he started showing me the ChatGPT prompts his design team had been using to speed up tolerance checks and material selection notes. Some of them were sharp. One was dangerous.
The dangerous one asked ChatGPT to “calculate the safety factor for this steel bracket given the load and dimensions below” and then dropped in six numbers. ChatGPT gave back a safety factor, formatted neatly, looking exactly like something an engineer would write. It was wrong by roughly 30 percent, because it had silently used the wrong yield strength assumption for the steel grade mentioned in passing three sentences earlier. Nobody caught it for two weeks because the output looked so tidy.
That is the thing nobody selling “perfect prompt” courses wants to say out loud: a well-structured prompt makes ChatGPT’s wrong answers look more convincing, not less. Structure improves clarity of output. It does not improve the accuracy of the calculation underneath it. If you want a useful piece on prompt structure that does not oversell what structure can fix, I wrote about what makes a ChatGPT prompt truly awesome, and the honest limits of that advice apply here more than anywhere else.
The five things every engineering prompt needs
After that scare, Dave’s team and I rebuilt how they used ChatGPT for problem solving. Not for calculations, for the thinking around calculations: narrowing down failure modes, comparing material options, drafting the reasoning section of a report, checking whether they had missed an obvious edge case. Every prompt that worked had these five parts.
- The role and the boundary. Tell it what kind of engineer to think like, and tell it explicitly not to trust its own arithmetic. Something like “think like a mechanical design engineer reviewing this problem, but flag any point where you are doing a calculation rather than reasoning, because I will verify those separately.”
- The constraints, all of them, in one place. Materials, dimensions, tolerances, environment (temperature, moisture, vibration), load type (static or cyclic), and the standard you are working to (BS, ISO, ASME, whatever applies). Vague constraints produce vague or wrong answers. This is the single biggest difference between a prompt that helps and one that wastes your afternoon.
- The failure mode you are worried about, named specifically. Not “is this safe” but “am I at risk of fatigue failure at the weld given roughly 40,000 load cycles a year” or “could thermal expansion cause binding in this fit at 60 degrees C”. Naming the failure mode forces ChatGPT to reason toward something specific instead of giving a generic reassurance.
- A required output format. Ask for assumptions listed separately from conclusions, and ask it to flag anything it is uncertain about, rather than presenting everything with the same confident tone. This one change cut Dave’s team’s review time roughly in half, because the uncertain bits were no longer buried inside confident-sounding sentences.
- A demand for alternatives, not just an answer. Ask for two or three approaches to the problem, not one. Engineering has trade-offs baked in, and a single answer hides the trade-off. Three answers show you the trade-off, which is usually the actual decision you need to make.
A real prompt, before and after
Here is roughly what one of Dave’s engineers was writing before we fixed anything:
“I need a bracket to hold a 50kg load, what material should I use?”
That gives ChatGPT almost nothing to work with, and it will answer anyway, because it always answers. It will probably say aluminium or mild steel and sound sure about it. Here is the version we built instead, which took about ten minutes to draft once and then got reused for every similar problem:
“Act as a mechanical design engineer reviewing a bracket design for a warehouse conveyor support. Constraints: static load of 50kg plus a dynamic factor of 1.5 for occasional impact, bracket length 200mm, ambient environment is a cold storage warehouse at 2 to 4 degrees C with high humidity, budget favours mild steel unless there is a strong corrosion argument for stainless. Failure mode I am concerned about: corrosion-driven section loss over a 10 year service life, not overload. Give me two material options with the trade-offs between them, list any assumptions you are making separately from the recommendation, and flag clearly if any part of your answer involves a calculation rather than a judgement call, since I will verify calculations independently.”
That prompt does not calculate anything for you, and it should not. What it does is force a structured comparison, surface the corrosion question Dave’s junior engineer had forgotten to consider, and separate the parts of the answer worth trusting from the parts worth checking. That is the realistic ceiling for ChatGPT on engineering problems right now, and treating it as anything higher than that is where people get burned.
Step by step: building your own engineering prompt
- Write down the constraints first, in a plain list, before you touch ChatGPT at all. Materials, dimensions, environment, standard, load type. If you cannot fill this list, you are not ready to prompt, you are ready to go back to the drawings.
- Name the specific failure mode or decision you are trying to resolve. “Is this okay” is not a prompt, it is a wish. “Will this weld survive 40,000 fatigue cycles at this stress range” is a prompt.
- Ask for the role, the reasoning, and a required format in the same message: assumptions separate from conclusions, calculations flagged as calculations, at least two options where a trade-off exists.
- Read the output looking specifically for the words “assume” or “assuming”. Every assumption ChatGPT makes silently is a place your real project can diverge from its answer. Pull those out and check them against your actual project.
- Take any number it gives you, treat it as a hypothesis, and verify it against a calculator, a standard, a colleague, or software built for the job. Never carry a ChatGPT number straight into a spec sheet or a client report.
- Feed back what you found. “The assumption about yield strength was wrong, the actual grade is S355, redo the comparison” gets you a better second pass, because now the constraint list is accurate.
Where this earns its keep
The engineering firms getting real value out of this are not using ChatGPT to replace calculation, they are using it to compress the thinking and writing that surrounds calculation, which is often the slower part of the job. Dave’s team now uses structured prompts to draft the reasoning sections of client reports, to generate a checklist of failure modes they might have missed before a design review, and to compare regulatory language across BS and ISO standards when a client asks which one applies. None of that touches the actual numbers. All of it used to take a junior engineer two or three hours a report. It now takes about 40 minutes, with the calculation verification done separately in the tools they already trust.
That is not a small saving. Across eleven staff doing a handful of reports a week, that is roughly ten to fifteen hours of engineering time back every week, hours that get spent on client-facing work instead of formatting. If you want the wider case for what AI can quietly take off a small team’s plate, including the admin side that eats just as much time, I have written about using AI for invoicing and admin which applies just as well to an engineering office as it does to a marketing one, the principle of automating the paperwork around the expert work is identical.
The mistake almost everyone makes
The single biggest mistake I see, and I saw it in Dave’s team before we fixed it, is asking ChatGPT one giant prompt that mixes “help me think about this” with “give me the final number”, then treating the whole reply with equal trust because it arrived in one confident block of text. Split those two jobs on purpose. Use ChatGPT to widen the thinking, narrow the options, and draft the words. Use your existing calculation tools, standards, and a second pair of human eyes for the number that gets built. If you find yourself pasting a ChatGPT figure straight into a design pack without a separate verification step, that is not a prompting problem you can fix with better wording, that is a process problem, and it is worth getting outside eyes on it before it costs you a project. This is exactly the kind of gap a proper AI consultant for small business should be closing with you, building the verification step into the workflow rather than leaving it to memory.
The uncomfortable bit, the part most advice on this topic avoids because it makes the tool sound less impressive, is this: the better your prompt, the more convincing the wrong answer looks. A vague prompt gets a vague, obviously hedgy answer that nobody trusts too much. A tight, well-constrained, professionally formatted prompt gets back something that reads exactly like a competent engineer wrote it, assumptions and all, and that is precisely when people stop checking. Good prompting does not reduce your need to verify. It increases it, because it removes the visible signs of uncertainty that used to remind you to check.
A quick checklist before you send any engineering prompt
- Have I listed every material, dimension, environmental factor, and standard, not just the ones that felt obviously relevant?
- Have I named the specific failure mode, not just asked “is this okay”?
- Have I asked for assumptions and calculations to be flagged separately from the recommendation?
- Have I asked for at least two options where a real trade-off exists?
- Do I have a plan to verify every number independently before it goes anywhere near a client or a build?
Related reading: funny chat gpt prompts.
Related reading: fun prompts to ask chatgpt.
Frequently asked questions
Can ChatGPT do engineering calculations correctly?
Sometimes, but not reliably enough to trust without checking. It tends to get the method right and silently substitute a wrong assumption, like the wrong material grade or yield strength, which produces a confident, well-formatted, wrong number. Treat every calculation as a hypothesis to verify, never as a final answer.
What is the single most important thing to include in an engineering prompt?
The full list of constraints, given all at once, in plain terms: materials, dimensions, environment, load type, and the relevant standard. Most bad answers come from a prompt that left one constraint out, not from a badly worded question.
Should I use ChatGPT for safety-critical calculations?
No, not as the final source. Use it to structure the problem, generate options, and draft the reasoning, then verify every number with proper calculation tools, standards, and human review before it touches anything safety-critical.
How do I stop ChatGPT from sounding more confident than it should?
Ask it directly, in the prompt, to flag any point where it is calculating rather than reasoning, and to list assumptions separately from conclusions. This does not fix accuracy, but it makes the uncertain parts visible instead of buried in confident-sounding sentences.
