AI is only as useful as the instructions you give it. That does not mean writing enormous prompts or learning a new technical language. In construction, a good prompt should look much like the instruction you would give a colleague.
Compare “Find the quantities.” with “Calculate all internal wall quantities. Separate them by floor and wall type, deduct openings and create an Excel file with the source reference for each quantity.” The second prompt is better for one simple reason: the task is clearer.
Give the system the context that matters
For most construction tasks, there are five things worth specifying:
- Task. What do you want done?
- Scope. What should be included or excluded?
- Sources. Which drawings, models or documents should be used?
- Checks. What should be verified or cross-checked?
- Output. How should the result be presented?
You do not need all five every time. But the more precise your prompt is, the more accurate and thorough the AI can be. This is especially useful for complex or critical tasks.
For example: “Review the project and calculate all internal wall quantities. Use the latest architectural drawings and wall schedules. Exclude exterior walls and deduct door and window openings. Separate the quantities by floor and wall type. Cross-check the result against the room schedule and flag any inconsistencies. Create an Excel file with the quantity, calculation basis and source reference for every item.” This is a precise prompt that leaves little room for error or misunderstanding.
Specify the scope and source documents
The most important context is usually two things: what should be included, and which documents to read. A construction project can contain thousands of pages across drawings, specifications, BIM models, schedules and revisions, so neither is obvious from the task alone.
An instruction like “Review the specification for deviations.” leaves several questions unanswered. Deviations from what, the standard, the tender documents or the client's own requirements? All trades, or only the ones under contract? Does an editorial difference count, or only something with cost consequences?
For example: “Review the concrete works specification against the referenced standard and list only deviations that affect cost or execution. Use the tender documents and the latest revision of the specifications as the primary source, and ignore editorial differences.”
Define only the details that would change the answer, for example, which document wins when two of them disagree. The same applies to quantities, estimates, risk analysis and other construction tasks.
Ask for something you can actually use
Instead of receiving a long explanation and restructuring it manually afterwards, tell the system what you need. Here are some examples we like from our customers:
“Create an Excel file with floor, wall type, length, height, gross area, deductions, net area and source reference.”
“Give me a short summary first, followed by all identified risks ranked as high, medium or low.”
“Structure the findings by trade and create a separate section for questions that should be clarified with the client.” The closer the output is to how you actually work, the more useful and efficient the AI becomes.
Make the result easy to verify
For important quantities, requirements or risks, ask the system to show where the information came from. For example: “Include the drawing number, page or model reference behind every quantity.” Or ask: “If you identify conflicting information, reference both sources and explain the difference.”
You can also ask the system to cross-check its work: “Compare the architectural drawings against the room schedule and flag differences you cannot explain.”
And if the documentation is unclear: “Do not guess. Flag anything that cannot be determined confidently and explain what information is missing.” We are working to make our AI flag uncertainty and missing information by default as part of our governance work. This makes the result easier for the person responsible for the project to review.
And when the answer misses the mark, you rarely need a better prompt. You need one more sentence. Say what was wrong or what was missing and run it again, the same way you would correct a colleague who misunderstood the first time.
You do not need perfect prompts
Prompting should not become another complicated part of the job. You already know how to give good instructions. You explain the task, define the scope, point someone towards the relevant information and tell them what you expect back. Using AI works the same way.
Start simple and add context where it matters. Clear instructions help you get useful work done faster.