AI Prompting for Construction

Here's a pattern almost everybody hits. Tuesday, you ask an AI tool something and the answer is genuinely impressive. You show it to somebody. Wednesday, you ask something similar. And what comes back as vague, generic, and a little bit wrong. So you decide the tool is unreliable and go back to doing it yourself. Video #4 was about what these tools can do. This one is about how to ask, because most of the time the difference between Tuesday and Wednesday isn’t that the tool. It's the request. Fair warning up front. This is the only video in the series about prompting and that is deliberate. It is a real skill and it's worth the next five minutes. It also isn't the thing that decides whether AI works at your company, and I'll come back to that at the end. First, let's start with a bad prompt and fix it in front of you. “Analyze this estimate.” Three words. You'll get something back because these tools always give you something back. It'll be a general description of what's in the estimate, which you already know because you have the estimate. Here's what's missing. Five things, and you can add them in any order. First. Role. Tell it who it should be. “You are an experienced civil construction estimator.” That one sentence changes the vocabulary and assumptions and what the tool treats as important. Two. Context. Tell it the situation. “This is a highway widening job, we self perform the earthwork and paving, and we're bidding it Thursday.” Three. Task. Say what you actually want done. “Review this estimate for potential scope gaps.” Not analyze. Review, for a specific thing. Four. Constraints. Point at where to look, but “focus on labor, equipment, production, assumptions, and missing pay items. Without that, it spreads itself evenly across everything and go shallow on all of it. Lastly, Output. Say what shape you want the answer in. “Return your findings in a table with three columns, the issue, why it matters, and what you'd recommend doing about it. Put those together and the request reads like this. “You are an experienced civil construction estimator. This is a highway widening job and we're self performing and bidding Thursday. Review this estimate for potential scope gaps. Focus on labor, equipment, production assumptions, and missing pay items. Return your findings in a table with the issue, why it matters, and recommended action.” Same tool, same estimate. The first version gives you a paragraph. In my experience, the second answer usually is better than the first, and the fourth is usually better than the second. And this isn't only for estimating. Those same five parts do a different job. “You are a project manager on a water line installation. We're three weeks behind on pipe. Read these daily reports from the last month and tell me where the time actually went. Focus on crew hours, weather delays, and equipment downtime. Give me a short list, ranked by hours lost.” Two habits matter as much as the structure does, and both take about ten seconds. The first is to keep going. Most people treat AI like a search box. You ask once, you get an answer, you're done. But it's a conversation. If the third item in that table looks off, say so. Ask why it flagged it. Ask what it would need in order to be more certain. The second habit is to check what matters. If an answer is actually going to affect a number you put in a bid, verify it against the source before it goes anywhere. That isn't just trust of the tool, it's what you'd actually do with a takeoff from somebody new. Now the honest part. It's a whole reason prompting gets one video here instead of four. A well-built prompt makes a real difference to to what a tool does with information it already has. It does nothing at all about information it doesn't have. You can write the most carefully structured request in the world asking whether your production on Cost Code 3200 is trending below your historical average. If your tool has no access to your production data and no access to your history, the structure of the question changes nothing. Best case, it tells you what it can't answer. Worst case, it makes something up and puts in a beautiful three column table. That's the ceiling on prompting. And it's a low one. Prompting improves the conversation, but it can't create access. So here's what to take away. Don't go looking for magic words. There's no secret phrase. And the ones that seem to work tend to stop working the next time these tools get an upgrade. The durable skill is simpler than that. Tell it who to be, what the situation is, what you want done, where to look, and what shape to give you the answer in. Then keep asking, and check anything that matters. Everything past that comes down to what the tool can actually reach. In the next video, we get concrete. Ten places AI is doing real work in civil construction right now across pre-construction, operations, fleet, field, and the business side