Back to Resource Guide.

Stop Asking AI to "Review Drawings": Why Your Prompts Are Creating Change Orders

You’ve probably done it.

You open up a public AI chatbot, upload a PDF of a complex structural or architectural detail, and type: "Find any mistakes or clashes on this drawing." The AI thinks for fifteen seconds, then spits back a beautifully structured, incredibly polite response: "The layout appears complete and compliant. All elevations are clearly labeled."

Two weeks later, your drywall sub points out that a structural column is slicing directly through a mechanical chase on Sheet M-102. You are looking at a $15,000 change order and a three-day delay. You stare at your screen and wonder: Why did the machine lie to me? I thought AI was supposed to be smart.

As someone who started their career swinging a hammer as a carpenter, finished a Bachelor's in Structural Engineering, and managed risk as a construction engineer on mega-projects before moving to ConTech (Vico Software, acquired by Trimble) and enterprise software (Anaplan through its IPO and Thoma Bravo acquisition)—let me tell you the cold, unhyped truth: General AI models do not "read" blueprints. They predict words.

The Anatomy of a Prompt Hallucination

When you upload a drawing PDF to a general LLM, the model doesn't build a 3D physical model of a building in its "mind." It converts the PDF into flat pixels, translates the text it can find via basic Optical Character Recognition (OCR), and then calculates the mathematical probability of what words should come next.

Because these models are optimized to be "helpful" and "agreeable," a vague prompt like "Are there any errors?" forces the model to default to compliance. It looks at the massive, noisy grid of lines and text, gets overwhelmed, and says: "Looks great, boss!"

When we secured our funding from Mark Cuban, we did so on a single, core premise: Construction professionals don't need conversational partners; they need deterministic outcomes. If you are going to use public AI tools to check drawings during estimating, you have to stop writing conversational prompts. You have to write Structural Prompts.

How to Write a Structural Prompt: The Three Rules

If you want to catch discrepancies before concrete cures, use this exact prompting blueprint:

1. Isolate the Dataset (Compare, Don't Analyze)

General AI, like OpenAI, Gemini or Claude, are mediocre at open-ended analysis, but are excellent at spot-the-difference. Do not ask it if a plan is "correct." Instead, give it two specific schedules and ask for a cross-reference.

  • The Wrong Prompt: "Does my door schedule on Sheet A-601 look right?"
  • The Structural Prompt: "Act as a strict pre-construction QA/QC manager. Cross-reference the 'Width' column in the Door Schedule on Sheet A-601 with the dimension annotations on Sheet A-102. Identify any door IDs where the schedule width is larger than the wall opening width. Present your findings in a clean markdown table: Door ID | Schedule Width | Plan Opening Width

2. Establish a Hard Negative Guardrail

Because LLMs are designed to be agreeable, they will often ignore small, ambiguous errors. You must explicitly give them permission to find nothing—or tell them how to handle incomplete data.

  • Add this line to every prompt: "If you cannot find a corresponding Door ID on Sheet A-102 for a door listed in the Sheet A-601 schedule, do not guess. List it under a dedicated table titled 'Unverified Openings' and state 'Missing from plan view'."

3. Use "Few-Shot" Examples

Show the machine exactly what a successful output looks like. If you want it to check sheet indexes against actual drawings, feed it a tiny example of an error first:

Example Input:
Index reads: Sheet S-101 - Foundation Plan Detail
Actual Sheet Title Block: Sheet S-101 - Foundation Plan Section

Example Output:
Discrepancy: Sheet S-101 title mismatch ("Foundation Plan Detail" vs "Foundation Plan Section").

The Friction Problem: Why Manual Prompting is a Trap

Writing structural prompts is highly satisfying, but let’s be honest: who has the time?

If you're a pre-con manager trying to bid four projects at once, you do not have the time to copy-paste coordinate grids, write complex prompt scripts, and upload massive 500MB PDFs sheet-by-sheet into a browser window. Furthermore, public AI models train on your data, which means uploading your proprietary client drawings is an immediate intellectual property liability.

This is why we built MarkedUp.ai to be completely invisible layer between where you store drawings and where you view drawings. We automate the audit your team doesn’t have time to do dramatically increasing the risk of losing a job, or worse, winning a job that is full of unknown risks.

We believe the best software is the software you never have to open. Our system integrates directly with your existing SharePoint, Procore, Bluebeam, or native network folders. You simply upload your unmarked drawings to your existing platform, our backend pipeline silently runs advanced, orchestrated checks, and it drops a fully redlined PDF back into your folder.

No new logins. No software fatigue. Just absolute risk protection, validated by our world-class engineering team and backed by Mark Cuban.

Stop prompting. Start engineering your pre-con pipelines.