Vibe Coding for engineers: build your own BIM tools without being a programmer

The AI amplifies whatever you give it. If your idea is fuzzy, the result will be fuzzy too.

Before opening Claude or ChatGPT, write on a sheet of paper:

• What specific task you want to eliminate. Not "automate the modeling," but "rename all alignments following the format AL-<level>-<sequence>".
• What the input is. What does the user select? The whole drawing? Only what's selected?
• What the expected output is. Does it modify the model? Generate a file? Show a report?
• What it should NOT do. The explicit boundary. "Don't touch alignments that already match the format."

The most common mistake: asking for code without saying which version and which tool you're using.

A context message you should reuse in every session:

Context for my environment:
- Civil 3D 2025 (.NET 8)
- Windows 11, Visual Studio Community 2022
- Goal: C# add-in using the official API
- Level: I have never coded in C#, explain what you do

When you propose code, also tell me:
- What each block does, in plain language
- What I need to install or reference
- How I test it

Without this, the AI assumes some random version and gives you code that won't compile. With it, it gets things right on the first try far more often.

Ask for the skeleton first: what files, what folders, what each part does. Only then move into detail.

An order that works:

1. Project structure — folders and empty files, with the responsibility of each one.
2. The simplest case — that does ONE minimal thing, and does it well.
3. The real cases — now your actual business logic, with its exceptions.
4. Error handling — what happens if nothing is found, if the data comes back empty.
5. Interface — button, window, options. Last, not first.

This order comes from the layered architecture we cover in detail in building a Civil 3D plugin with Claude Code.

One change at a time. Verify each result before asking for the next.

The right loop:

1. You ask for one small thing.
2. You read what it proposes (even if you don't understand every line, understand the intent).
3. You test it on a copy of a small model.
4. If it works: save the progress. If not: give it the complete error, not a summary.
5. Next small thing.

When something fails: don't ask it to "fix it." Paste the exact error message, say what you expected and what happened instead. The difference in response quality is enormous.

You don't need to write code from memory. You do need to understand what the code in front of you does.

The minimum viable knowledge so you don't depend blindly on the AI:

• Variables and types — what each thing stores.
• Conditionals and loops — how it decides and how it repeats.
• Functions — how logic gets organized and reused.
• Specifically in BIM: what a Transaction is (the block that groups changes to the model and lets you undo them) and why every change goes inside one.

If you're missing those fundamentals, the shortest path is our guide on learning to code from scratch for civil engineers. It takes a few weeks, and it multiplies what you can do with AI.

This is where Vibe Coding alone stops being enough, and you need to add discipline.

What separates an experiment from a real tool:

• Version control. Install GitHub Desktop, create a repository, make a "commit" every time something works. It's your infinite backup. 20 minutes to learn, permanent value.
• Decent error handling. Show "No alignments found in the drawing" instead of crashing with unreadable red text.
• Tested with odd data. Empty model, 10,000 elements, names with accents and emojis, incomplete data.
• Minimal documentation. Three lines: what it does, how to install it, when NOT to use it.
• Distribution. A loose file works for you. For the team, package it as a bundle.

If the tool is going to be critical to the operation, it's time to decide between maintaining it yourself or commissioning it. The factors that matter are in how much a Revit/Civil 3D add-in costs.