What vibe coding means
"Vibe coding" is the informal name for building software by describing what you want to an AI model and iterating on the result, rather than writing the code line by line. The term started as a joke about not reading the code at all. For engineering tools, a stricter version works better: let the agent write the code, but you write the specification and you verify the output.
That division of labour suits engineers. Writing JavaScript or Python is not your core skill. Knowing that a bearing outer-race frequency on a 6205 at 1,487 rpm should be about 89 Hz, and noticing when a tool says 178 Hz, is.
Classical baseline · how engineering tools got built
A spreadsheet that grows macros until nobody dares touch it, a request to IT that waits behind higher priorities, or a vendor licence for a feature you need once a quarter. Vibe coding does not remove the need for professional software in critical systems, but it closes the gap for the many small tools that never justified a project.
The build loop
- Idea. Start from a task you understand completely, ideally one you already do by hand. You need to be able to tell a right answer from a wrong one.
- Specify. Write down the inputs (file format, units, sampling rate), the method (formulas, standards, thresholds), and the outputs (charts, tables, a report). The prompt playbook has templates.
- Generate. Give the specification to an AI agent and let it build a first version.
- Run. Test on data where you know the answer: a synthetic signal with a known fault, a hand calculation, a result from a trusted tool.
- Review. Check numbers, units, edge cases and wording. Read the assumptions the agent made. Verify before you trust lists what to check.
- Refine. Report problems precisely ("the envelope spectrum peak is at 2× BPFO, expected 1×"), not vaguely ("it's wrong"). Repeat until the checks pass.
What engineers build this way
Bearing and gear frequencies
Defect frequencies from geometry and speed, with harmonics and sidebands, for a whole bearing list.
Signal analysis apps
Spectrum, envelope and fault rules on uploaded files. The Vibration Analyzer is an example.
CMMS and historian clean-up
Merge work orders with sensor exports, normalise asset tags, flag duplicates, build failure timelines.
Report generators
Turn a week of measurements into a formatted Word or PDF report with charts and findings.
Model baselines
Feature extraction plus Isolation Forest or Random Forest, with honest train and test splits and a confusion matrix.
Fleet overviews
A single page that ranks assets by anomaly score and links each one to its trend chart.
Kinds of AI tools
| Tool type | How you work with it | Best for |
|---|---|---|
| Chat assistants that write and preview code | Describe the tool in a conversation; get a runnable page or script back | Single-file apps, calculators, quick analyses |
| Agentic coding tools | The agent works in a project folder: reads files, writes code, runs it and its tests, fixes errors | Multi-file projects, data pipelines, anything that must run repeatedly |
| Editor and notebook assistants | Suggest and edit code inside your editor or Jupyter notebook | Analysts who already work in Python or MATLAB |
Capabilities change quickly, and your company may approve some tools and not others. Check your organisation's policy before uploading any plant data to an external service.
A first project in an afternoon
A good first project is small, useful and easy to check. For example: a page where you enter bearing geometry and shaft speed and get BPFO, BPFI, BSF and FTF with their first three harmonics.
- Why it is a good start: the formulas are standard, the answer for a common bearing is published, and it fits in one file.
- How you check it: compare with the bearing manufacturer's frequency factors for two or three bearings.
- What you learn: how to specify, how the agent handles units, and how to report a mismatch.
From there, move to tools that read files, then to tools with ML inside. The case study walks through a full analyzer.
Key takeaways
- You write the specification and verify the output; the agent writes the code.
- Start with tasks where you can tell a right answer from a wrong one.
- Test with known answers: synthetic signals, hand calculations, trusted tools.
- Report problems precisely, with expected and actual values.
Next
Build your first prompt with the interactive prompt builder.