Pick your first AI build: a job you already know cold

AI can turn a description into working software fast. It can't tell you what your business needs. If you don't know the job, you can't tell when the result is nonsense, so AI builds the wrong thing, really, really fast.

For years I knew exactly what our systems needed. I just couldn't code them. AI fixed the coding part. My experience didn't come with the subscription.

1. Pick a job you know cold. Something you've done yourself or watch done every week: booking an inspection, sending an estimate, chasing a no-show. Not "the whole business." One job.
2. Name the annoying part. One sentence: what goes wrong, and who it bugs. "The dispatcher retypes every web request" is good. "Operations are inefficient" isn't.
3. Write what "better" looks like. What would that person see or do differently? This sentence is your test later, so make it something you can check.
4. Describe it like you're training a new hire. Who uses it, what goes in, what comes out, what must never happen. Ask the AI to ask you questions before it builds. The prompt is below.
5. Build a small version and try it with fake data. Run real-looking cases through it. Because you know the job, you'll spot the wrong answers. That's the whole point.
6. Get help before it touches real customers. I still make mistakes. I still get help. Have someone who knows the tech look at anything that handles customer details or money before it goes live.

Paste this into the AI tool you build with and fill in the brackets:

I run a [roofing / plumbing / HVAC] company. I want to build a small tool for one job.

The job today, step by step: [...]
Who does it (job title): [...]
The annoying part: [...]
Better looks like: [what they'd see or do differently]
What goes in: [the info we already have]
What comes out: [what the next person needs]
Never: [what must not happen, e.g. a customer's details going to the wrong person]

Before you build anything, ask me about anything unclear.
Then build the smallest version I can try with fake data.

Bayside Plumbing's dispatcher, Marco, retypes every web request onto the schedule. The annoying part: requests never say how old the water heater is, so techs sometimes show up with the wrong parts.

Better looks like: every request arrives with the heater's age and type, or a clear flag that it's missing. The owner fills in the prompt, answers the AI's questions, and gets a small intake form that sends Marco a summary.

They try it with five fake requests. On the third, the form treats a tankless heater like a tank and asks for the tank size. Marco catches it in a second, because he knows the job. Someone who didn't would have shipped it.

• Picking a job because it sounds impressive. If you can't do the job yourself, you can't check the build.
• Building the whole system first. Small version, try it, then grow it.
• Trusting it because it runs. Running and right aren't the same thing.
• Measuring yourself against someone else's speed. A fast build from someone who's done the job for years isn't a beginner benchmark. Mine included.

You don't need my background. You just need to know the job. Otherwise AI builds the wrong thing, really, really fast.