Sprint 1 · ~11 min · includes a hands-on task on your own work
What AI is (and isn't), in plumber's terms
If you think the model 'knows' code and specs, it'll burn you. Knowing what it actually is — pattern prediction, not a rulebook — is the difference between leverage and a confident wrong answer in front of a customer.
The idea
An LLM like ChatGPT is a prediction engine: it guesses the next word based on patterns it learned, not a database of verified truth. That's why it's genuinely good at language-shaped work — drafting a customer email, summarizing a long spec, reformatting your messy notes, turning a voicemail into a clean message — and genuinely unreliable on exact facts: code sections, model numbers, dimensions, prices, and anything new. The dangerous part is it sounds equally confident either way; a made-up IPC section reads exactly like a real one. Meet it where you already are: open ChatGPT (free) and treat it like a sharp apprentice — fast and useful for drafts, never the authority on anything load-bearing. Two habits start here and carry through everything: verify any fact before you act on it, and never paste customer or money data into a public model.
Walkthrough
- 1Open ChatGPT (the free version is fine) and ask it something you already know cold — a code requirement or a spec on a job you just did.
- 2Notice where it's confidently wrong or vague — that's the bluff zone you'll learn to catch.
- 3Now ask it to do a language task instead: 'rewrite this rough note to a customer in plain, polite English' — notice how much better it is at that.
- 4Write down one task from your week that's language-shaped (drafting, summarizing, reformatting) — that's your first safe AI win.
Worked example
A new HVAC tech wants to see the difference between asking AI for a fact versus asking it to draft. They run both on something they already know.
The prompt
I'm an HVAC tech. Answer two ways so I can compare: 1) What is the minimum clearance to combustibles for a typical residential gas furnace vent? (I'll check this against the manufacturer's installation manual and code — just give your best answer.) 2) Rewrite this rough customer note in plain, polite English: "furnace short cycling, found dirty flame sensor, cleaned it, running fine now, recommend they replace 16x25 filter monthly."
What good output looks like
1) Clearances vary by appliance and vent type — commonly noted as 1 inch for B-vent, but this is exactly the kind of number you must confirm in the unit's installation manual and local code. Do not rely on this figure. 2) "Hi [Name] — your furnace was short-cycling. We found a dirty flame sensor, cleaned it, and the system is running normally now. To keep it running well, we recommend replacing the 16x25 filter monthly. Let us know if you'd like us to set a reminder."
Where it burns you
- Treating a confident answer as a correct one — fluent and right are not the same thing with an LLM.
- Asking it for a code section or model number and pasting that into a quote or work order without opening the actual document.
- Pasting a customer's name, address, or card info into a public model just to 'clean up' a note — strip that out first.
Do it now
In ChatGPT, ask it one thing you already know the answer to (a code or spec) and watch where it bluffs. Then run one real rough note through it and see how clean the rewrite comes back.
Before it touches a price — verify
- Did I confirm whether I'm asking for a FACT (verify it) or a DRAFT (review it)?
- If it gave me a code/spec/model number, did I trace it to the manual or code before using it?
- Did I keep customer PII and money data out of the prompt?
- Am I treating this as an apprentice's draft, not the final word?
You should now have: A gut feel for where AI helps (language work) vs. where it bluffs (facts) — and the two habits that keep you safe from day one.
Guardrail: AI gives you a draft and a second opinion — never the final word on code, safety, pricing, or contracts. Verify facts; keep customer and money data out of public models.