All posts

ChatGPT is a brand new hire every time you open it

Why generic AI chat keeps giving you generic answers, and how a few written pages about your business change what it can do.

Most business owners I talk to have tried ChatGPT. They asked it to write a customer email or summarize a contract, got something that looked fine, and then quietly went back to doing the work themselves. The answer was close, but it wasn't how their business does things.

That reaction is correct. And it points at the real problem, which isn't the model.

A brilliant employee with no memory of your business

Think of AI as a very capable new hire. It writes well, reads fast, knows a lot about almost everything and never gets tired. But every time you open a new chat, it's their first day. They don't know your customers, your pricing, your tone, which supplier is always late or why you never quote a job without seeing it first.

You wouldn't hand a first-day employee a complicated job with a one-line instruction and expect it done your way. Yet that's exactly how most people use AI: a short prompt, a generic answer, and disappointment.

The fix is the same one you'd use for a person. Write things down once, and make sure they're read every time.

What "writing it down" looks like

In my own work, every project has a context file. It's a plain document, usually one to three pages, that the AI reads before it starts any task. A typical one covers:

  • Who the business is. What it sells, who buys it, where it operates.
  • The words you use. Your product names, the internal terms staff use, the words you never use with customers.
  • Your rules. "Never promise a delivery date without checking stock." "Always copy the office manager on refunds." "Prices include GST unless stated."
  • Your systems. Where the data lives, which spreadsheet is the real one, what the reports are called.
  • What good looks like. An example of a finished output you were happy with.

None of this is technical. It's the stuff you'd tell a good employee in their first week, if you had the time.

Then turn each recurring job into a procedure

A context file tells the AI who you are. For each job that repeats, I also write a short procedure, which some tools call a "skill". It names the inputs, the steps, the checks and the format of the finished output.

For example, a procedure for a weekly sales summary might say: read these two exports, ignore internal test orders, compare against the same week last year, flag anything that moved more than 15%, and produce a one-page summary in this exact layout.

Once that exists, the job runs the same way every week. It doesn't depend on someone remembering the right prompt, and it doesn't drift when a different person asks.

Why this matters more than which AI you pick

People spend a lot of time comparing ChatGPT, Claude, Gemini and the rest. The differences are real, and I test them carefully. But for most small business work, the gap between a well-briefed model and a poorly briefed one is far bigger than the gap between brands.

A mid-range model with a good context file and a clear procedure will beat the most expensive model working from a one-line prompt, almost every time.

A quick test you can run this week

Pick one task you do every week and have tried with AI before. Then:

  1. Write half a page about your business: who you serve, your tone, two or three rules that always apply.
  2. Write the steps for that one task, the way you'd explain them to a new staff member.
  3. Paste both in before your request, and include one example of a finished version you liked.

Compare the result to what you got before. In my experience the difference is obvious on the first try, and it's the clearest way to see why custom-built AI beats generic chat for real work.

Where it stops being a quick test

Pasting documents into a chat window works for one person and one task. It gets fragile when several people use it, when the rules change, or when the work needs to pull from your actual systems. That's when it's worth setting it up properly: context files that stay current, procedures that are tested, and a review step before anything reaches a customer.

But start with the half page. It's the cheapest improvement you'll make to your AI results, and it costs nothing but an hour of your time.

Have work that needs doing?

Send me the role, the problem, the process or the team you want trained. I'll come back with what I'd deliver, how I'd run it and a fixed fee.