Most bots aren’t broken by the AI.
They’re broken by vague instructions, content written for a web page, and decisions nobody made. Here’s how I find them, and what I hold every assistant to.
Why it’s broken.
Usually it’s one of these.
“Be helpful” isn’t a brief.
Most assistants run on a few vague lines. Nobody told it how long to talk, what it can’t do, or who it’s meant to sound like.
Your help pages weren’t written for chat.
Your help pages were written to be scanned. Fed to an assistant, they come back as walls of text and half-right answers.
Nobody decided.
What does it do when it doesn’t know? When does it hand over? What will it never say? If you didn’t decide, it guesses.
You might already know what’s wrong with your assistant. Knowing how to fix it is the part that takes a writer.
How I judge an assistant.
Every assistant I work on gets held to these.
- Short beats complete.Answer what was asked, then stop. If they want more, they’ll ask.
- “I don’t know” is a good answer.A confident wrong one isn’t. Air Canada had to honour a refund policy its chatbot made up.
- Every reply goes somewhere.Each one ends with something the customer can do next, even if that’s talking to a person.
- If it can’t do it, it doesn’t say it did.No refunds it can’t process, no call-backs nobody will make.
- It sounds like your best staff member.On a good day, talking to a customer they like. That’s the voice to write to.
- Use their words.If they say “the veggie one”, so does the assistant. Nobody asks for “product variant B”.
Send me your assistant. I’ll try to break it, so we can mend it.
It’s free. You get the three problems that matter most and one exchange rewritten.