From the book · Unleash Your Business With AI
Chapter 2: What AI is good at, what it is bad at, and how to tell
The 2 prompts from Chapter 2, ready to copy, with the chapter's mistakes to avoid and three tasks for this week.
A five-question test for any task
Before you do any work, help me decide whether this task suits you. The task: [describe it, e.g. "draft replies to retailers asking about delivery of pending orders"] What I can give you: [e.g. "the pending orders list with customer names replaced by codes, our dispatch days, two replies we were happy with"] Who will check your output: [role, e.g. "dispatch in-charge"] Do not start the task yet. Answer these four points in short bullets: 1. Which parts of this task are language or pattern work that you can do from what I give you. 2. Which facts you would need that I have not listed. Be specific. 3. Where you would have to guess if I do not give you those facts. 4. Any calculation involved, and how I can verify it myself (for example, an Excel formula I can run). Rules: be plain about your limits. Do not say you can do something you would have to guess at. Do not invent anything about my business, my products or my customers.
What you should get back: four short lists. The second and third are the valuable ones. They tell you what to paste in before you begin.
Check: if the list of missing facts is empty for a task that involves rates, stock or policy, do not believe it. Ask again: "What would you need to know about my business to do this without guessing?"
That's the first one
Open the rest, free
Enter your email once and every prompt page and guide on this site opens.
- A five-question test for any task (2)
- Mistakes to avoid
- Do this week
A five-question test for any task (2)
Answer the question below. Then go back through your answer and tag every factual statement with one of three labels: [GIVEN] it comes from the information I pasted in this chat. [GENERAL] it is general knowledge that I should verify before relying on it. [GUESS] you are not sure, or you filled a gap. Do not delete a statement to avoid tagging it. If you assumed something about my business, say what you assumed. Do not invent sources, figures, names, laws or case references. If you do not know, write "not known". Question: [your question] Information: [paste the facts, with customer names, phone numbers and addresses removed]
What you should get back: the answer, with a label after each statement.
Check: the labels are a help, not a guarantee. AI can label a guess as given. Compare every [GIVEN] line with your own document, and treat every [GENERAL] line on tax, law or compliance as a question for your CA or lawyer.
Mistakes to avoid
- Judging AI from one answer. One impressive reply or one silly mistake tells you about that task, not about AI.
- Trusting the feeling of speed. The developers in the METR study felt faster and were slower [5]. Time the task, including the checking.
- Asking for facts you did not supply. If your price list is not in the chat, any rate in the answer is invented.
- Letting AI do the totals. Use it to organise and explain numbers. Do the calculation in Excel or Tally.
- Taking big advice on thin inputs. The Kenyan owners whose revenue fell had acted on generic advice about hard problems [7]. Give AI your real figures and constraints, and treat its answer as one opinion.
- Giving the checking to whoever is quickest with the phone. The checker has to know the work. Chapter 19 explains who should learn first.
Do this week
- Write down three tasks from your own week and run each through the five questions on paper. Mark each one: draft with AI, supply facts first, or keep with a person.
- Pick one of the three and run the first prompt in this chapter before giving AI the task. Paste in whatever it says is missing.
- Time one task twice: once the usual way, once with AI, counting the minutes spent checking and correcting. Write both numbers down.