Delegate · Describe · Discern · Take responsibility

What you will learn You have been through your first try and getting the answer you wanted. This chapter is the map that shows where those sit in the whole. There are four moves in working with AI, and the difference between people who use it well and badly is usually not the prompt but the other three.

"Just write good prompts" is not even half of it

Almost everyone who sets out to learn AI starts hunting for prompt tricks. But collect the incidents that actually happen in the field and the order reverses.

The prompt (Describe) is one move out of four. And real incidents mostly come from the third and fourth.

graph TD
  A["1. Delegate<br/>What to hand over"] --> B["2. Describe<br/>How to ask for it"]
  B --> C["3. Discern<br/>How to look at what came back"]
  C -->|"describe again"| B
  C --> D["4. Take responsibility<br/>What happens after it ships"]

1. Delegate — what to hand over

The decisions you make before asking. Is this work suited to AI at all; do I do it with AI assisting, or do I hand the whole thing over?

There are three ways to hand over. Naming them makes meetings tractable.

The same task carries different risk depending on which of the three you choose. Contract review as augmentation means a person looks every time. As automation, nobody looks. Same work, different decision.

• Choosing the work → Three questions for choosing work
• How far to hand over → Deciding how much to hand over

2. Describe — how to ask for it

Getting what is in your head out into the open. The model does not know what you know — your company, this document's readers, what was decided in yesterday's meeting.

More often than not this is a question of thoroughness, not technique. You would have to say this much to a person too.

• The basics → Getting the answer you wanted
• In depth → The four parts of a good instruction
• Why instructions slip → Why instructions do not simply stick

3. Discern — how to look at what came back

The most frequently skipped move, and where most incidents happen.

An AI answer is well written. Well-written text looks correct. So you read it and move on.

There are three questions, and none substitutes for another.

• What to watch for → Things to watch out for
• Invention → Why it is wrong so plausibly
• Leaning → Agreeing with you, and leaning

Describe ↔ Discern is a loop

A good answer on the first try is rare. The way this actually works is discerning what is missing, then describing that back.

Skip discerning and the loop stops turning. You then have a bad answer with no idea why, and no amount of prompt fiddling improves it.

To run it yourself → Rewriting instructions

4. Take responsibility — what happens after it ships

The first three you do even alone. The fourth appears when the work goes to someone else.

Three parts:

None of the three are visible when you work alone. Which is why lifting a personal way of working straight into an organisation snags right here.

• Disclose → Saying what AI made
• Protect → Tools are permissions
• Record → Audit log

Where is your team weak?

Usually the third box is ticked first. And an organisation missing the third is missing the fourth too — if you never look, there is nothing to record.

These four moves are also the skeleton of this whole manual. The machine's side of it continues in The four properties of the machine; the organisation's side continues in the "Why transform" part.

Check

1. Which of the four moves do real incidents mostly come from?

2. Why does distinguishing automation, augmentation, and agency matter?

3. What does "Describe ↔ Discern is a loop" mean?

If the words are the obstacle, start here → Glossary