# Ask for the same thing three times

> **What you will learn**
> Not a chapter to read but one to **do**. You will ask for **exactly the same
> thing** three times, adding one technique from the last chapter each round, and
> see with your own eyes which technique actually changes the answer.

## What you need

```
□ 20 minutes
□ One piece of writing you actually owe today — an email, a notice, a reply
   · pick one you have not written yet
□ Somewhere to put the three answers side by side (a scratch file is fine)
```

**"Not written yet" matters.** Use something you already wrote and you will mark
the answer that resembles yours too generously.

---

## Round 1 — Bare (5 min)

Use no technique at all. Throw it in **the way you would type a search**.

```
write an apology email to a client about a late delivery
```

### What to look for

```
□ Is it longer than what you would have written?
□ Does it contain phrases you would never use?
□ Could you send it as is?
```

**The third is usually "no."** This is where most people conclude "AI is not
there yet." Round 2 changes that conclusion.

---

## Round 2 — Add the situation (5 min)

Same request, plus **[the situation](/guide/ai-how-to-ask)** and nothing else.
Leave the format alone for now.

```
Write an apology email to a client about a late delivery.

Recipient: a major client of three years; I am on easy terms with the contact
Cause: parts shortage. Not our fault, but it must not read as an excuse
New date: 15 March
Context: this is the first delay we have had with this client
```

### What to look for

| Compared to round 1 | Check |
|---|---|
| Did the tone match the relationship? | □ |
| Does the cause avoid sounding like an excuse? | □ |
| Did the new date make it into a sentence? | □ |

**This is where most of the gain appears.** Of the five techniques, **the
situation alone does about half the work.**

---

## Round 3 — Add the format (5 min)

Now add the format and the constraints.

```
(everything above, unchanged)

Format:
- five sentences maximum
- open with the apology, close by asking them to confirm the new date
- do not use filler like "we will do our very best"
```

### What to look for

```
□ Did it hold to five sentences?
□ Are the first and last sentences what you asked for?
□ Did the banned filler actually disappear?
```

**The third is the point of this round.** Watching a "do not" get obeyed tells
you what is worth writing down next time.

---

## Round 4 — Score them side by side (5 min)

Stack the three answers and mark them.

| | Round 1 | Round 2 | Round 3 |
|---|---|---|---|
| Could you send it as is? | | | |
| How many fixes needed | | | |
| Faster than writing it yourself? | | | |

```
Scoring rule: two fixes or fewer counts as "usable"
```

### One of three conclusions comes out

**It was already usable at round 2** — the common case. Write the situation and
stop; save the format work for documents that are fussy about it.

**It only became usable at round 3** — format mattered for this document. Save
that format block and reuse it verbatim next time.

**None of them were any good** — the problem is not the request, it is **the
task you picked**. Work you lack the information for is work the model lacks it
for too. That is the subject of [things to watch out for](/guide/ai-cautions).

---

## Self-check

```
□ You actually placed the three answers side by side
□ You can say in one sentence what changed between rounds 1 and 2
□ You checked whether the "do not" was obeyed in round 3
□ You saved one fragment of instruction to reuse
```

That last line is the real output of this exercise. **Save the instructions that
worked and use them again** — this is where the gap opens up against people who
start from scratch every time.

---

## Check yourself

**1. If you could use only one of the five techniques, which would it be?**

<details>
<summary>Answer</summary>

**The situation.** Round 2 produced the largest jump. Saying who the reader is
and what position you are in aligns the tone and the content at the same time.
</details>

**2. Why check whether a "do not" was obeyed?**

<details>
<summary>Answer</summary>

**So you know what is worth writing down.** Once you have seen prohibitions
land, you can pre-empt the things that catch you every time — filler, guessing,
overstatement.
</details>

**3. All three rounds were poor. What should you suspect?**

<details>
<summary>Answer</summary>

**The choice of task.** Work that does not improve with better instructions is
usually work you lack the information for. Check the task before polishing the
request further.
</details>

---

Finally, the limits and cautions you must know before using it →
[Things to watch out for](/guide/ai-cautions)