The four parts of a good instruction

What you will learn The techniques from “New to AI”, organised into a system. What the four parts are, and exactly how things go wrong when each is missing.

Why a system is needed

“New to AI” gave you techniques: be specific, give the situation. For personal use that is enough.

But when you hand work to an agent, or several people repeat the same task, techniques are not enough. Agents go just as hard in the wrong direction, and if instructions differ by person, so do the results.

So you need a checklist you can use to confirm nothing is missing.

The four parts

Part	The question it answers	If missing
Goal	what are we making	it makes the wrong thing
Scope	how far may it reach	it touches what it should not
Done condition	how do we know it is finished	people disagree on whether it is done
Context	what already exists	it rebuilds what is already there

One at a time

1. Goal — what are we making

The most obvious one, and the most often vague.

text

✗ "Improve the login"   → "improve" is undefined. Speed? Security? The screen?✓ "Fix the problem where no message appears on failed login"   → both the problem and the resolution are clear

Tip: write the goal as "here is what is currently wrong" rather than "here is what I want to do," and it usually becomes clear.

2. Scope — how far may it reach

When this is missing, good intentions cause incidents. You said improve, so it touches everything that looks related.

text

✓ "Do not change the authentication logic itself — only the on-screen messages"

Scope is most effective written as what not to do. Listing everything to do is hard; the things that must not be touched are few.

3. Done condition — how do we know it is finished

The most frequently omitted, and the slowest to reveal itself when omitted.

text

✗ "Make it work properly"✓ "Done when: wrong password / non-existent account / locked account   each pass their own test"

Without a done condition, a person has to judge "it's done" every time. That eats into the gain from automating.

At organisational scale this part becomes evaluation (evals) directly. An individual can judge by eye; an organisation needs a standard.

4. Context — what already exists

Without it, it reinvents the wheel.

text

✓ "The coupon model already exists in the Coupon class, and discounts use   PriceCalculator. Use those rather than building new ones."

Context also includes reasons not to do something. Tell it "we already tried this and it didn't work" and it deletes that direction.

All four together

text

[Goal]Fix the problem where no message appears on failed login.[Scope]- Do not change the authentication logic- Only the messages shown on screen[Done condition]- Distinguish wrong password / non-existent account / locked account- But do not reveal which one was wrong (prevents account-existence leakage)- Each of the three cases passes its own test[Context]- Message strings are managed in messages/en.json- We were flagged in a security review once for leaking account existence

It looks long; it takes two minutes to write, and it is faster than giving it vaguely and asking three times.

Checklist

Before you send an instruction:

text

□ Goal            did you write what is wrong and what would resolve it□ Scope           did you write what must not be touched□ Done condition  is there a way to judge completion□ Context         did you say what exists and what already failed

If all four are filled in and the result is still bad, it is a different problem — a limit of the model, or context being pushed out.

Common misconceptions

"Am I supposed to write this much every time?"

No. Use it for repeated work and work handed to an agent. A one-off question is fine with the “New to AI” techniques.

For repeated work, saving this format as a template and reusing it is the practical move. That is exactly what a sub-agent's instructions are for.

"How do I set a done condition in advance?"

If you cannot, that work is not ready to be handed over. Hand AI a job whose completion even a person cannot judge, and judging it stays a human job forever.

Check yourself

1. Why is it effective to write scope as "what not to do"?

Answer

Because listing everything to do is hard, while the things that must not be touched are few. Without scope, good intentions reach into everything that looks related and cause incidents.

2. Why does a missing done condition eat into the gain from automation?

Answer

Because a person has to judge "it's done" every time. Production is automatic but judgement is manual, so that much human time remains.

3. All four parts are filled in and the result is still bad. What do you suspect?

Answer

A limit of the model, or context being pushed out. It is not the instruction itself, so look at moving to a new conversation or splitting the work further.

Now some practice rewriting → Rewriting instructions