Write clear instructions

A clear instruction says what to do, for whom, and what good looks like — in words a colleague could follow without asking you a single question. Most “the AI misunderstood me” stories are not model problems; they are instructions missing load-bearing walls.

The two failure modes

• Vagueness. “Make it better.” Better how? The model substitutes its guess for your intent.
• Overload. “Shorten this, add statistics, keep the tone, make it professional and maybe more exciting.” Five requests in one sentence become five half-done requests.

The cure for both is the same: one job per instruction, stated so a stranger could check it.

The specificity ladder

Watch one request climb:

1. “Improve this paragraph.” — no direction at all.
2. “Shorten this paragraph.” — a direction, no target.
3. “Cut this paragraph to 120 words.” — countable.
4. “Cut this paragraph to 120 words; keep every commitment and the apology; add no new facts.” — countable, bounded, checkable.

Climb one rung more than feels necessary, and your results stop being lotteries.

Why it works

Instructions narrow the space of plausible tokens — that is the only mechanism there is. Each specific, checkable requirement removes a swath of outcomes you did not want. Vague words like “professional” remove almost nothing; “apologetic but confident, five sentences” removes almost everything except the target.

A bad example

Ask: “Help me write an email to my boss about the project.”

The model knows the topic, not the situation: status update or request? Good news or bad? Formal or casual? You will spend three rounds teaching it what one sentence could have said.

A better example

Ask: “Write a five-sentence update email to my manager about Project Atlas. The integration milestone slipped to October 4; everything else is on track; I need two more days of QA support. Direct, solution-oriented, no jargon. End with the ask.”

One job, fully specified. The reply is checkable line by line against the brief.

One last test before sending: read your instruction as a skeptic. “Summarize briefly” — how would anyone prove you wrong? “Summarize in five bullets, one per section, no adjectives” — now the output either matches or it does not, and the next round tells you which word to tighten. Instructions you cannot audit are wishes; the auditability is what turns them into instructions.

Practice

Transfer

• Any “improve this” request: replace the verb with a countable one.
• Multi-part wishes: split into separate instructions — they are separate decisions.
• The stranger test before sending: would someone with zero context produce the right thing?

Next

Next: Add constraints — bounding length, scope and exclusions deliberately. The prompt glossary entry covers the basics.