Foundations 5 min

Writing prompts that work

The shape of a good prompt. Iterate; don't restart. The one skill that compounds.

Every AI tool eats prompts. The same model gives very different results depending on how you write the request. This is the one skill that transfers between products and compounds over time.

The four-part shape

Part	What goes in it
Goal	What "done" looks like, in one sentence.
Constraints	What it must / must not do. Length, tone, files to not touch.
Verify	How you'll know it's right. "Tests pass." "Renders on mobile."
Clarify	"Ask one question if anything is missing." Forces the model to fill its own gaps.

Side by side

Weak

"Write me an email about the launch."

Goal vague (launch of what?)
No audience, tone, length
No way to know if it's good

Strong

"Draft a 4-sentence internal email announcing the Q4 launch (a self-serve onboarding flow). Audience: engineering team. Tone: warm, direct, no marketing language. End with one concrete ask: read the spec by Friday. Ask one question if anything is missing."

Iterate; don't restart

First draft will be close-but-wrong. Don't rewrite the prompt — react to the output. These three phrases will rescue 80% of bad outputs:

"More direct. Drop the corporate language."
"Same content, half as long."
"What's a different angle you could try?"

Tactics that move the needle

Few-shot

Paste 2–3 examples of "good" before your request. Models match the pattern.

Role priming

"You are a senior editor with 20 years at The Economist." Still works in 2026.

Show the format

Don't describe the format — give a one-line template. "Reply as: Decision: ___ · Why: ___"

Force "I don't know"

"If you don't have enough information, say so. Don't guess." Cuts hallucinations.

Try it

Send the weak version. Then send the strong version. Notice the difference.

Claude New chat

What can I help with?

This is a faithful imitation. Click a suggestion or type your own message.

The compounding skill. This pattern works the same in claude.ai, ChatGPT, Cursor, Cowork, and any open-source agent. Get good at it once; it pays off forever.

Do I need 'please' and 'thank you'?

Doesn't hurt; doesn't really help. Models don't have feelings to flatter. But politeness sometimes nudges them toward a friendlier register — useful if you're drafting something that should sound human.

What about chain-of-thought ('think step by step')?

Less needed in modern models — they reason by default. Still worth it for hard math, code review, or any task where you want to see the working before the answer. Phrase: "Reason out loud first, then give the final answer."

How long is too long for a prompt?

For chat: a few paragraphs is fine. For agents: longer is OK if it includes real reference material. The model's "attention" weakens at extreme lengths — if you're past 5,000 words and still elaborating, split into two prompts or attach a file instead.