How to write a good prompt.

Good prompting isn't a formula. It's a habit of noticing what's missing when you ask for something and adding it before you hit send.

The five things worth specifying

**Audience.** Who's this reply for? "Explain X to my board" gets a different reply than "explain X to a technical co-founder" than "explain X to my mom."

**Goal.** What are you trying to accomplish? "Rewrite this email to be more direct so the recipient replies yes/no today" beats "make this better."

**Constraints.** What must be true? Length, tone, forbidden words, required elements. "5 names, one word each, easy to spell, not fintech-sounding."

**Context.** What does the model need to know that isn't in the raw question? "This is a hot path called 1000x/sec — correctness matters more than elegance."

**Format.** Table? Markdown list? Prose? Code fenced with a specific language? Say it — you'll get the format you want on the first try instead of the third.

A working template

You don't need all five every time. Some tasks are self-explanatory. But when a reply keeps missing what you wanted, the fix is usually adding whatever you left implicit.

Rough template that works for most professional tasks:

"[GOAL sentence]. [Context: 1-2 sentences about what matters]. [Constraints: what must/mustn't be there]. [Format: what the output should look like]. [Content: the actual thing]."

Example:

"Draft a client update email. Context: this project is 2 weeks behind schedule because of a vendor delay outside our control. Constraints: professional but human, no jargon, three paragraphs max, ends with a clear next step. Format: subject line + body. Content: the delay is a shipping issue with imported hardware; we've reworked the timeline to catch up by end of month."

What to skip

**Role-play framing.** "You are a senior copywriter with 20 years..." Doesn't reliably work. State what you want.

**Please and thank you.** Not harmful, but doesn't help either. Neutral.

**Overloading with 15 rules.** Split into two prompts if you have that many. The model attends to some and drops others.

**Chain-of-thought unless it matters.** "Think step by step" helps on hard reasoning; on drafting or classification it adds nothing.

When to iterate vs when to accept

If the first reply is close, refine in a follow-up: "Good draft. Cut the second paragraph and make the closing more direct." That's often faster than re-prompting.

If the first reply is off in a way that suggests you didn't specify enough upfront, close the chat, spend 30 seconds on a better prompt, start fresh. The re-prompt is usually better than five follow-up refinements on a bad starting point.

The compounding effect

Users who prompt well get 3-5× more value from the same subscription. Model quality is a ceiling; prompt quality determines how much of that ceiling you use.