Prompt craft for everyday work

A useful first draft isn’t luck — it’s a prompt that leaves the model nothing to guess.

The idea

A vague prompt forces the model to guess — who it’s for, how long, in what shape, using what facts. Each guess is a place your first draft goes wrong. Good prompting is just removing those guesses, one at a time.

Five ingredients cover almost every work task: state the goal, name the audience, paste the real context, set constraints and output format, and show one example. Then, when it’s close, you re-steer the draft instead of starting over.

How it works

Think of a prompt as a stack of decisions. Every decision you leave out, the model makes for you — usually blandly. Here’s the anatomy, and the exact ambiguity each part removes:

Goal        what to actually do        removes "reply? summarize? translate?"
Audience    who reads it                removes "how formal, how warm?"
Context     the real facts, pasted      removes "which order, what dates?" (stops invention)
Constraints length, rules, what to offer removes "how long? what can I promise?"
Format      the shape of the output     removes "a wall of text, or a subject + two paragraphs?"
Example     one sample in your voice    removes "in whose tone?"

Then iterate: re-steer, don't restart.
  good first draft --> "Good. Now lead with the fix and cut it to three sentences."
  (keeps everything that worked; changes one thing)

Context does the heaviest lifting — it’s the difference between a real order number and an invented one. Show, don’t tell, for voice: one example beats three adjectives like “friendly.”

When to use it

Watch out for

• Telling instead of showing for tone. “Make it friendly” is weak. One pasted example in your actual voice transfers the register far better.
• Skipping the real context. With no order number or dates, the model invents plausible-looking ones. Pasting the facts is what stops confident fiction.
• No output format. Without it you get a shape you didn’t want — an essay when you needed three bullets — and you reformat by hand.
• Restarting on every miss. If the draft is close, re-steer it. Starting fresh drops everything that already worked and re-rolls the parts that were fine.
• Templating a one-off. A reusable template pays off only when the task recurs. For a genuine one-time ask, a good single prompt is enough.

Worked example

An interviewer asks: “How do you use AI tools in your actual work?” A shallow answer says “I ask ChatGPT and clean it up.” A strong one shows method: “For recurring writing — say a delay notice to a customer — I write one prompt with the goal, the audience, the pasted order details, hard constraints (under 120 words, offer the standard credit, don’t promise an earlier date), the output format, and one example in our voice. That gets me a usable first draft. If it’s close but too formal, I re-steer — ‘good, now warmer, lead with the fix’ — rather than starting over. Because that task recurs, I saved the whole thing as a template and just paste fresh context each time.” That answer signals judgment, not just tool use.