🤖 Part 4: The Anatomy of a Good Prompt

Why the Prompt Does the Heavy Lifting

Ask an assistant "write me an email" and you'll get an email — generic, medium-length, medium-formal, addressed to no one in particular. It's not that the tool did a bad job; it did the best it could with almost nothing to go on. Remember from Part 1 that the model predicts fitting text. When your request is vague, "fitting" means "average" — the blandest, safest response that could follow a prompt that thin. The model fills every gap you leave with a guess, and its guesses are aimed at the middle of the road.

A good prompt simply stops leaving gaps. Every detail you add — who's writing, who's reading, what it's for, how long it should be — narrows the target from "any email" to "this email." That's the whole game. You're not learning secret command words; you're learning to say out loud the things you already know in your head but forgot to mention. The assistant can't read your mind — but it will happily use every clue you give it.

The Four Ingredients

Almost every strong prompt contains four things. You won't always need all four, and they don't have to appear in any special order — but if an answer disappoints you, the fix is nearly always a missing ingredient. Here they are, with the plain-English question each one answers.

Let's take them one at a time, because each one changes the answer in a different way.

🎭 Role — who it should be

Opening with "Act as a…" quietly shifts the whole reply. "Act as a plain-spoken tax preparer" produces different word choices, examples, and assumptions than "Act as a witty copywriter." You're not tricking the tool; you're telling it which of its many voices to use. Roles are especially powerful when you want a particular level — "act as a patient teacher explaining to a beginner" versus "act as a peer reviewer being rigorous."

🎯 Task — the one specific ask

The task is the verb: rewrite, summarize, brainstorm, compare, draft, explain, fix. Keep it to one clear job per prompt when you can. "Summarize this and translate it and turn it into a tweet" often produces a muddle; asking for each in turn produces three good results. Be concrete about the thing itself: not "help with my email" but "rewrite this email to sound warmer."

🗂️ Context — what it needs to know

This is the ingredient people skip most, and it's the one that moves the needle most. Who's the audience? What's the background? What tone fits? What must it avoid? "For my team's Monday stand-up," "for a customer who's already frustrated," "keep it jargon-free for non-engineers" — each phrase steers the answer somewhere generic prompts never reach. Part 5 is devoted entirely to context, because it's that important.

📐 Format — the shape of the answer

If you don't say how you want the output, you'll get the assistant's default — usually flowing paragraphs of medium length. Say the shape and you get it: "as a five-item bulleted list," "in a table with columns for pros and cons," "under 100 words," "in a warm, casual tone," "with a bold heading for each section." Format is the cheapest ingredient to add and one of the most satisfying, because the reply suddenly looks like what you pictured.

Weak vs Strong, Side by Side

Nothing sells this like a direct comparison. Below are three everyday requests, each shown as most people first type it — and then rewritten with the four ingredients. Read them out loud; you can practically hear the answer getting better.

⚠️ Important: The strong prompts aren't harder to write — they're just more honest about what you already wanted. Every extra clause came from a question you were going to ask yourself anyway when the generic answer disappointed you. Front-loading those answers into the prompt saves the back-and-forth. And remember the golden rule from Part 1: even a great answer to a great prompt still deserves a check on anything that matters.

Specificity Beats Politeness

A lot of first-timers agonize over being polite — "Could you please, if it's not too much trouble, maybe help me with…" The assistant doesn't need the courtesy (though there's no harm in it), and manners aren't what improve the answer. Specificity is. The words that change the result are the concrete ones: the audience, the length, the tone, the constraint, the example. "Please" is invisible to the quality of the reply; "under 150 words, for a nervous first-time buyer" is decisive.

🛠️ How To: Rewrite a Vague Prompt Into a Strong One

Best Practices

✅ Do's

• Name the audience. "For a 10-year-old" or "for my boss" reshapes the whole answer in one phrase.
• Ask for a shape. Bullets, a table, a length, a tone — say what you want the output to look like.
• Use one task per prompt. Do one job well, then ask for the next thing in a follow-up.
• Add numbers. "3 options," "under 150 words," "a 5-step plan" — concrete targets get concrete results.

❌ Don'ts

• Don't rely on vague adjectives like "good" or "professional" — say what you mean by them.
• Don't confuse politeness with clarity — "please" is fine, but it's the specifics that change the answer.
• Don't cram five requests into one prompt — you'll usually get a watered-down version of each.
• Don't forget to verify — a well-crafted prompt makes a confident answer, not a guaranteed-correct one.

📓 Learning Journal

📝 Part Summary

🎉 What You Can Now Do

You can look at any request and instantly see what's missing from it — and add it. That's the core skill that separates people who find AI "hit or miss" from people who get useful results almost every time. From here on, when an answer disappoints you, you won't shrug and give up; you'll know exactly which ingredient to add. That small habit compounds into a big difference over hundreds of conversations.

❓ Common Questions

🔭 Up Next

In Part 5: Giving Context & Examples, we'll go deep on the ingredient that matters most — context — and add a powerful new move: showing the assistant an example of what you want so it matches your style exactly, instead of just describing it.

📚 Additional Resources

• OpenAI — Best practices for prompting
• Anthropic — Prompt engineering overview
• Try your prompts in Gemini (Google)