How to Prompt Claude AI Effectively: The Method We Teach

Here's how to prompt Claude AI effectively in one sentence: give it context (the background and material), a specific task with constraints, the format you want back, an example of what good looks like, and then iterate on the first draft instead of accepting it. That five-part structure is the entire method we teach in our class, and it fixes the number-one complaint we hear from students: "Claude's answers feel generic." They feel generic because the prompts were. This article walks through each part with real before-and-after rewrites.

Why most prompts fail (and it's not the model)

When a new student shows us a disappointing Claude conversation, the prompt is almost always some version of "write a marketing email" or "summarize this" or "give me ideas for my business." Claude did exactly what was asked — and what was asked could only produce an average answer, because the prompt contained nothing that distinguishes this email, this document, this business from a million others.

The mental model that fixes it: treat Claude like a smart new hire on their first day. Capable, fast, well-read — and knowing absolutely nothing about your situation until you tell it. You wouldn't hand a new hire a sticky note reading "write a marketing email" and expect brilliance. You'd brief them: who the audience is, what you're selling, what tone the brand uses, what the last email that worked looked like. A prompt is that briefing, written down.

One honest caveat before the method: prompting well raises the floor and the average of what you get, dramatically. It does not make Claude infallible. It will still occasionally state something false with confidence, so facts, figures, and citations in any output you rely on need a human check — better prompts don't remove that step, they just make everything around it faster.

The five-part prompt: context, task, format, example, iterate

1. Context — the part almost everyone skips

Context is the background Claude needs to give your answer instead of the answer: who you are, who the output is for, what the situation is, and — most importantly — the actual material. Paste the document. Paste the email thread. Paste the job posting. Claude's long context window is one of its genuine strengths; it can hold a very long report or a whole contract in a single conversation, so there's rarely a reason to summarize your own source material before handing it over.

Weak: "Write a follow-up email to a client."

Strong: "I run a small landscaping company. Below is the email thread with a client who got our quote 10 days ago and went quiet [thread pasted]. Write a follow-up that's friendly, doesn't sound desperate, and gives them an easy reason to reply — we have one crew slot opening the week of August 10."

The strong version is longer to type once — and saves the four rounds of "no, more casual… no, mention the quote…" the weak version guarantees. If you find yourself repeatedly correcting Claude in the same direction, that correction belonged in the prompt.

2. Task — one specific verb, real constraints

"Help me with my resume" is a topic, not a task. "Rewrite my three most recent bullet points to emphasize measurable results, without inventing any numbers" is a task. Good task lines share three traits: a concrete verb (rewrite, compare, extract, critique, draft), a defined scope (these three bullets, not "my resume"), and explicit constraints (no invented numbers, under 150 words, plain English, no exclamation points).

Constraints matter more than most people expect, and negative constraints — what to avoid — are perfectly legitimate: "don't use the words 'delve,' 'leverage,' or 'game-changer'" is a real instruction Claude will follow. If you've read our 40 business prompt templates, you'll notice every one of them is built this way: verb, scope, constraints, always with a slot for your context.

3. Format — say what the output should look like

If you don't specify format, Claude picks one, and its default is often longer and more essay-shaped than you want. Say it directly: "a table with columns for clause, plain-English meaning, and risk level." "Five bullet points, each under 15 words." "A numbered checklist I can print." "Just the rewritten paragraph — no preamble, no explanation of what you changed."

That last one is worth memorizing. Adding "no preamble" to a prompt removes the "Certainly! Here's a draft that…" framing and gives you output you can paste straight into a document. Small instruction, outsized quality-of-life improvement.

4. Example — show, don't describe

Describing a style takes paragraphs and still misses; showing it takes one paste and lands. If you want an email in your voice, include one you actually wrote: "Here's an email I sent last month — match this tone." If you want a report section formatted a particular way, paste a section from a past report. Anthropic's own prompt-engineering guidance makes the same point — examples (what the docs call multishot prompting) are among the most reliable ways to steer output, more reliable than adjectives like "professional" or "casual," which mean different things to every reader.

This is the core of the voice-sample method we cover in depth in our guide to using Claude for writing: three samples of your real writing teach Claude your voice better than any paragraph of self-description ever will.

5. Iterate — the first answer is a first draft

People rate their prompting by the first response and quit. The skill that actually separates effective users is the second message. Treat the first output as a draft from that smart new hire and mark it up. The follow-up moves we teach, in rough order of how often we use them:

• Tighten: "Cut this by half without losing the second point."
• Redirect: "Too formal — write it the way you'd text a colleague you like."
• Push back: "You said X — argue against your own conclusion. What's the strongest case that it's wrong?"
• Ask for options: "Give me three versions: safe, bold, and weird."
• Make it check itself: "List anything in your answer that's an assumption rather than something I told you."

Two or three rounds like this routinely turns a 6/10 first draft into something usable. And because each round is cheap and fast, iterating beats trying to write one perfect mega-prompt — with one caution: very long conversations re-read everything above them on every turn, which burns through usage faster. Our message-limits guide covers the habits that keep iteration from eating your allowance — the short version is: iterate within a task, but start a fresh chat for a new task.

Three upgrades once the basics are habit

Make Claude interview you

The single highest-leverage sentence we teach: "Before you answer, ask me any questions you need to do this well." It flips the hardest part of prompting — knowing which context matters — onto Claude. For any meaty task (a plan, a difficult email, a document from scratch), Claude's questions will surface exactly the details you didn't think to include.

I need to write a proposal for a client who's comparing us against a cheaper competitor. Before you draft anything, ask me the five questions whose answers would most change what you write.

Separate your material from your instructions

When a prompt mixes a pasted document with your instructions, Claude occasionally treats text inside the document as something you're asking for. The fix is to label the sections. Anthropic suggests XML-style tags — <document>…</document> around the pasted material, instructions outside it — but even a plain line like "=== DOCUMENT BELOW, INSTRUCTIONS ABOVE ===" does the job. This matters most on long-document work, the kind we lean on in spreadsheet analysis and legal-document review.

Give roles a job, not a costume

You'll see "act as a world-class copywriter with 20 years of experience" in every prompt listicle. Honest assessment: the costume adds little. What does work is a role with a functional job — "review this as a skeptical CFO would: what would you cut?" or "read this as a first-time customer who's never heard of us: where do you get confused?" The role earns its place when it changes the perspective of the analysis, not when it just flatters the model.

What doesn't matter (despite what the listicles say)

• Magic words and secret templates. There is no incantation. Every "SECRET PROMPT FORMULA" post is the five parts above wearing a trench coat.
• Threats, tips, and emotional appeals. "This is very important to my career" and its cousins circulate endlessly; results are inconsistent at best. Context and constraints beat drama every time.
• Perfect grammar in your prompt. Claude reads typos fine. Spend the effort on substance, not polish.
• Politeness — for effectiveness, anyway. "Please" doesn't change output quality measurably. Type whatever keeps you sane.

The pattern behind all four: prompting is not a trick you play on the model, it's a briefing you give a collaborator. Anything that would make a human briefing better — more context, clearer ask, an example — makes a prompt better. Anything that wouldn't, doesn't.

A full worked example

Here's the whole method in a single prompt, the kind we build live in class:

CONTEXT: I'm a physical therapist opening a second clinic location in September. Below is my announcement email from when I opened the first location in 2023, which got a great response [pasted], and the details of the new location [pasted].

TASK: Draft the announcement email for the new location. Lead with what it means for existing patients (shorter drives for the north side), not with us celebrating ourselves. Don't invent any details not listed.

FORMAT: Under 200 words, one clear call to action to book at the new location, subject line included. No preamble.

Then: ask me anything you'd need to make this better before I send it.

Notice what's doing the work: the old email is the example, the new-location details are the context, the task has a priority order and a "don't invent" constraint, the format is exact, and the last line opens the iteration loop before the first draft even lands. Nothing here is clever. It's just complete — and if you're brand new and don't have Claude set up yet, our install guide gets you from zero to trying this prompt in about fifteen minutes.

How to actually get good at this

Reading about prompting is like reading about swimming. The method sticks when you run it on your own work with someone watching who can say "your context is thin here" or "that task line has three verbs in it — pick one." That feedback loop is the whole reason our class exists: we take the five-part structure above and drill it on the documents, emails, and problems you actually brought with you, until writing a complete prompt stops being a checklist and becomes the way you naturally ask.