Anthropic's Prompt Playbook

Stop settling for Claude's average answer.

The prompt framework Anthropic's own team demonstrated on stage, turned into seven copy-paste prompts. Six levers, a skeleton that makes Claude read your prompt as labelled containers, and one template you paste at the top of every serious conversation.

Get the prompt pack free

Free Prompt pack

Anthropic's Prompt Playbook

Pull all six levers, so Claude makes decisions instead of filling gaps with the average guess
Seven prompts you can paste today, including the full template with self-correction built in
A system that compounds — rules files, saved templates, and a 15-minute weekly review

Works with Claude desktop app, Claude Cowork, Claude on the web.

Unlock it now

Why this exists

Same Claude. Same question. Three very different answers.

Most people type prompts the way they text a friend. “Write me a script about AI for small business owners.” What comes back is word soup: vague, flat, occasionally confident and wrong. The instinct is to blame the model. It isn’t the model. The prompt skipped every part of what a serious prompt looks like.

Anthropic’s Applied AI team showed this on stage by running one question through Claude three times, changing only how the prompt was built. A lazy one-liner got a car crash described as a skiing accident. Adding a role and a tone got the accident right but the model wouldn’t name who was at fault. The full framework, stacked properly, named the driver in a format you could parse. Same model. Same question. The prompt was the entire difference.

That framework is this page: six levers, a skeleton that makes Claude read your prompt as labelled containers instead of a wall of text, five multipliers, and a system that compounds over months. I’ve rewritten it in plain language and left the prompts themselves exactly as demonstrated, because the wording matters. Fill in the brackets and run them.

What the difference looks like

See what one run produces before you try it.

The same request, sent two ways. The second one pulls all six levers and ends with the self-grade from Prompt 03.

PROMPT  Write me a LinkedIn post about our new one-hour response guarantee.

Excited to announce our new one-hour response guarantee! In today's
fast-paced world, customers expect speed, and we're here to deliver.
Reach out today to experience the difference.

─────────────────────────────────────────────────────────────────

PROMPT  Same ask. Role, context, task, examples, guardrails, format, bar.

Last Tuesday a client emailed at 4:52pm and had an answer by 5:14.
That's not a policy. It's what happens when support stops being a queue.
So we're putting it in writing: one hour, every message, or that month
is on us. Not because it's a great line. Because we've been doing it
for six months and nobody noticed until we said it out loud.

SELF-GRADE   voice 9 · specific 8 · would ship as-is 8

Who it's for

You pay for Claude and mostly get answers that need a rewrite before you can use them
You've noticed the same person gets far better output from the same model, and want to know what they're doing differently
You'd rather learn one framework properly than collect another hundred prompts
You write, analyse, or decide things for a living and want a system that improves every month

What you need

A Claude account on any plan
Any Claude surface. Everything here works in the desktop app, Claude Cowork, and the web app
Ten minutes to fill in the brackets the first time. After that it's a paste.

The prompt pack

All 7 prompts. Copy whichever one matches the job.

Wrap every section in tags The single biggest upgrade you can make. Claude was trained on text labelled with tags, the way a database is labelled with column names.
Show, don't describe A couple of real samples beat a paragraph of adjectives. Paste the voice or structure you're after, and Claude matches the pattern instead of guessing at the description.
Make Claude grade its own work The most underused move in prompting. Every first response is a draft, so make Claude critique it and rewrite in the same turn.
Force three perspectives For any decision with real tradeoffs, make Claude argue from three sides before it answers. It has to defend each angle instead of picking its favourite.
Let Claude write your prompt When you can't figure out how to word a prompt, stop trying. Describe what you want and let Claude build the prompt for you.
Load your rules file Keep a few short text files outside any single conversation, and paste the relevant one at the top of anything important. From that point on, every reply in the chat is held to those rules.
The full template All six levers, the tag skeleton, and the self-correction step, in one prompt. Paste it at the top of every serious Claude conversation and fill in the brackets.

Why most prompts come back average

Anthropic's Applied AI team ran the same question through Claude three times on stage, changing nothing but how the prompt was built. The lazy one-liner got a confidently wrong answer. Adding a role and a tone fixed the facts but the model wouldn't commit. The full framework got a clear call in a format you could parse. Same model, same question. The prompt was the whole difference.

The thing to internalise is that Claude is a reasoning engine, not a mind reader. Hand it a vague prompt and it fills every gap with the most average guess it can make. Hand it a scaffolded prompt and there's nothing left to guess about.

The six levers

Think of a prompt as a control panel. Most people pull one lever, maybe two. The people getting the good answers pull all six, in this order.

WHO — the role. Tell Claude who it is, and make it narrow. Never "a helpful assistant." Something like "You are the senior copy editor on my newsletter who hates passive voice." The role sets the tone and depth of everything that follows.
WHERE — the context. Describe the world Claude is working inside: your business, your audience, what this is for, what came before. Leave that out and Claude has to assume it; put it in and Claude can decide instead.
WHAT — the task. Not "help me with my LinkedIn post." Something like "rewrite this post so it reads like a founder sharing a lesson, not a consultant pitching." Claude should know exactly which finish line you're pointing at.
SHAPE — the output format. Say what the finished answer should physically be: a three-column table, a 300-word block, five bullets, a single tweet. Skip this and you get Claude's default shape, which is rarely the one you had in mind.
AVOID — the guardrails. What not to do. "No em dashes. No phrases like In today's world. Do not use the words unlock, leverage, or delve." A short list of don'ts closes off the usual exits into generic AI prose before Claude can take them.
BAR — the quality standard. Define shippable. "Good enough to post to my audience without a rewrite." "Specific enough that I can make a pricing decision in under 5 minutes." One line, and it sharpens everything above it.

The six-lever test: before you send anything, scan what you've written and count the levers you actually pulled. Only named the task? You'll get an average answer.

Two cheat codes from the demo

Label everything with tags — every distinct section, not only the context and the output. Inputs, examples, guardrails, format, reminders, each in its own container. Claude doesn't take your prompt in as one block of text; it navigates by the labels, so the more of them there are, the less it has to infer. Prompts 01 and 07 show the shape.
Write the first characters of the reply yourself. Whatever you put at the start of Claude's turn, it has to carry on from — so there's no preamble and no wandering intro, only the format you asked for. Want JSON? Start the response with [ . Want a clean final answer? Start with <final_answer>. Want a table? Start with | . A language model can't contradict what's already on the page, so if the reply begins inside a structure, it stays inside it.

Two multipliers that don't need a prompt

Chain, don't cram. Ask for five things in one prompt and you get five half-done answers. Break it into sequential prompts instead — each one stays focused, each output is deep enough to use, and you can correct course at every step instead of at the end. A 2,000-word post in four prompts (outline, opener, body, close) beats anything a one-shot prompt produces.
Give your rules a reason. A bare rule gets obeyed to the letter; a rule with a why behind it gets applied with judgement, and Claude spots the edge cases the bare version would have sailed past. Weak: "Keep it under 200 words." Strong: "Keep it under 200 words because this goes on a carousel slide and anything longer gets cut off." Weak: "Do not use em dashes." Strong: "Do not use em dashes because they are one of the biggest tells that AI wrote the text."

Anthropic's 10-part stack, for reference

This is the full assembly order the Applied AI team walked through. The six levers and the skeleton map onto it, with the two cheat codes at the end. Stack top to bottom, and skip parts on purpose when they don't fit the task — never by accident.

Task context — who Claude is and the situation it's in
Tone context — the register to write in, and where it should decline
Background data, documents, images — the material that doesn't change
Detailed instructions — the steps, in the order they happen
Examples — a few worked pairs, each inside tags
Conversation history — earlier turns Claude needs to keep in mind
Immediate task reminder — say the task again, right before you ask
Precognition and guidelines — the guardrails and the reminders
Output formatting — the exact shape of the answer, in tags if you can
Prefilled response — begin Claude's reply so the structure is forced

The skeleton

Wrap every section in tags

The single biggest upgrade you can make. Claude was trained on text labelled with tags, the way a database is labelled with column names.

Wrap a section in tags and Claude stops treating it as prose and starts treating it as a labelled container. The names are up to you — whatever describes the job. What matters is that no distinct section is left outside one.

One more rule from the demo: put the big stuff above the ask. If you're pasting a document, a spreadsheet, or a transcript, it goes above your question. Claude reads top to bottom, so material on top builds understanding and the question at the bottom lands with full context loaded. Put the question first and the answer gets measurably worse.

Prompt 01

<context>
I run an AI education brand. My audience is non-technical
founders and content creators who use Claude Cowork to build
personal software. They do not write code.
</context>

<task>
Rewrite this paragraph so it sounds like me, not like AI.
</task>

<constraints>
- No em dashes
- No 'In today's world' openers
- Do not say 'unlock' or 'leverage' or 'delve'
</constraints>

<output_format>
Just the rewritten paragraph. Nothing else.
</output_format>

<paragraph_to_rewrite>
[PASTE PARAGRAPH HERE]
</paragraph_to_rewrite>

What you get: A rewritten paragraph in your voice, and nothing else. Swap the context block for your own world before you paste.

Show, don't describe

A couple of real samples beat a paragraph of adjectives. Paste the voice or structure you're after, and Claude matches the pattern instead of guessing at the description.

Claude matches patterns far better than it follows descriptions. If you want a certain voice back, paste one. If you want the opposite of a certain style, paste that too and say "not like this." The example below teaches an analysis style in two pairs, then hands over a fresh data point.

Prompt 02

<examples>
<example>
<input>We grew the list from 2,000 to 2,400 this month.</input>
<output>List up 20% this month. Worth checking what drove it:
the lead magnet or the podcast appearance?</output>
</example>
<example>
<input>We missed our revenue target by 12%.</input>
<output>Short of target by $X. Before blaming the team,
check if this is a close rate issue or a top of funnel issue.</output>
</example>
</examples>

Now analyze this: [YOUR NEW DATA POINT]

What you get: Analysis in the exact register of your examples. Replace the two pairs with samples from your own work.

Five multipliers

Make Claude grade its own work

The most underused move in prompting. Every first response is a draft, so make Claude critique it and rewrite in the same turn.

Paste this as a follow-up after any answer you'd otherwise start editing by hand. In testing, the second version came back measurably better more than 80% of the time. Change the three criteria to whatever "good" means for the task.

Prompt 03

Now re-read what you wrote.
Score it 1-10 on three things:
- Does it sound like me, or like AI?
- Is it specific, or is it generic?
- Would I actually ship this without editing?
Anything scoring below 8, explain what is weak and fix it.
Show only the improved version.

What you get: The improved version only, with anything that scored under 8 fixed.

Force three perspectives

For any decision with real tradeoffs, make Claude argue from three sides before it answers. It has to defend each angle instead of picking its favourite.

Use it on pricing, hiring, positioning, anything where the honest answer is "it depends." Rename the three angles to whoever actually has a stake in your decision.

Prompt 04

Look at this decision from three angles:
1. The growth-minded founder chasing market share
2. The operator watching margins and runway
3. The customer who wants fair value
Give me 2 sentences from each angle, stating their case.
Then pick the angle I should weight most, and tell me why.
Flag any case where two angles disagree sharply so I can think about it.

What you get: Three short cases, a recommendation on which to weight, and a flag wherever two angles disagree sharply enough that you should think about it yourself.

Let Claude write your prompt

When you can't figure out how to word a prompt, stop trying. Describe what you want and let Claude build the prompt for you.

Your blind spots in prompting are the exact spots Claude is good at, because it's seen millions of prompts and knows which patterns produce which outputs. This uses its strength to patch your weakness in one move. Run the prompt it gives you in a fresh conversation.

Prompt 05

I want to: [YOUR GOAL]
Here is what I know about the situation: [CONTEXT]
Here is what good output would look like: [EXAMPLE OR DESCRIPTION]
Write me the best possible prompt to get that output.
Add anything I forgot. Structure it in XML tags.
I will run it on a fresh conversation.

What you get: A prompt, structured in XML tags, that's almost always better than the one you'd have written.

The compounding system

Load your rules file

Keep a few short text files outside any single conversation, and paste the relevant one at the top of anything important. From that point on, every reply in the chat is held to those rules.

Four files cover most of it. voice.md holds your writing tics, sentence length, vocabulary and tone. banned-words.md is every word or phrase you never want to see. audience.md says who reads your work, what they care about, and what makes them scroll past. project-state.md is what you're building, its status, and the decisions already made.

Then save every prompt that produces a great result. Swap the specifics for brackets and leave the bones intact. In a few months you'll have a template for every kind of task and never start from scratch again.

The part people skip is the weekly review. Every Friday, 15 minutes: look at the week's AI output, find what you had to rewrite before shipping and what kept missing in the same way, and turn those edits into new rules in the files above. Three months of that and your prompting is on a different level. Skip it and you stay where you are.

Prompt 06

Read this file fully before starting.
Follow every rule. If you are about to break one,
stop and ask me before continuing.

<voice_rules>
[PASTE FILE HERE]
</voice_rules>

What you get: A conversation that follows your rules without being reminded — and stops to ask before it breaks one.

The template

The full template

All six levers, the tag skeleton, and the self-correction step, in one prompt. Paste it at the top of every serious Claude conversation and fill in the brackets.

Run the checklist before you send it. You named who Claude is, specifically. You explained your world and who this is for. The task is narrow and named. There's at least one example of good output, in tags. Every guardrail has a reason attached. You said the exact shape the answer should come back in. You defined what "shippable" means for this task. The self-correction step is at the end. And you've asked yourself whether this is really one prompt, or a chain of three.

Here's the trajectory if you actually do this. Week one, the answers tighten up and you notice. A month in, the rules files carry the context and you stop repeating yourself at the start of every chat. Three months in, each recurring job has its own saved template. Half a year in, you're getting more out of Claude than almost anyone around you — and the model didn't change. Your scaffolding did.

Prompt 07

<role>
You are [SPECIFIC PERSONA WITH EXPERIENCE AND QUIRKS].
</role>

<context>
[MY WORLD: BUSINESS, AUDIENCE, WHAT THIS IS FOR,
WHAT HAS COME BEFORE.]
</context>

<task>
[THE SPECIFIC ASK, NAMED CLEARLY.]
</task>

<examples>
<example>
<input>[SAMPLE INPUT]</input>
<output>[IDEAL OUTPUT]</output>
</example>
</examples>

<guardrails>
- [WHAT NOT TO DO + WHY]
- [WHAT NOT TO DO + WHY]
</guardrails>

<output_format>
[EXACT SHAPE OF THE ANSWER, IDEALLY IN TAGS]
</output_format>

<bar>
[WHAT 'SHIPPABLE' MEANS FOR ME ON THIS TASK.]
</bar>

Now complete the task.
After you finish, re-read and score 1-10 on
voice match, specificity, and whether I would ship it as is.
Anything below 8, explain what is weak and fix it.
Show only the final version.

What you get: The finished task, self-graded on voice, specificity and whether you'd ship it, with anything under 8 already fixed.

Based on the prompt framework Anthropic's Applied AI team demonstrated on stage.

How to use it

Run the six-lever test on your next prompt

Before you send it, count the levers. Who, where, what, shape, avoid, bar. If you only named the task, add the other five before you hit enter. That habit alone changes the answers you get back.

Paste the template at the top of one serious conversation

Prompt 07 is the whole framework in one block. Fill in the brackets for a task you actually care about this week and compare the result to what you'd normally get. Then save your filled-in version as a template.

Start a voice.md and load it every time

Ten lines on how you write is enough to begin. Paste it in with Prompt 06 at the start of each session, and every Friday add whatever you had to fix that week. This is the part that compounds.

Get the prompt pack

Want the version that does the work?

Beyond the Chatbox

The 90-minute Claude training we run inside companies. Everything from how the model actually works to projects, skills, connectors and keeping your data safe — so you stop getting chatbot answers and start handing Claude real work.

The prompt anatomy. Role, goal, context, examples — and the one closing line that improves almost every answer you'll ever get.
Projects as AI employees. The three I actually run my business on, and how to build your own.
Skills, connectors and MCP. What they are in plain language, and which to set up first.

$27

One-time purchase. Yours forever.

See Beyond the Chatbox Watch immediately after checkout

Questions

Do I need to be technical to use this?

No. Nothing here involves code. The tags are just labels in angle brackets — you type them like any other text. If you can copy and paste, you can use every prompt on this page.

What is actually in the pack?

Seven copy-paste prompts, the six-lever framework that explains why they work, two cheat codes from Anthropic's own demo, and the 10-part stack their Applied AI team uses as an assembly order. All of it stays on this page. There's nothing to download.

Why do the prompts use XML-style tags?

Because Claude was trained to read labelled text, and a tagged section reads as a container rather than as prose. It's the single biggest upgrade you can make to a prompt, and it costs nothing. The tag names can be anything that makes sense for the job.

Do these work in Claude Cowork?

Yes. The framework is about how you structure a request, so it holds anywhere you talk to Claude. Cowork is where it pays off most, because a well-scaffolded prompt is what lets you hand over a whole job and walk away.

Where did this come from?

Anthropic's Applied AI team walked through their internal prompting framework live on stage, running the same question three ways to show the gap. This page is that framework rewritten for people who'd rather have the prompts than the talk.

Why do you ask for my email?

So we can send you the link back to this page, tell you when we add prompts to the pack, and let you know when the paid skills ship. Everything here stays on the page — there is nothing to download. 1–2 emails a month helping you learn more about AI. Unsubscribe any time.