The Prompt Engineering Playbook For Claude

Anthropic's Applied AI team walked through their internal prompt framework live on stage. They ran the same question through Claude three times, changing only how the prompt was built, and the gap between the first answer and the third is the entire skill. This is the full playbook: the 6 levers, the skeleton, 5 multipliers, the compounding system, and two cheat codes. Non technical, no code required. Just copy and adapt.

What You'll Learn

• The 6 levers every good prompt pulls, in the order to stack them
• The XML tag skeleton that upgrades every prompt
• 5 multipliers that compound quality (chaining, self grading, reasons, perspectives, meta prompting)
• The compounding system: rules files, templates, and the weekly review
• Two cheat codes from Anthropic's own demo, plus the full 10 part stack

Why Most Prompts Suck

Typing prompts like you are texting a friend ("Write me a script about AI for small business owners") gets you word soup. Vague, flat, sometimes confidently wrong. The model is fine. The prompt is skipping every part of what a serious prompt actually looks like.

In Anthropic's demo, the same question ran three times:

1. One sentence, lazy: Claude called a car crash a "skiing accident"
2. Added role + tone: correct accident, but refused to name a fault driver
3. Full framework stacked: confidently named Vehicle B as at fault, in parseable format

Same model. Same question. Three prompts.

The real insight: an AI model is a reasoning engine, not a mind reader. When you hand it a vague prompt, it fills the missing pieces with the most average, generic guess. When you hand it a scaffolded prompt, it has nothing to guess about. That is the whole game.

The 6 Levers Every Good Prompt Pulls

Think of a prompt as a control panel. Beginners pull one, maybe two levers. Experts pull all six. In stacking order:

1. WHO (the role). Tell Claude who it is. Never "a helpful assistant." Something narrow: "You are the senior copy editor on my newsletter who hates passive voice." "You are a short form video strategist who has studied 10,000 viral Reels." The role sets the entire tone and depth of the response.
2. WHERE (the context). Paint the world Claude is working inside. Your business, your audience, what you are shipping this for, what has come before. Without context, Claude makes assumptions. With context, Claude makes decisions.
3. WHAT (the actual 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." The difference is whether Claude knows exactly what finish line you are pointing at.
4. SHAPE (the output format). What does the answer look like when it lands? Three column table? 300 word block? Bullet list of 5 items? A single tweet? If you do not say, Claude guesses, and the guess rarely matches.
5. AVOID (the guardrails). The list of what NOT to do. "No em dashes. No phrases like In today's world. Do not use the words unlock, leverage, or delve." Negative rules cut off the most common ways Claude drifts into generic AI writing.
6. 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." This one line sharpens everything above it.

The 6 lever test: before you hit send on any prompt, scan your text and check which levers you actually pulled. If you only named the task, your output will be average. Pull all six and watch the quality change.

The Skeleton (How To Stack A Prompt)

The 6 levers are what goes in. This is how to arrange them on the page so Claude actually uses them.

Wrap every section in tags. This is the single best upgrade you can make to your prompts. Claude was trained to read text that is labeled with tags, the same way a database is labeled with column names. When you wrap a section, Claude stops reading text and starts reading a labeled container.

<context>
I run an AI education brand. My audience is non-technical
founders and content creators who use Claude Code 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>

Tag names can be anything. Use whatever labels make sense for the task. The point is that every distinct section is inside tags.

Put the big stuff above the ask. If you are pasting a document, a spreadsheet, a voice memo transcript, anything long, put it ABOVE your actual question. Claude reads top to bottom. Material on top builds understanding. Question at the bottom lands with full context loaded. Flipping the order measurably drops quality.

Show, do not describe. Two examples do more for Claude than ten sentences of description. If you want a certain voice or structure back, paste a sample. If you want the opposite of a certain style, paste that too and say "not like this." Claude matches patterns way better than it follows descriptions.

<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]

5 Multipliers That Compound Quality

Once your prompt has all 6 levers and a clean skeleton, these five moves compound the quality further.

Multiplier 1: chain, do not cram. If your prompt is asking for five things, you are going to get five half done answers. Break it into sequential prompts instead. Each one stays focused. Each output is deep enough to actually use. And you can correct mistakes at every step instead of at the end when it is too late. Example: a 2000 word blog post in four prompts instead of one. Prompt 1 built the outline. Prompt 2 wrote the opener with the hook. Prompt 3 wrote the body sections. Prompt 4 tightened the close and added the FAQ. Every output was sharper than anything a one shot prompt would have produced.

Multiplier 2: make Claude grade its own work. The most underused move in prompting. Every first response is a draft. Force Claude to critique itself and rewrite, all in the same turn. The second version is measurably better more than 80% of the time in testing.

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.

Multiplier 3: give your rules a reason. Raw rules get followed literally. Rules with reasons get followed intelligently, and Claude catches edge cases the bare rule would miss.

• 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."
• Weak: "Make it sound casual." Strong: "Make it sound casual because my audience is non technical creators who disengage the moment a piece feels corporate."

Multiplier 4: force multiple perspectives. For any decision with real tradeoffs, ask Claude to argue from three sides before it answers. You get better strategic thinking and cleaner recommendations, because the model has to defend each angle instead of picking its favorite.

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.

Multiplier 5: let Claude write your prompt. When you cannot figure out how to word a prompt, stop trying. Describe what you want and ask Claude to build the prompt for you. The generated prompt is almost always better than what you would have written, because Claude has seen millions of prompts and knows which patterns produce which outputs.

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.

Why this trick wins: your blind spots in prompting are the exact spots Claude is good at. Meta prompting uses Claude's strength to patch your weakness, in one move.

The Compounding System

Everything above is about individual prompts. This is what makes prompting compound over months, so that a year from now your baseline output is higher than most people's best days.

Rules files you reuse everywhere. Keep a few short text files that live outside any single conversation. Every time you start something important, paste the relevant one at the top of the prompt. Claude now works under those rules for the rest of the chat.

• voice.md: your writing tics, sentence length, example vocabulary, tone
• banned-words.md: every word or phrase you never want to see in your output
• audience.md: who reads your content, what they care about, what makes them scroll past
• project-state.md: what you are currently building, status, key decisions already made

At the start of the session, paste:

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>

Save every winning prompt as a template. Any prompt that produces a great result should get saved. Replace the specifics with brackets, keep the structure. Over months you build a library for every kind of task. Writing, analysis, editing, strategy. You never start from scratch again. You pull the template, swap the brackets, ship.

The weekly review. Every Friday, take 15 minutes. Look at the AI output from the week. What did you have to rewrite before shipping? What kept missing the mark in the same way? Turn those edits into new rules in your voice or banned words files. In three months of doing this, your prompting is a different level. Skip it and you stay still forever.

Two Cheat Codes From Anthropic's Demo

Cheat code 1: label everything with tags. Not just the context. Not just the output. Every distinct section. Inputs, examples, guardrails, formats, reminders. Claude does not read your prompt as one wall of text, it reads it as a set of labeled containers. More labels means better navigation means higher quality output.

Cheat code 2: start Claude's answer for it. This is the one most people have never heard of. You can literally type the first few characters of Claude's response, and the model has to continue from there. No preamble. No wandering intros. Just the format you demanded.

• Want JSON? Start Claude's response with [
• Want a clean final answer? Start with <final_answer>
• Want a table? Start with |

Why this works: a language model cannot contradict what is already on the page. If the response begins inside a structure, the model keeps going inside it. This trick alone turns Claude from a chatty assistant into a predictable structured output engine.

The Full Template

Paste this at the top of every serious Claude conversation. All 6 levers, XML skeleton, self correction tacked on at the end. Replace the brackets and you have a top tier prompt.

<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.

Pre send checklist:

• I named who Claude is, specifically (not "assistant")
• I explained my world and who this is for
• The task is narrow and named, not vague
• I pasted at least one example of good output in tags
• I listed guardrails, and each one has a reason attached
• I said the exact shape the answer should come back in
• I defined what "shippable" means on this specific task
• I added a self correction step at the end
• I checked: is this really one prompt, or should it be a chain of 3?
• Everything is wrapped in XML tags Claude can label

What changes when you do this: the first week, Claude's output gets noticeably sharper. By the end of month one, your rules files start running the show and you stop re explaining yourself. By month three, you have a template for every recurring task and your output compounds. By six months in, you are using Claude at a level most people will not reach, not because the model got better, but because your scaffolding did.

There is no magic phrase. There is no secret prompt somebody is hiding. Six levers. A skeleton with tags. Five multipliers. A system that compounds. Two cheat codes. One template. Copy the template. Fill it in. Run it.

Appendix: The 10 Part Stack From Anthropic

For completeness, here is the full 10 part stack Anthropic's Applied AI team walked through on stage. The 6 levers plus the skeleton map onto this stack, with the two cheat codes added on. Use it as the assembly order for your prompts.

1. Task context: the role and the scenario
2. Tone context: how Claude should sound, when to refuse
3. Background data, documents, images: everything static
4. Detailed instructions: step by step, in order
5. Examples: few shot examples in tags
6. Conversation history: prior turns Claude should remember
7. Immediate task reminder: re state the task right before the question
8. Precognition / guidelines: guardrails and reminders
9. Output formatting: exact structure, ideally in tags
10. Prefilled response: start Claude's answer to force structure

How to use the stack: the playbook is the strategy. The 10 part stack is the assembly order. Stack top to bottom. Skip parts on purpose when they do not fit your task, never by accident. This is how Anthropic's own team builds prompts with enterprise clients.

Go Deeper

Prefer a printable version? Grab the PDF: https://drive.google.com/file/d/1hROihM2gaYt5Exs0vzPrLo6Np2Utqfal/view?usp=drivesdk

Want the whole system installed live, with Cooper on the call? Build Your Own AI, the two week sprint runs the first two weeks of August.

Or get a new working tool every Monday inside The AI Operator community on Whop.