A clearer brief. A more useful answer.

How Claude reads a prompt

Section one: how Claude reads a prompt.

The idea in this section is simple, and it explains almost every disappointing answer you've ever gotten from an A I tool. Claude is like a brilliant colleague on their first day. It knows a great deal. But it knows nothing about you.

The new-colleague model

Picture a very capable new colleague who walked in this morning.

Look at what they bring. Broad knowledge of language, business, and most subjects. Fast drafting, summarizing, and rewriting. Patience for the fifth rewrite, without a sigh. And good judgment, once they know what you're after.

Now look at what they don't know, unless you tell them. Your company, your customers, and your audience. What good looks like to you. Why you need this, and what happens to it next. And anything outside the conversation, or anything newer than its training, unless you give it the source or let it search.

Anthropic, the company that makes Claude, has a golden rule for this. Show your prompt to a colleague with minimal context on the task, and ask them to follow it. If they'd be confused, Claude will be too.

Try it on your own work. Think of a task you'd hand to a smart temp on their first day. What would you need to tell them? That's what belongs in your prompt.

What happens when facts are missing

This is the most important idea in the module, so it's worth slowing down.

Claude doesn't guess wrong on purpose. A language model predicts the most likely response. When you leave a blank in your prompt, most likely means most typical. And typical means generic.

Take this prompt: write an email about the new schedule. Now count the blanks. Who is it from? Who reads it? What changed, and when? Why did it change? How long should it be? What tone? That's six blanks, and six decisions you handed over without meaning to.

So what do you get? A polite, generic email full of placeholders, like date in brackets and reason in brackets. It's correct, and it's useless.

The fix isn't a magic phrase or a secret template. It's closing the gaps. Every blank you fill is a decision Claude no longer has to make for you. The six ingredients in the next section are simply a checklist for closing those gaps.

Where your prompt lives

Here's a point that surprises many people. Your prompt is more than the box you type in.

The whole conversation is part of the prompt. Every earlier message still counts, so a correction you made three turns ago is still shaping the answer.

Attachments are part of the prompt too. Files and pasted text are what Claude works from.

You can also edit a message you already sent. Hover over it, choose edit, fix the prompt itself, and send it again. Claude keeps both versions, and you can flip between them with the arrows under the message. You'll see that in the second demo.

When the job changes, start a new chat. Long, wandering chats blur the task, because all those earlier detours are still part of the prompt.

And finally, Projects give you a workspace with its own files and standing instructions, so you don't have to repeat the background every time. A later module covers Projects in detail.

The six ingredients

Section two: the six ingredients.

Role, source, task, constraints, format, and courtesy. They're the same parts every time. Put together, they read like a good assignment note to a colleague. In this section you'll learn each one, see them at work in two real examples, and then use them yourself in Lab one.

Role, Source, Task, Constraints, Format, Courtesy

Here are the six ingredients of a strong prompt. The strip across the top is the recipe, and each numbered card gives you a tip for that ingredient.

One is role. Say who Claude should be. Act as an H R generalist, or act as a skeptical editor. A role sets the vocabulary, the depth, and the tone.

Two is source. Give it the material. Attach the file or paste the facts. The actual document beats any description of it.

Three is task. Ask for one clear thing. For example, would you please write the announcement email? One task, phrased as a request, so Claude knows exactly what it's making.

Four is constraints. Say who it's for, and the limits. For example, for part-time staff reading on their phones, so under one hundred fifty words. This is where audience, purpose, length, and anything that must not change all live.

Five is format. Describe the result, and what done looks like. A three-column table. Or, done when every date matches the fact card. A clear finish line makes the result checkable.

And six is courtesy. Ask like a colleague: please, and thank you, and invite questions, such as, before you draft, ask me up to five questions. This sixth ingredient is Mark's practice, not an official rule from Anthropic. In his experience, a collaborative ask tends to draw out the flags and caveats you want.

You rarely need all six in full. The test is always the golden rule: would a new colleague have what they need? The two ingredients people leave out most often are source and format.

This slide has the camera, so it’s worth a screenshot.

Explain the reason behind a constraint

Anthropic's own prompting guidance makes a point that changes how you write instructions: context beats commands.

Look at the first pair. The less effective version is a bare rule: never use ellipses. The more effective version explains why: your response will be read aloud by a text-to-speech engine, so never use ellipses, since the engine won't know how to pronounce them.

Now the second pair. Less effective: do not use markdown in your response. More effective: say what to do instead. Your response should be composed of smoothly flowing prose paragraphs.

Why does this work? Claude generalizes from the reason. Tell it an email will be read on a phone, and it will shorten paragraphs you never mentioned. A bare never teaches one rule. A reason teaches the goal.

One nuance. Do-not instructions still have a place, for hard limits. Do not change any dates is a perfectly good instruction. Just pair your limits with the positive version, and a reason, wherever you can.

The camera marks this pattern as one worth saving.

A stronger schedule email prompt

Let's see the ingredients at work. Here's a staff email, weak versus strong.

The weak prompt is: write an email about the new schedule. There's no sender, no audience, no facts, and no length. You'll get a template full of brackets.

Now the strong version. The numbers in brackets mark each ingredient.

One, the role: please act as the operations manager of a three-café coffee company.

Two, the source: here are the facts. Starting Monday, November second, weekday opening moves from six to six thirty in the morning at all three cafés. Weekend hours don't change, and openers now start at five forty-five.

Three, the task: would you please write the email announcing this to our forty-two staff?

Four, the constraints: many are part-time students reading on their phones, so keep it under one hundred fifty words and friendly, and explain that the first half hour averaged four customers. Notice that one reason, reading on their phones, does the work of three separate rules about length and layout.

Five, the format: subject line first, then the email, ending with who to ask.

And six: thank you.

Yes, it's longer. And it saves you three rounds of fixing.

A stronger editing prompt

Here's the second example: improving a draft. It's the request people make most often, and get the least from.

The weak prompt is: make this better. Better how? Shorter? Warmer? Fewer errors? Claude will guess. And while it's improving things, it may quietly change your dates.

The strong version: please act as a careful copy editor. Below is our draft holiday-hours announcement for the website. Would you please tighten it and fix any grammar? Please keep every date and time exactly as written, stay under one hundred twenty words, and keep it warm and plain. Give me the revised text first, then a short list of what you changed and why. Thank you. And then the draft is pasted below.

Look at two parts in particular. Keep every date and time exactly as written is a hard limit on the facts, and it's exactly where a do-not instruction earns its place.

And the short list of what you changed makes the edit reviewable. You can see what moved, instead of rereading every line to find it.