How Claude Actually Thinks

Think of every Claude conversation as briefing a brilliant new hire. This person is incredibly smart, has broad knowledge, and can execute anything — but they only know what you tell them in this conversation.

That mental model unlocks everything else in this Playbook.

The Context Window: Your Briefing Room

Every time you start a conversation with Claude, you're opening a "context window." This is Claude's working memory — everything it can see and reference during your conversation.

Here's what matters:

• Claude can see everything in the conversation — your messages, its responses, any files you've uploaded, and any project instructions you've set up
• More context = better output — the more relevant information Claude has, the better it performs
• Context has a limit — think of it as a very large document (roughly 150,000 words for most Claude models)

Tokens: The Unit of AI Thinking

You'll hear the word "tokens" a lot. Here's all you need to know:

A token is roughly 3/4 of a word. "The quick brown fox" is about 4 tokens. A typical email is about 200 tokens. A 10-page document is about 3,000 tokens.

Why does this matter? Two reasons:

1. Pricing — Claude subscriptions include a token budget. More complex conversations use more tokens.
2. Context limits — the context window is measured in tokens, not words.

The Briefing Principle

Here's the most important concept in this entire Playbook:

The quality of Claude's output is directly proportional to the quality of your input.

This isn't about "prompt engineering tricks." It's about giving Claude the context it needs to do excellent work — just like you'd brief a consultant or a new team member.

Write me a cold email for my SaaS product.

How Claude Processes Your Input

When you send a message, Claude processes the entire conversation in sequence:

This is why project instructions matter so much. They're the first thing Claude sees, every single time. We'll cover this in depth in the Projects, Memory & Skills playbook.

The Three Levels of Context

Think of context as having three layers, and your results improve dramatically as you use more of them.

Level 1: The Conversation

What you type right now. This is where most people stop — and it's the least powerful layer on its own.

Every message you send is context. Every response Claude gives becomes context for the next message. This is why conversations get better over time within a session — Claude accumulates understanding as you go.

Level 2: Project Instructions

Persistent context that loads automatically every time you start a new conversation. Your brand voice, your company info, your preferences, your audience details.

This is where the magic starts. Instead of re-explaining who you are and what you need, Claude already knows. You skip straight to the work.

Level 3: Referenced Files & Data

Documents, spreadsheets, transcripts, and other materials you upload or reference. This is where Claude becomes genuinely useful for real work.

A financial model. A competitor's marketing page. A customer interview transcript. A 50-page strategy document. Claude can process all of it and work with the specifics — not just generic advice, but recommendations grounded in your actual data.

The "New Hire" Test

Before you send any important prompt, run this mental test:

If you handed this exact text to a brilliant new hire on their first day, would they have enough information to do great work?

If not, add what's missing. Usually it's:

• Who the audience is — "Write this for CTOs at mid-market companies" vs. just "Write this"
• What success looks like — "The email should get them to book a call" vs. "Write an email"
• What tone to use — "Casual and direct, like texting a colleague" vs. leaving it to chance
• What constraints exist — "Under 200 words, no jargon, must include pricing" vs. open-ended

Practical Example: The Difference Context Makes

Let's look at the same task with increasing levels of context:

Write an apology email about downtime.

This will produce a generic, template-sounding email. Usable in a pinch, but not good.

Write an apology email about today's downtime.\n\n(Claude already knows from Project instructions: company name, product, customer base, communication tone, brand voice)

Better — Claude matches your voice and knows your audience. But it's still guessing about the specifics.

Draft a customer communication about today's service outage. Details: 45-minute downtime from 2:15-3:00 PM EST, affected 12% of users (our East Coast infrastructure), root cause was a database migration that hit an edge case, fully resolved with no data loss. Tone: empathetic, transparent, and confident. Include: what happened, who was affected, what we did to fix it, and what we're doing to prevent recurrence. Under 200 words.

Now Claude has everything. The output will be specific, accurate, and on-brand. One draft, minimal editing.

What This Means for You

Three practical takeaways you can use right now:

1. Don't start from scratch — Set up Projects (Chat) or CLAUDE.md (Code) with your persistent context. You'll cover this in the Projects, Memory & Skills playbook.

2. Front-load context — Give Claude the important information at the beginning of the conversation, not as an afterthought. The order matters: Claude weighs earlier context heavily.

3. Be specific about what you want — Claude can't read your mind. Tell it the format, the audience, the tone, the constraints, and the goal. Every missing detail is a place where Claude will guess — and guessing means generic output.

Next Up

Now that you understand how Claude processes information, the next playbook covers which Claude product you should actually be using — Chat, CoWork, or Code. The answer depends entirely on what you're trying to accomplish.