Phase 1, Session 2

What Claude Sees at Session Start — The Boot Sequence

The Big Idea

When you open a new session, Claude doesn't just "wake up." There's a precise loading sequence. Things arrive in a specific order, and that order matters because it creates a priority stack: earlier = more foundational, later = more immediately relevant.

Analogy
Think of it like booting a computer. The BIOS loads first (Anthropic's system prompt). Then the OS kernel (CLAUDE.md). Then drivers and services (rules files, supplements). Then the user logs in (your first message). Each layer depends on the ones below it. You wouldn't expect to configure Wi-Fi before the network driver is loaded. Same principle.

Understanding this order is practical, not theoretical. It tells you where to put different kinds of instructions, why some things "stick" better than others, and why your Team Claude architecture works the way it does.

The Boot Sequence — Step by Step

Click each step to see exactly what loads and who controls it.

Step 1: Anthropic's System Prompt ~4,000 tokens
Anthropic controls this

Before you say a word, Anthropic has already given Claude its baseline instructions. This includes safety guidelines, formatting rules, how to use tools, the knowledge cutoff date, and behavioral guidance.

In Cowork specifically, this also includes: your user preferences (the ones you set in settings), your name and email, the list of available skills, instructions for how to handle files, and how to use computer-use tools.

What you can't change: Safety behavior, core personality, tool schemas.

What you CAN influence: Your user preferences setting ("Use quick and clever humor," "don't use em dashes," etc.) gets injected into the system prompt. This is why those preferences work across sessions without being in CLAUDE.md.

Step 2: CLAUDE.md Loads ~1,500 tokens
You control this

Your CLAUDE.md file is read automatically. It typically sets identity ("You are [role], working for [company]"), defines the organization, links to a knowledge map, lists connected tools, and establishes session rules.

Critically, CLAUDE.md also contains @-import references that pull in additional files.

Why this matters: CLAUDE.md is the ONLY file that loads automatically. Everything else requires either an @-import or an explicit read action. This is why CLAUDE.md is the most important file in your entire setup. If it's wrong, everything downstream is wrong.

Token trade-off: Every line in CLAUDE.md costs you on every single message. A 50-line CLAUDE.md vs. a 500-line CLAUDE.md is the difference between ~500 tokens/turn overhead and ~5,000 tokens/turn overhead. Yours is lean by design. The knowledge map alone saves thousands of tokens by pointing to files instead of including their content.

Step 3: @-Imported Rules Files Load ~2,000 tokens
You control this

The four @.claude/rules/ files referenced in CLAUDE.md load as part of the boot context:

  • hard-rules.md — Non-negotiable behaviors (response rules, verification requirements, safety gates)
  • session-protocol.md — Startup sequence, close-out steps, context lookup order
  • team-structure.md — Who does what across sessions, messaging, role boundaries
  • domain-rules.md — Domain-specific rules for the session's primary focus area

Why separate files instead of one big CLAUDE.md? Three reasons:

1. Maintenance: Update marketing rules without touching core identity.

2. Reusability: hard-rules.md could be shared across sessions (same rules, different roles).

3. Clarity: When Claude violates a rule, you know exactly which file to check.

The downside? They ALL load at boot, so modularity doesn't save tokens. It saves your sanity.

Step 4: Your First Message Arrives varies
You control this

Now Claude has its foundation (system prompt + CLAUDE.md + rules) and receives your first message. At this point, Claude "wakes up" — it can see everything loaded so far and your request.

What Claude knows at this exact moment:

  • Who it is (its role and identity from CLAUDE.md)
  • Who you are (your name, role, and context from CLAUDE.md)
  • All hard rules and session protocols
  • What tools are connected
  • Your format preferences

What Claude does NOT know yet:

  • Product-specific details (supplements not loaded yet)
  • What happened last session (no session memory loaded)
  • Current system state (hasn't run any tools yet)
  • What's in the KB (hasn't searched yet)
  • Messages from other sessions (hasn't checked inbox)

This is exactly why the session startup protocol exists — to fill in these gaps in a deliberate order.

Step 5: Session Startup Protocol Runs ~8,000-15,000 tokens (tool results)
Claude executes this

This is Claude following the startup protocol from session-protocol.md. Each step is a tool call, and each tool result lands in the context window. A typical startup might look like:

  1. Identify environment (what machine/interface am I on?)
  2. Load supplements (read product/project context files)
  3. KB orientation (run knowledge base searches for current state)
  4. Read open tasks (what's in progress?)
  5. Check for new messages or alerts
  6. Review active priorities
  7. Wait for direction

The token cost: A thorough startup sequence consumes 8,000-15,000 tokens. That's the price of Claude being properly informed. Without it, Claude would be flying blind on current state. With it, Claude starts every session knowing what matters right now.

Key design choice: Good startup protocols use on-demand retrieval (pull from KB) rather than loading entire files into CLAUDE.md. Each query returns a small, targeted result. Loading the same docs into boot context would cost 10x more tokens on every turn for the rest of the session.

Step 6: Session Is Live ~20,000 tokens consumed so far
Ready to work

At this point, Claude has consumed roughly 20,000 tokens (10% of the 200K window) just getting oriented. The remaining 180,000 tokens are available for actual work.

This is the most productive moment of the session. Claude has full context, maximum thinking room, and strong attention on all loaded instructions.

The takeaway: The boot sequence is an investment. You're spending ~20K tokens to make the remaining ~180K tokens count. A session without this startup would have 200K tokens available but produce worse results because Claude wouldn't know the current state of anything.

Watch It Boot

This simulates what happens when a Cowork session starts. Watch the context window fill up in real time.

Waiting to start...
Context used: 0 / 200,000 tokens Available: 200,000

Three Interfaces, Three Boot Experiences

Claude starts differently depending on which interface you're using. Click each to compare.

Cowork

  • System prompt with skills list
  • User preferences injected
  • CLAUDE.md auto-loads
  • @-imported rules load
  • Connected MCP tools listed
  • File access (workspace folder)
  • Computer use available
  • Sandbox shell available

Claude Desktop

  • System prompt (different from Cowork)
  • User preferences
  • CLAUDE.md auto-loads
  • @-imported rules load
  • Connected MCP tools listed
  • No direct file access
  • Computer use available
  • No sandbox shell

Claude Code

  • System prompt (minimal)
  • No user prefs injection
  • CLAUDE.md auto-loads
  • @-imported rules load
  • MCP tools if configured
  • Full filesystem access
  • No computer use
  • Native terminal/shell
Cowork: The richest boot experience. Anthropic's system prompt includes skill descriptions, file handling instructions, computer-use guidance, and your personal preferences. The trade-off is this system prompt is the largest of the three, eating more of your token budget before you say anything. Your CLAUDE.md then layers on top with your specific identity and rules.

The Priority Stack

Where something appears in the boot sequence affects how much "weight" Claude gives it. This isn't a bug — it's how attention works in transformer models.

Priority LevelWhat Goes HereExample
Highest: System Prompt Core behavioral rules that should never be overridden Safety rules, tool schemas, user formatting preferences
High: CLAUDE.md + Rules Identity and operational rules you want followed consistently "You are [role]," response rules, scope boundaries, knowledge map
Medium: Supplements Product context that frames the work but doesn't need constant enforcement Product feature list, current sprint state, pricing details
Active: Recent Messages The current conversation — highest attention, highest recency "Fix the login bug" or "Write a blog post about X"
Practical implication: If you want something followed every time, put it in a rules file (high priority, always loaded). If you want something followed right now, say it in your message (highest recency). If you want it available when relevant, put it in a knowledge base (zero boot cost, retrieved on demand). That's the decision tree for "where does this instruction live?"

What the Other Models Do Differently

Quick comparison so you start building a mental model of the landscape.

ConceptClaudeChatGPTGemini
Boot context file CLAUDE.md — automatic, file-based, version-controlled "Custom Instructions" or "GPT Instructions" — entered in a form, limited to 1,500 chars "Gems" — similar to Custom GPTs, less mature
Rules files @-import any .md file from .claude/rules/ Not supported natively Not supported natively
Tool connections MCP — open protocol, any tool can plug in Actions (OpenAPI-based), more restricted Extensions (Google ecosystem mostly)
Multi-session Independent sessions share nothing by default — you build the bridges (KB, messages, handoffs) Single conversation memory — "remembers" across chats but you can't control what it keeps Limited memory, tied to Google account
Your control Maximum. You control the files that control behavior. Moderate. Form fields, no file system. Minimal. Google controls the integration points.
Why this matters: Claude's approach (file-based configuration, open protocol tools, explicit context management) is more work to set up but gives you dramatically more control. ChatGPT's approach is easier to start but hits a ceiling fast. A multi-session setup with a shared knowledge base and structured handoffs is literally impossible in ChatGPT. Not hard. Impossible.

Key Takeaways

1. Boot order = priority order. System prompt outweighs CLAUDE.md outweighs supplements outweighs conversation. Put critical rules where they'll have the most weight.
2. Auto-load is expensive but powerful. CLAUDE.md and rules files cost you tokens on every turn. That's by design. They're the instructions you need followed consistently enough to pay the tax.
3. The startup protocol is an investment. Those 10 steps consume ~20K tokens, but they mean Claude starts every session with real situational awareness instead of just its rule book.
4. Architecture is everything. Modular rules, lean CLAUDE.md, knowledge base for retrieval, supplements for product context — these aren't arbitrary choices. They're each solving a specific context window constraint. Understanding why it works gives you the framework to make better decisions about where to put new instructions.

Next: Phase 1, Session 3

Tokens: The Real Currency — what a token physically is, why "200K context" is not 200K words, and counting real tokens from our CLAUDE.md.

We'll use a real-time tokenizer to see exactly what different types of content cost — and why your CLAUDE.md is pricier than it looks.