Why Tokens Are Your Real Currency
- Explain what a token is and why the context window is the fundamental resource in Claude Code
- Identify the six components that load into context automatically before you type your first prompt
- Apply the token-budget mental model to predict which session patterns will be cheap vs expensive
Tokens Are Not Just About Cost
Most developers learn about tokens when they get an unexpectedly high bill or hit a context limit mid-session. That is the wrong moment. Understanding tokens from the start changes how you structure sessions, write prompts, and choose models — and it prevents the habits that silently drain your budget.
A token is roughly four characters of text: a word, a punctuation cluster, a fragment of code. Claude does not read files the way you do — it loads their raw text into a shared pool alongside everything else in the conversation. Every file read, every command output, every tool result, every prior message occupies space in that pool. The pool has a ceiling, and once full, performance degrades and costs climb.
What Loads Before You Type Anything
Before your first prompt appears, Claude Code has already filled a meaningful share of the context window. This startup stack happens automatically on every session:
Core behavioral instructions, tool definitions, and response formatting rules. Always loaded first. You never see it, but it is always there.
Notes Claude wrote to itself in previous sessions: build commands it discovered, patterns it noticed, project-specific context. The first 200 lines or 25KB, whichever comes first.
Working directory, platform, shell, OS version, git branch, status, and recent commits. Injected automatically so Claude understands where it is.
Tool names and server instructions for configured MCP servers. Full schemas stay deferred until Claude needs a specific tool — a significant saving over loading everything upfront.
One-line summaries of available skills so Claude knows what it can invoke. Full skill content loads only when Claude actually uses one.
Your global preferences plus the project-root CLAUDE.md. Every line you write here costs tokens on every session — the most controllable part of the startup cost.
The Mental Model: Context Is a Budget, Not a Buffer
The useful mental model is not "context window = how much Claude can remember" — it is "context window = your working budget for a session." Every file read, bash output, and message you send draws from that budget. The difference between a $0.10 session and a $2.00 session is almost always how much the context window grew, not how many replies you got.
Two patterns drain the budget faster than anything else:
- Kitchen-sink sessions: mixing unrelated tasks in one long session. Every new task piles context on top of context that is no longer relevant to what you are currently asking.
- Broad exploratory prompts: "review this codebase" or "improve performance" triggers wide file scanning. A single debugging session that reads 30 files can consume most of the available space before any coding begins.
The rest of this track covers every lever you have for managing that budget: session commands, model selection, prompt habits, hooks, and more. This lesson is the foundation — understand what fills the window and the rest follows naturally.
- The context window is a shared pool for everything: system prompt, memory, environment info, CLAUDE.md, your messages, file reads, and tool outputs — all compete for the same space.
- Before you type your first prompt, 6,000–8,000+ tokens are already loaded automatically; your CLAUDE.md length is the most controllable part of that startup cost.
- Token cost does not scale with the number of replies — it scales with how large the context grows; a brief question in a large session costs more than a detailed question in a fresh one.
- Two patterns drain the budget fastest: kitchen-sink sessions (mixing unrelated tasks) and vague exploratory prompts that trigger wide file scanning.
- Understanding the startup stack is the prerequisite for every other cost-control technique in this track.