How big is the Claude Code system prompt?

The Claude Code system prompt is approximately 15KB (15,000 characters) and is assembled from 7 distinct layers on every API call. The first portion is static and cached, while the bottom portion is dynamic and unique to each user.

What is the Claude Code cache boundary trick?

Claude Code splits its system prompt at a cache boundary marker. Everything above the marker — the core instructions and Anthropic rules — is identical for all users and gets cached by the Anthropic API at a 90% discount. Only the user-specific section below the boundary (your MEMORY.md, environment info, MCP tools) is sent fresh each time.

How does CLAUDE.md get injected into Claude Code?

Claude Code reads CLAUDE.md files from your project directory and home directory, then injects their contents into the user message — not the system prompt. This is done so the dynamic content does not break the system prompt cache boundary, preserving the 10x cost savings on input tokens.

CHAPTER 03 — SOURCE ANALYSIS

The Hidden Instructions

Every time you talk to Claude Code, it reads a secret playbook first — 15,000 characters of rules, memory, and context, assembled in real-time. This chapter shows you exactly what's inside, layer by layer.

What is a system prompt?

Before Claude Code sees your message, it receives a set of hidden instructions from Anthropic. Think of it like a briefing document that a new employee reads on their first day — it tells the AI who it is, what it can do, what to be careful about, and everything it knows about you and your project.

The clever engineering: this briefing is split into two halves. The top half is identical for every user in the world and is cached (shared) — saving 10x on cost. The bottom half is unique to you — your memory, your environment, your tools.

15KB

total size

layers

10x

cost savings

uncached section

The 7 Layers

Click any layer to explore what’s inside. Layers above the green boundary are cached globally.

Core Identity

Who the AI is and how it should behave

~1KBconstants/prompts.ts

Coding Philosophy & Rules

How to approach tasks, use tools, and communicate

~7KBconstants/prompts.ts

Above: $0.50/MTok cached

10xsavings

__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__

utils/api.ts

Below: $5.00/MTok freshper-session, unique to you

Your Memory (MEMORY.md)

Everything Claude Code remembers about you across sessions

0–25KBmemdir/memdir.ts

Environment Info

CWD, git branch, OS, model, knowledge cutoff

~0.5KBprompts.ts

MCP Server Instructionsuncached

The only section that recomputes on every single turn

variessystemPromptSections.ts

CLAUDE.md (Project Instructions)user message

Injected into the user message — NOT the system prompt

variesutils/claudemd.ts

How Claude Code Assembles the Prompt

Every time you send a message:

1Load static rules (cached, ~8KB)$0.50/MTok

2Add your MEMORY.mdunique to you

3Add environment infoyour CWD, OS, model

4Check MCP serversrecomputed every turn

5Inject CLAUDE.md into user messagenot in system prompt!

6Send to APItotal ~15KB of context

Deep Dive

How the system works under the hood — click to expand

Source Files

constants/prompts.tsSection functions, assembly, boundary marker

constants/systemPromptSections.tsSection registry, memoized vs uncached

utils/claudemd.tsCLAUDE.md discovery, @include, conditional rules

memdir/memdir.tsMEMORY.md loading, truncation (200 lines / 25KB)

context.tsgetSystemContext, getUserContext

services/api/claude.tsbuildSystemPromptBlocks, cache control