How Claude Code Builds a System Prompt

中文 | English

Source URL: https://www.dbreunig.com/2026/04/04/how-claude-code-builds-a-system-prompt.html Captured: 2026-04-16

Summary

This article explains Claude Code's system prompt as a dynamically assembled context rather than a single static string. Using an accidental source-code leak as evidence, it maps the component sections that can enter the prompt, including global instructions, coding philosophy, tool-use policy, communication rules, subagent guidance, skills, memory behavior, MCP instructions, scratchpad rules, git context, and other conditional suffixes.

Key Points

• System prompts in serious agent products are assembled from many conditional sections, not authored as one constant paragraph.
• Claude Code's prompt builder includes both always-on sections and conditionally injected sections.
• The article treats system prompts as one layer inside a larger harness context that also includes tool definitions, user content, conversation history, attachments, and skills.
• The assembly logic itself is important product evidence because it shows how much context engineering happens outside the base model.
• The article does not republish the leaked source code, but it distills its structure into a visualization and descriptive breakdown.

Notable Components Mentioned

• Global system framing and response style
• Tool-use policy and preference for dedicated tools
• Coding philosophy and minimal-change norms
• Risk confirmation guidance for destructive or external actions
• Communication and verbosity rules
• Dynamic cache boundary markers
• Clarification and user-question rules
• Interactive-shell guidance
• Subagent and fork guidance
• Skills discovery and invocation
• Independent verification requirements
• Auto-memory behavior
• Runtime / model / environment suffixes
• User language preference and output style
• MCP server instructions
• Scratchpad and function-result clearing behavior
• Git status context and appended system-prompt suffixes

Extracted Excerpt

System prompts generally aren’t static strings; they’re dynamically assembled contexts with many conditional statements determining what makes it in the prompt.

When you type instructions and hit ‘Enter’, Claude Code assembles a rich context to increase the odds that it obtains a successful response from Opus or Sonnet. As we can see, agents are more than just models. Context engineering is critical.

Practical Notes For This Vault

• This source complements reverse-extracted prompt inventories by explaining how a harness decides which prompt fragments appear at runtime.
• It is especially useful beside sources that catalog Claude Code prompt fragments, because it adds assembly logic and conditional structure rather than only prompt contents.
• The article strengthens the case that prompt/control surfaces, skills, memory systems, MCP instructions, and tool definitions should be treated as one interconnected operating layer.