The Context Engineering Handbook

Chapter 5. Case Study: How to Build a Great System Prompt

by Srihari Sriraman

A Note for Early Release Readers

With Early Release ebooks, you get books in their earliest form—the author’s raw and unedited content as they write—so you can take advantage of these technologies long before the official release of these titles.

This will be the 18th chapter of the final book. Please note that the GitHub repo will be made active later on.

If you’d like to be actively involved in reviewing and commenting on this draft, please reach out to the editor at [e-mail].

When designing an AI application, the system prompt is the first piece of context you control. It sets the baseline instructions that are present in every call to the model, preceding every skill, file, tool result, and user input.

System prompts are artifacts you can study. Model weights are frustratingly opaque, but system prompts can be extracted, studied, and compared. They reveal what an application needs the model to know, the behaviors its product team desires, and the model tendencies the team is trying to correct. Simon Willison notes that system prompts “double as better written documentation than any of the official sources.”

To understand their range, we compared system prompts from six coding agents (Figure 5-1): Claude Code, Codex CLI, Cursor CLI, Gemini CLI, OpenCode, and Pi. Each application essentially performs the same function: given a task, gather information, understand the codebase, write code, ...