I Reverse-Engineered Claude Code: Steal These Agent Tricks

Summary

In this video, the creator explores the inner workings of Claude Code, revealing how prompt engineering drives its superior performance compared to other coding agents. By reverse engineering Claude Code's interactions with the Anthropic API, the creator uncovers how detailed system prompts, tool definitions, and task management workflows allow Claude to execute tasks with high precision. Key insights include the importance of clear workflows, the role of sub-agents, and the significance of prompt tuning specific to model families. The video also highlights how prompt engineering continues to shape the development of coding agents in 2025.

Takeaways

• 😀 Claude's code performs exceptionally well due to sophisticated prompt engineering, despite using the same model architecture as other coding agents.
• 😀 A detailed reverse engineering process revealed that Cloud Code's prompt and LM interactions are dynamically constructed rather than static text strings.
• 😀 The use of a proxy allowed the interception of requests made between Cloud Code and the Anthropic API, revealing key interactions and tools used by the system.
• 😀 The system prompt plays a crucial role in shaping the agent's responses and task management, providing detailed instructions and examples on handling tasks.
• 😀 Key workflows, such as task management, are reiterated across the system prompt to ensure consistency and reliability in how tasks are completed.
• 😀 Cloud Code employs a unique approach by using tools like the 'to-do' tool for task planning, which ensures the model consistently remembers tasks and follows the workflow.
• 😀 Prompts in Cloud Code are designed to be highly human-readable and formatted to provide semantic meaning to the LLM, enhancing task execution accuracy.
• 😀 The introduction of sub-agents allows the main agent to delegate tasks to specialized agents, maintaining a clear separation of responsibilities and message histories.
• 😀 Sub-agents are created dynamically with their own system prompts, and their results are integrated into the main agent’s workflow once tasks are completed.
• 😀 The tool section defines available agents and their workflows, providing detailed instructions for triggering sub-agents, ensuring that the agents perform complex tasks autonomously.
• 😀 Successful prompt engineering for coding agents is highly model-specific, with prompt tuning being crucial for tool calling accuracy, especially when switching models or external providers.

Q & A

Why does Claude Code feel more efficient than other coding agents despite using the same models?

Claude Code feels more efficient because it utilizes sophisticated prompt engineering, which dynamically constructs prompts and manages tasks effectively. Its key to efficiency lies in detailed, well-structured prompts and tool orchestration that maximize the LLM's performance.

What tool did the creator use to reverse engineer Claude Code?

The creator used a tool called WebCrack to unbundle and deobfuscate the compiled .js file of Claude Code, which revealed over 443,000 lines of code for analysis.

What was the creator's initial assumption about Claude Code, and how did it turn out?

The creator initially assumed Claude Code would be an open-source project, but it turned out to be a proprietary tool with a bundled 9 MB CLI .js file for analysis.

How did the creator capture the interactions between Claude Code and the anthropic API?

The creator used a proxy tool, ProxyMan, to intercept requests between Claude Code and the anthropic APIs, enabling them to capture the full interaction, including the user message, system prompt, and tool calls.

What role does the system prompt play in Claude Code's functionality?

The system prompt defines the core behavior, workflow, task management, and tool usage for Claude Code. It guides the LLM’s responses, ensures consistency in task execution, and structures how tasks are planned and tracked.

Why does Claude Code frequently remind the model about the to-do tool?

Claude Code uses constant reminders about the to-do tool in the system prompt and message history to ensure consistent task management and planning, which is key to accurate and efficient task execution.

What is the significance of sub-agents in Claude Code's design?

Sub-agents in Claude Code are triggered by the main agent to handle specific tasks independently, using their own system prompts and message history. This separation ensures that complex, multi-step tasks are managed autonomously while maintaining the integrity of the main agent’s workflow.

How does Claude Code ensure that sub-agent tasks remain independent from the main agent’s memory?

Claude Code ensures that the sub-agent's message history is discarded after completing its task. This prevents memory sharing between the main agent and sub-agent, ensuring that each agent remains focused on its own tasks and avoids unnecessary rework.

Why is the detailed description of tool calls important for Claude Code's accuracy?

Detailed descriptions of tool calls, including when and how to use each tool, ensure that Claude Code performs tasks with high accuracy. These descriptions help the model to understand the correct application of tools in various situations, making it more reliable in executing tasks.

How does Claude Code handle formatting within its system prompt?

Formatting within Claude Code’s system prompt, such as using all caps or XML tags, plays a crucial role in guiding the model’s comprehension. It helps structure the information semantically, enabling the model to interpret and execute tasks more effectively.