Past the Prompt - Claude Code as an extensible system

Past the Prompt: Claude Code as an Extensible System

Most developers who use Claude Code stop at the chat window. They type, Claude responds, they copy the output. That is a perfectly functional workflow, and it leaves most of what the tool can do completely untouched.

This book is for developers who want to go further. Not further in the sense of better prompts, but further in the sense of a different mental model: Claude Code is not a chat interface with a code bias. It is an extensible agent platform, configurable at multiple layers, capable of running without a human in the loop, and designed to be composed into systems that dwarf what any single conversation can produce. Past the Prompt is a map of that platform.

The book is organized in three parts. The first builds the conceptual foundation and walks through the five mechanisms that let you extend the system: hooks, which make Claude react to events in your environment; MCP servers, which give the model access to tools and data outside its context; subagents, which introduce parallelism and context isolation; skills, which encode your conventions and load themselves when relevant; and plugins, which package all of the above into distributable units. The second part steps outside the single-agent frame and into the ecosystem around it: persistent memory, Plan Mode, coordinated agent teams, headless and CI/CD operation, and the Agent SDK for building custom agents to specification. The third part opens the hood on the implementation itself, drawing on Claude Code's leaked TypeScript source and a parallel academic analysis to explain how the agent loop, permission model, and context management actually work.

The organizing metaphor is John Searle's Chinese room, and a village of them. It is not ornamental. The room is the right model for thinking about what an LLM is and is not; the village is the right model for thinking about what a multi-agent system produces and why. Neither understands. Both are useful. The tension between those two facts runs through every chapter, and resolving it is not the point. Working productively inside it is.

This is not a book from inside Anthropic. I use Claude Code the way you do, with the same rate limits, the same context drift, the same occasional output that would have been faster to write by hand. The perspective is that of a daily practitioner who found it useful enough to spend a book on, and honest enough to say where the limits are.

What you get: 400+ pages across 20 chapters and 2 appendices, covering everything from first principles to the internals of the agent loop. Originally published as two Italian volumes; this English edition unifies them and adds a third part of new material.