12 prompt-design patterns · 7 agent failure modes · the internal mechanics behind the world's leading agentic coding tool, inferred from how it behaves.
🌍 English · 简体中文 · 日本語 — all three editions included.
💖 Sponsor $25 to unlock the full book + toolkit · Read 2 free chapters ↓
Sponsor the $25 Inside Claude Code tier — you get the full book (EN + 中文 + 日本語 PDF) + toolkit, delivered within hours.
📘 The full book ships through GitHub Sponsors. Read the two free chapters below, then sponsor the $25 Inside Claude Code tier to unlock all 10 chapters + the toolkit. ⭐ Star the repo too.
Everyone is building agents. Almost no one has taken apart one that actually works.
Claude Code is the reference implementation of the autonomous coding agent — it runs for hours, edits dozens of files, tests its own work, and ships. Underneath it is a body of hard-won engineering: a prompt layer built as many small, single-purpose units rather than one block, a precise request-assembly pipeline, a context engine that survives long sessions, and a set of guardrails where the majority of the instructions describe what the agent must not do.
This book reverse-engineers that machine and hands you the patterns — organized, translated into plain English, and turned into things you can paste into your own agent today.
If you've ever watched an autonomous agent confidently declare a half-finished task "complete," there is a known fix for that. It's in Chapter 6.
- Engineers building agents on the Claude or OpenAI APIs who want production-grade patterns, not blog-post platitudes.
- Power users of Claude Code / Cursor who want to understand — and bend — the tool they live in.
- Anyone shipping LLM features who is tired of agents that stub out work, inflate their own grades, or panic near the context limit.
60+ pages · ~14,000 words · 10 chapters · a copy-paste toolkit.
| # | Chapter | You walk away with |
|---|---|---|
| 1 | The Five-Layer Prompt Architecture | Why a great agent's prompt is a filesystem, not a string |
| 2 | The Prompt Loading Flow | How one message becomes an API request — and where the cache boundary stops you re-billing the whole static prompt every turn |
| 3 | Twelve Patterns I — Behavior & Architecture | Constraint-first, single-responsibility, event-driven, cognitive boundaries, delegation, progressive safety |
| 4 | Twelve Patterns II — Runtime & Knowledge | Compaction, output efficiency, mode switching, the memory trinity, skillification, observability |
| 5 | Print Mode & Autonomous Loops | Signal handling, the auto-compaction threshold, designing a headless loop that won't wedge |
| 6 | The Seven Agent Failure Modes | Each failure, its root cause, and the verified fix |
| 7 | Sub-agents & Layered Delegation | The isolation contract — and the messaging asymmetry that silently kills most multi-agent setups |
| 8 | The Autonomous Development Pipeline | A full prompt-to-verified-build loop with an expert-review panel |
| 9 | Designing Your Own Harness | The whole book reduced to five buildable subsystems |
| 10 | The Checklist & Templates | A pre-flight checklist + paste-ready templates |
- 🗂️
CLAUDE.md.template— a constraint-first operating-rules file for any repo - 🤖
subagent-prompt.template.md— brief sub-agents so they don't silently fail - 🔄
context-handoff.template.md— survive a context reset without losing the thread - 📄 Patterns cheat sheet (1 page) + Failure-modes checklist (1 page) — print and pin
Every chapter is this concrete. Chapter 6 names all seven failure modes and gives each a verified fix. The modes:
- One-shot impulse — does everything at once, exhausts context, leaves fragments
- Premature "done" — declares victory with most of the work unfinished
- Context anxiety — rushes to wrap up near the context limit, with capacity to spare
- Self-evaluation inflation — grades its own broken work 9/10
- Skipping E2E — unit tests pass; the actual button does nothing
- Stub-ification — the UI looks complete; the interactions are hollow
- Spec cascade — a planner's one wrong detail poisons everything downstream
You've probably met at least three. The fixes are in the book.
No email required. Read the first two chapters right here:
- Introduction — Why Take a Coding Agent Apart?
- Chapter 1 — The Five-Layer Prompt Architecture
- Chapter 2 — The Prompt Loading Flow
🇨🇳 简体中文免费试读： 引言 · 第 1 章 · 五层提示词架构 · 第 2 章 · Prompt 加载流程
🇯🇵 日本語 無料サンプル： はじめに · 第1章 · 5層アーキテクチャ · 第2章 · プロンプト読み込みフロー
If the mechanism-level depth is what you've been looking for, the other eight chapters and the toolkit are one click away.
Sponsor the $25 Inside Claude Code tier → you get a private-repo invite with the full PDF (3 languages) + toolkit, usually within hours.
These patterns were distilled from careful observation of how Claude Code behaves, inference about the request pipeline that behavior implies, and Anthropic's own published guidance on long-running agent harnesses. It is independent analysis and commentary on observed behavior — it describes patterns in original words and does not reproduce any proprietary source code or prompt text.
Is this just the Claude Code docs? No. The docs tell you how to use the tool. This book reverse-engineers how it's built, and turns that into patterns for your own agents.
I don't use Claude Code — is it still useful? Yes. The patterns are model- and tool-agnostic. Chapters 3, 4, 6, 7, and 9 apply to any agent on any API.
What format? A PDF plus the raw Markdown, and the toolkit files. Yours forever, free updates.
How do I get it after sponsoring? Sponsor the $25 Inside Claude Code tier on GitHub Sponsors; you'll get an invite to the private repo with the full PDF + toolkit (usually within a few hours). Any tier at or above the book tier works.
Not happy? Message me and I'll make it right.
The two free preview chapters and the README are shared for evaluation. The full book and toolkit are commercial — please don't redistribute. © 2026. All rights reserved.
Study the machine that works, and you stop guessing at the one you're building.