Ch. 1–2
Prompts
Powerful for one answer — but fragile, unrepeatable, and rebuilt from memory every time.
Out November 10, 2026PrinteBookPackt14 chapters
Claude Code Beyond the Prompt
Build reusable skills and agentic workflows. Move from one-off prompts to skills, workflows and supervised agents you can test, share and trust.
By Martin Krause · A practical field guide to agentic engineering with Claude and Claude Code.
Companion code on GitHub, free for every reader.
The shift
From prompt engineering to agentic engineering
Prompting collapsed the search-and-assemble loop into a single question. But a prompt is tied to one problem, depends on the context you remember to give it, and doesn’t come out the same way twice. This book follows one path from there to supervised agents, one level at a time.
Ch. 3–5
Skills
The method behind a recurring prompt, packaged in a SKILL.md that Claude recognizes and applies on its own.
Ch. 6–8
Workflows
Several reliable steps with inputs, checks, decisions and a stopping point — tested, maintained and shared.
Ch. 9–14
Agentic systems
Agents that gather, act and verify, connected to real tools through MCP — inside a supervised workbench.
Don’t start by asking “Can I make this an agent?” Ask “What loop does this task need?”
— from Chapter 2
What you’ll learn
Skills you keep when the tools change
Claude is the lens, but the patterns carry over to other agents too, including Copilot and Codex.
01
Replace ad hoc prompting
Spot the stable structure under your prompts and turn it into an explicit, repeatable method.
02
Turn prompts into testable skills
Write, trigger-test and refine skills that activate on the right requests — and stay silent on near-misses.
03
Build context-aware workflows
Capture inputs, steps, checks and decisions in a WORKFLOW.md your whole team can rely on.
04
Orchestrate agents and subagents
Know when a task needs a prompt, a skill, a workflow or an agent — and pick the smallest structure that fits.
05
Connect tools through MCP
Give agents real evidence through MCP while keeping retrieved content as data, not instructions.
06
Design your own AI workbench
Combine CLAUDE.md, permissions, hooks and validators into a supervised setup you control.
Preface
Before you begin
This book is hands-on from page one. Each chapter adds a level to the same project, so you finish with working artifacts, not just notes.
Who this book is for
- Developers, tech leads and product-minded builders who already use Claude or Claude Code and want consistent, repeatable results instead of lucky one-off answers.
- Engineers who keep pasting the same prompts and want to package that know-how as skills and workflows the whole team can share.
- Managers, founders and consultants who want a clear, practical picture of where agentic work is heading inside real teams.
You don’t need a machine learning background. If you’ve used Claude or another chat assistant and tried a few prompts, you’re ready.
To get the most out of this book
- Claude Code installed and signed in. The chat examples work in any Claude interface.
- Node.js 24 LTS or higher, for running ToDo Mate and the scripts in later chapters.
- Docker to run Kamori, the log server that ships with the example app.
- Git and a terminal to clone the companion repository and follow along chapter by chapter.
The example project
All examples use ToDo Mate, a small Vite and TypeScript todo app that saves
to localStorage. It comes with real bugs to investigate, from a blank page on
load to tasks that vanish after a refresh.
It runs alongside Kamori, a lightweight, self-hosted log server with a native MCP interface. Early on, you paste Kamori’s logs in as evidence. Later, Claude Code queries them directly.
Download the example code
The repository has one folder per chapter. Each folder holds the app as it stands in that chapter, plus the skills, workflows and Claude Code configuration you build there.
Conventions used
Skills, workflows and CLAUDE.md files are all Markdown. New to Markdown? Start with
our Markdown basics reference.
CLAUDE.md | Code in text — file names, commands, folder names and user input. |
```markdown | Code blocks — skills, workflows and prompts are plain Markdown files. |
trc_51y8un | Timestamps and trace IDs in log listings are examples. Yours will differ, so don’t copy them verbatim. |
Contents
14 chapters, one project
Every chapter leaves you with an artifact that the next one builds on. Open a chapter to see what you’ll make.
Part 1
Why prompting is no longer sufficient
Chapter 1 From Prompt to Skills: The Limits of One-off AI Use
From Prompt to Skills: The Limits of One-off AI Use
Why one-off prompting degrades through inconsistency, drift and context loss — and the ladder from prompt to pattern, method, template, skill, workflow and agent.
You build A structured diagnostic prompt
Chapter 2 What Agentic Engineering Actually Means
What Agentic Engineering Actually Means
A precise vocabulary for model, tool, agent, subagent, context and memory — and the agentic loop that connects them.
You build A workflow map of a recurring task
Part 2
Building reusable skills
Chapter 3 What a Skill Is: Instructions, Context, and Reuse
What a Skill Is: Instructions, Context, and Reuse
When a prompt earns promotion to a skill, the three-layer structure of metadata, instructions and resources, and progressive disclosure.
You build Your first SKILL.md skeleton
Chapter 4 Your First Skill
Your First Skill
The smallest complete loop: a narrow task, a description that is both trigger and promise, concrete steps, light guardrails and real tests.
You build A working bug-ticket-from-logs skill
Chapter 5 Designing Skills That Actually Help
Designing Skills That Actually Help
Scope with Always / Ask first / Never, move depth into reference files, and tighten triggers so a skill activates only when it should.
You build A refined, more reliable skill
Chapter 6 From Personal Prompt to Reusable Workflow
From Personal Prompt to Reusable Workflow
Separate the moment from the method and expose inputs, steps, checks, decisions and outputs as a shareable process.
You build A WORKFLOW.md for a repeated repo task
Chapter 7 A Skill, an Agent, or a Connected Workflow?
A Skill, an Agent, or a Connected Workflow?
A decision framework where the task defines the structure: the smallest shape that fits, worked through one bug at every level.
You build A task decision matrix
Chapter 8 Testing, Maintaining, and Sharing AI Workflows
Testing, Maintaining, and Sharing AI Workflows
Behavioral tests, maintenance notes and review boundaries that make a workflow explain itself once others depend on it.
You build A tested, shareable workflow pack
Part 3
How modern AI agents work in practice
Chapter 9 The Agentic Loop in Practice
The Agentic Loop in Practice
Gather, act, verify: run a real investigation end to end and fix the ToDo Mate persistence bug at its root.
You build An agentic debugging path
Chapter 10 Chat Assistant, IDE Agent, and Background Agent
Chat Assistant, IDE Agent, and Background Agent
Choose the environment from the work itself — what the agent must see, do, and when you need to step in.
You build An environment comparison for one task
Chapter 11 Claude Code and the New Working Environment
Claude Code and the New Working Environment
The agentic harness and its control plane: CLAUDE.md, context coordination, reusable capabilities, extensions and permissions.
You build A project-level Claude Code setup
Part 4
From local workflow to connected capability
Chapter 12 From Response to Action: Designing Tool-Backed Workflows
From Response to Action: Designing Tool-Backed Workflows
Action contracts, output contracts and a deterministic validator — with tighter approval the further a tool can reach.
You build A validator-backed action step
Chapter 13 MCP, RAG, and the Rise of Connected Agent Systems
MCP, RAG, and the Rise of Connected Agent Systems
Connect the investigation to Kamori through MCP, tell retrieval from connection, and treat what comes back as data.
You build A read-only Kamori MCP connection
Chapter 14 Designing Your Own AI Workbench
Designing Your Own AI Workbench
Assemble subagents, MCP, hooks and a human approval gate into one supervised setup — instruction steers, the harness enforces.
You build Your supervised AI workbench
About the author
Martin Krause
Martin Krause is an AI consultant, full-stack solution architect and engineering leader with over 20 years of experience delivering scalable digital platforms and advising global organisations.
Trained at Oxford Saïd Business School and LSE-certified in AI ethics, he applies AI and modern architectures to real-world business challenges. He writes regularly for Germany’s largest IT magazine, is the author of The Complete Developer, and builds AI-driven products that bridge cutting-edge technology with practical impact.
More at mkrause.info
Stop re-prompting. Start engineering.
Out November 10, 2026 from Packt. Pre-order now and get the book on release day.
PrinteBookPackt14 chapters