First edition — 2026
A practical handbook for the Claude Code era.
A field manual for engineering leaders who have to ship real software with AI agents — without making the codebase, the org, or the budget worse.
Most published guidance on AI-assisted coding is either vendor marketing or doom-mongering from people who haven't opened a terminal in a decade. This handbook is neither.
It's an operating manual for engineering leaders running real codebases through Claude Code and the tools that will replace it next quarter. It covers the patterns that work — legible repos, deterministic hooks, scoped PRs, explicit autonomy levels — the failure modes that keep recurring — AI slop, prompt injection, runaway token bills — and the conversations you have to have with your CEO, your board, and your security team.
Anchored on Claude Code, but the principles port to whatever ships next.
A dozen chapter highlights — not a comprehensive list.
The seven canonical signatures of AI-generated code that passes review and breaks production: mocked-implementation tests, deleted edge cases, silent error swallowing, weakened validation, removed security checks, gratuitous abstractions, diff bloat.
Your CLAUDE.md, skills, and hooks are code too — so benchmark them. A small golden-task set that catches regressions when you change the harness or the model, because deterministic verification is the one thing the agent's self-congratulation can't corrupt.
Why the harness, not the model, is what keeps an agent from doing the wrong thing — and how to wire hooks that block dangerous actions before the LLM can rationalize them.
The line item your CFO has not yet found, with the math on why "we'll just use Opus everywhere" turns into a six-figure surprise.
A graduated framework for what an agent is allowed to do unsupervised, supervised, or never — so "AI autonomy" stops being a vibe and becomes a policy.
The exceptions the rest of the book owes you: hard real-time, novel cryptography, ML experiment loops, compilers, database internals, and driver code — domains where today the agent's output costs more to repair than to write by hand, and why most of that list will shrink.
Why every major 2025–2026 agent incident has the same root cause — and how to architect around it.
The questions to actually ask AI vendors, the answers you can write yourself, and which contractual terms are non-negotiable.
The higher-leverage half of the senior week: incident triage, on-call digests, vendor reviews, RFPs, architecture-review prep, performance reviews. A 6x speedup on a vendor review changes whether the review happens at all — draft, never publish.
The steelman chapter, where the author argues against his own thesis — that the model eats the harness, that local LLMs win sooner, that vibe-coding is fine in narrow domains — and states the falsifiable conditions under which the book would need a second edition.
Week-by-week rollout for a mid-size org, from baseline measurement through CLAUDE.md adoption to measurable PR-throughput gains.
The four-slide deck, the talking points, and the specific pushbacks for when leadership demands a 50% headcount cut by Q4.
Every artifact referenced in the book lives in the open-source companion repo: CLAUDE.md and AGENTS.md templates, the twelve starter skills, the starter subagent roster, the hook library, the prompt pattern library, the AI code-smell and test-review checklists, the fillable AI-readiness scorecard, the runnable prompt-injection test suite, the autonomy ladder, the Do-Not-Automate catalog, the 90-day plan, the incident postmortem templates, and the executive kit.
Join the waitlist for launch notification, sample chapters, and the companion starter kit — CLAUDE.md templates, hooks, and the 90-day rollout plan.
Or email ryan@codingwithaibook.com directly.