A great model with a sloppy harness is a junior dev with amnesia. The Claude Code Setup Playbook is the production harness, packaged so you can install it in an afternoon — not a tutorial, a copy-paste bundle of real, tested files. WHAT IT IS Three working guardrail hooks + four skills + a behavioral-contract playbook, all of which run on YOUR machine with no servers, no API keys, and no subscription. THE THREE HOOKS (real Python, fail-open, each with a kill-switch) • Auto model-router — when your main loop is on a premium tier and you spawn a subagent without setting a model, it auto-routes the subtask: recall/sweep work to a cheaper tier, judgment work (review, security, architecture, diagnosis) to a capable tier. It never overrides a model you set, and stays silent on cheaper main loops. Cuts the bill on the work that doesn't need the expensive model. • Verify-gate — blocks a "done / it works / fixed" claim when you changed code but ran no verification that turn. Stops the agent from declaring success it didn't earn. Fail-open: it can never wedge you, and one env var turns it off. • Recall-gate — surfaces your own prior work from your memory notes before the agent rebuilds something you already built. Quiet until you have notes; then it stops the re-research tax. ALSO IN THE BUNDLE • settings.json that actually wires the three hooks (copy-paste or merge) • Four skills: a self-contained dark-neon PDF tool (pandoc + WeasyPrint, fully offline — no external services) plus verify-before-done, recall-before-build, and route-by-cost • A ~27-page playbook PDF that explains the architecture and the judgment behind every file, so you can tune it instead of fighting it • Real dogfood receipts: terminal output showing the router sending a sweep to a cheaper tier and a security review to a capable one, and the verify-gate blocking a premature "done" • INSTALL, CUSTOMIZE, and a knowledge-index template THE MODEL-ROUTING ECONOMICS (honest) The playbook includes the current published per-token rates and shows the arithmetic: routing the work that doesn't need a premium model to a cheaper tier costs a fraction of the premium input rate. These are illustrative calculations from published rates, not a measured benchmark — your actual savings depend on your token counts. REQUIREMENTS (please read before buying) • Claude Code (the hooks use its hook system; the same patterns adapt to other agents) • macOS or Linux — Windows works via WSL • Python 3.8+ (already on Mac/Linux) • For the PDF tool only: pandoc + WeasyPrint installed (one-time, free, offline) • No API keys, no servers, no subscription for any of the core WHAT THIS IS NOT This is the transferable harness — three hooks and the contract behind them — not a clone of one person's entire private setup. Every file ships in a form that runs on your machine; nothing depends on infrastructure you don't have. Built and run by an operator, not a course-seller. 30-day refund, no questions. Claude is a trademark of Anthropic PBC. This product is not affiliated with, endorsed by, or sponsored by Anthropic.