A production harness · copy-paste · runs on your machine
The Claude Code Setup Playbook — The 3 Hooks That Stop Fake "Done", Wrong Models & Wasted Tokens
A great model with a sloppy harness is a junior dev with amnesia — smart enough to write production code, but no memory, no policy on "done," no guard against runaway model spend. This is the harness that fixes it: three working hooks, a settings.json, and four skills, as a copy-paste bundle. Not a tutorial.
Instant download · no servers, no API keys, no subscription · ~27-page PDF + installable bundle
The problem
The model isn't the problem. The harness is.
A CLAUDE.md that reads like a wish list does nothing. Real production behavior comes from enforcement — hooks that block, redirect, and route.
▌ A sloppy harness
- Re-researches and rebuilds work you already did
- Every subagent bills at your premium model's rate
- Claims "done" on code it never ran
- A contract the model agrees to, then ignores under pressure
▌ This kit
- Surfaces your prior work before the agent rebuilds it
- Auto-routes cheap work to cheaper models
- Blocks a "done" claim with no verification
- Enforcement, not requests — hooks the model can't override
What it does
Three hooks that enforce the contract
Real Python. Fail-open (a broken hook never blocks you). Each has a one-env-var kill-switch.
Auto model-router
On a premium main loop, a subagent spawned with no model set gets auto-routed: sweeps → a cheaper tier, judgment (review/security/architecture) → a capable one. Never overrides a model you set.
Verify-gate
Changed code and claimed "done" with no test, run, or check this turn? It blocks once and asks you to actually verify. Catches the silent "it should work."
Recall-gate
Detects build/research intent and surfaces your own prior work from your memory notes before the agent starts from scratch. Kills the re-research tax.
Dogfood proof
Real receipts — the hooks actually running
Stock sellers ship prose. This ships terminal output from the real hooks.
# subagent-model-router — Opus main loop $ echo '{sweep task: "grep the CI logs"}' | router.py → routed to model: haiku # cheap work, cheap model $ echo '{judgment task: "security review of auth"}' | router.py → routed to model: opus # hard work, capable model # verify-gate — Stop hook $ echo '{edited auth.py, claimed "done", ran no test}' | verify-gate.py {"decision": "block", "reason": "you modified code and claimed it works, but ran no verification this turn..."}
What's in the box
Everything installs and runs — nothing fake
- ✓3 hook scripts — model-router, verify-gate, recall-gate. Real Python, fail-open.
- ✓settings.json that actually wires all three.
- ✓A self-contained dark-neon PDF tool (pandoc + WeasyPrint, fully offline).
- ✓3 more skills — verify-before-done, recall-before-build, route-by-cost.
- ✓~27-page playbook PDF explaining the architecture + the judgment behind every file.
- ✓Dogfood receipts — terminal output proving the hooks fire.
- ✓INSTALL + CUSTOMIZE guides and a knowledge-index template.
- ✓Model-routing economics — current per-token rates + the arithmetic.
⚠ Requirements — please read before buying
- Claude Code (the hooks use its hook system; 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 (one-time, free, offline)
- No API keys, no servers, no subscription for the core
Stop shipping on "it should work"
Give your Claude Code a production harness
Three hooks, a contract, and the judgment to tune them — installed in an afternoon. The harness an operator actually runs, not a course.
Instant download · 30-day refund, no questions
See all my Claude Code tools → expressive446.gumroad.com