Vibe coding
Vibe coding that only moves forward
Vibe coding is steering AI builders with clear goals while a control plane proves every step against the live product. This hub is how we run it: the loop, the rules, and the guide.
Step 1
Get set up first
Do this once, before reading anything else: download the free guide pack, open it in your AI coding tool, and paste one setup prompt.
Download the system zip into your project folder
Pick (or create) the folder where you want to work with your AI coding tool, and download Ratchet system guide v1.2 directly into it — no login, no form.
Open your AI coding tool in that same folder
Launch your AI coding tool (Claude Code, Cursor, or similar) with that folder as its working directory — the same one the zip just landed in.
Run the setup prompt
Paste this into your AI coding tool. It unzips the pack and finishes initial setup for you:
promptUnzip ratchet-guide-v1.2.zip in this folder, then read the pack in order: README.md, then overview → architecture → principles → layout → loop-and-missions. Follow the product contracts strictly: - Live deploy gate via an honest public version signal (tester judges the live URL only) - Builder proof-of-work from git state only (ignore agent claims) - Secrets only via a credentials boundary — never in builder env - Multi-step goals → multiple queue items - Optional infra ensure is fail-closed; prefer bound cloud project identities - Overnight helpers may observe only; they never implement product features Start with: loop + mock roles + mission shape validation. Then: goal capture + queue. Then: real builder/tester roles and a credentials boundary stub. Do not invent machine-specific install paths or operator runbooks.
Get the guide
Get the guide: Ratchet system guide v1.2
The complete Ratchet system guide, v1.2, as one free zip: overview, architecture, the loop contract, Composer, Vault, design principles, and the phase A–E rebuild checklist.
- The guide is free — the full v1.2 pack, no signup and no paywall.
- It contains no API keys, no vault passwords, and no host credentials.
- It teaches public-safe product concepts only — not install trees or host operations.
- It is product-design documentation, not access to anyone's running VPC.
Start free downloads the zip directly — no login, no form.
Want to be notified when there are updates?
Optional signup for Ratchet guide and product updates. Reading and downloading the guide never requires this form.
What vibe coding is — and what it is not
What it is
- Setting goals and constraints while an AI builder writes and pushes the code.
- Iterating in small, verifiable steps against the deployed product, not a local hope.
- A control loop: build, pass a live deploy gate, get tested, repeat until a streak of passes.
What it is not
- Not one mega-prompt expected to produce a finished product overnight.
- Not trusting an agent's claim of “done” — only the live site counts.
- Not a sandbox: no secrets in the builder environment, no unverified merges.
The loop
Every mission runs the same ratchet. It never moves backward:
1
Goal
A human states the outcome.
2
Multi-step missions
Queued as ~4–8 verifiable steps.
3
Build
The AI builder pushes code.
4
Live deploy gate
/version must report the new SHA.
5
Test
A tester grades the live site.
6
Streak of passes
Consecutive passes close the loop.
A FAIL sends the mission back to Build with the tester's report. Nobody babysits; the ratchet just holds.
Non-negotiables
1.Live is truth
The tester grades the deployed site at its live URL. Local trees and agent claims do not count.
2./version must report the deploy SHA
Every deploy answers with the actual git SHA, so the gate can prove what is really live before anything is graded.
3.No secrets in the builder environment
Credentials stay in Vault and are brokered per task. The builder environment never holds them.
4.Multi-step goals (~4–8 steps), never one mega-prompt
Real product work is queued as multi-step missions, each step small enough to verify on its own.
The mental model
Composer is the factory office where goals become queued missions, Ratchet is the factory floor that runs the build–deploy–test loop, and Vault is the key cabinet that keeps credentials out of the builder's hands.
Topics
Every card below renders from a single module list — adding a topic means appending one entry. Start with the guide; the remaining topics publish here as they ship.
Ratchet system guide
AvailableThe full documentation pack: overview, architecture, the loop contract, Composer, Vault, design principles, and the Mermaid diagram gallery.
Open module →
Rebuild checklist
Coming soonGreenfield product milestones in phases A–E: foundations, config rules, credentials boundary, first product, then hardening.
Published here when it ships.
Writing missions
Coming soonScoping goals into 4–8 verifiable steps with acceptance criteria a live tester can actually check.
Read the stub →
Live verify & testing
Coming soonHow the deploy gate and read-only tester grade the live product, and why only a streak of passes counts.
Read the stub →
Models & costs
Coming soonBuilder/tester model tiers, what a mission costs end to end, and where FAIL cycles add up.
Read the stub →
Case studies
Coming soonReal missions annotated from goal to streak of passes, with deploy-gate evidence and honest numbers.
Read the stub →
Composer walkthrough
Coming soonA guided tour of the goal-capture surface: drafting multi-step missions, product shells, and queue status.
Published here when it ships.
Vault deep-dive
Coming soonHow credentials stay behind a brokered boundary and never enter the builder environment.
Published here when it ships.
Build with a ratchet, not a leap of faith
Tell us what you are building. We will show you the loop running against your own live URL.