From plan to merged code
How it works
A plan everyone agrees on, cut into tasks small enough for one AI agent, built on contributors’ own computers and checked by someone else before it merges.
The seven steps
One public roadmap per target
Every product we replace gets one public roadmap, for example the “Zoom Replacement Roadmap”. Anyone can suggest changes to it. Nobody starts a competing roadmap, so all the effort goes into one plan.
The roadmap is a single canonical GitHub pull request. Contributors propose changes to it; forks of the roadmap are not used.
Two AIs from two labs must agree
Two AI models from two different companies review the roadmap: Fable (Claude, by Anthropic) and Astra (ChatGPT, by OpenAI). Each one tries to prove the roadmap is incomplete, without seeing the other’s answer. We revise and repeat until both say there are no material gaps. That is roadmap consensus.
Both reviewers run at maximum reasoning effort and review independently. A round ends only when both return “no material gaps”.
Every feature gets a contract
Each feature on the roadmap gets a Feature Contract: a precise description of what the feature must do. It is reviewed the same way, by both AIs, until they agree.
Feature Contracts go through the same independent two-model review loop as the roadmap.
Contracts are cut into small tasks
A contract is broken into Atomic Build Units: tasks small enough for one AI agent to finish. A map of which tasks depend on which lets many people build at the same time without getting in each other’s way.
Units form a dependency graph. Independent units are scoped so they never touch the same files, which lets them be built in parallel.
You press BUILD
Open wOS Desktop (or the wos command-line tool), pick a target, a feature and a task, and press BUILD. wOS reserves the task for you, sets up a clean copy of the code, tells the Claude Code on your computer exactly what to do, keeps it to the files it is allowed to change, and runs the tests.
wOS leases the unit to you, creates an isolated git worktree, hands your local Claude Code a context manifest for that unit, enforces a file allow-list, and runs the unit’s tests. It runs on your own Claude subscription.
Someone else checks your work
Your work is reviewed by Fable and Astra running on other contributors’ computers. You never review your own work. Then automated checks run everything again. Only after that does wOS open the real pull request on GitHub.
Review runs on other contributors’ machines, never the author’s. CI then reruns every check before wOS opens the GitHub PR.
Merged work moves the needle
When your work is merged, the public progress numbers go up, the tasks that were waiting on yours unlock, and you earn WOS tokens.
Merges update MAPPED, SPECIFIED and BUILT, unlock dependent units, and credit tokens. WOS tokens are in-app credits with no cash value.
How progress is measured
Each target shows three numbers. They are independent: a product can be fully mapped while almost nothing is built.
Today every target is at 0% on all three. See the Sniper List.
- Mapped %
- How much of the product is written down on the public roadmap.
- Specified %
- How much has an agreed, reviewed plan for exactly how each feature works.
- Built %
- How much has been built, checked and merged.
Why two AIs?
One AI can miss what it does not know. Two models from two different labs, trained differently and working without seeing each other’s answer, are much less likely to miss the same thing. Nothing moves forward until both agree.
Fable is Claude, made by Anthropic.
Astra is ChatGPT, made by OpenAI.
Both run at maximum reasoning effort for every review.