The system
The whole company is a folder of markdown files, organized as a seven-station assembly line and operated by AI agents. This is what the stations are, how often each one runs, and what it has to hand over before anything moves.
It is not a waterfall
S1 to S4 is a chain each product walks exactly once. Everything after that is loops. The moment a product is live, S5 and S6 both start and neither ever finishes — growing it and keeping it alive are orthogonal jobs, not consecutive ones, and they run against each other’s findings rather than in sequence. S7 then cuts vertically through every live product once a month.
So several products sit at different stations at the same time, and a product that has shipped is never “done” — it is in two loops until the month it gets killed. Upstream of all of it sits a standing research engine that feeds S1 and never joins the chain itself.
The stations
- S1discover
- Scan lead sources for early signals and drop the survivors into a radar pool.Fortnightly · hands over: The idea radar
- S2validate
- Five-step evidence chain — demand, competitor revenue, distribution, platform risk, social proof — ending in one of four verdicts.Event-driven · hands over: A dated feasibility file, and a numeric prediction S7 can hold it to
- S3commit
- Candidates compete for an open slot; the winner gets a PRD, a revenue model and a broken-down backlog.When a slot opens · hands over: A product folder and its issues
- S4build
- Implementation and launch, in the product's own repo behind an engineering checklist.From commit until it ships · hands over: A live product and a signed acceptance sheet
- S5growloops forever
- One launch sequence, then acquisition loops — SEO, communities, social, email — that never end.Daily to weekly, per channel · hands over: A GTM plan, channel logs, and a monthly review
- S6operateloops forever
- Monitoring and repair. Once a product is live this is the only route to production, which is what keeps a fix from skipping the checks.Daily to weekly, per product · hands over: A monitoring log and closed issues with proof they were re-checked in production
- S7reconcile
- Predictions against actuals for every live product, ending in scale, iterate, maintain — or kill.Monthly, whole portfolio · hands over: One verdict line per product
One human gate
Agents draft everything — the research, the code, the SEO, the kill memos. Nothing reaches the public without me pressing the button. That gate is the reason this is a company and not a content farm, and it is also the bottleneck: the honest limit on how fast any of this can go is how much I can read in a day.
Why any of this is public
Because a pipeline that only ever gets described in the abstract is unfalsifiable. The station that matters most here is S7, and it is the one most build-in-public accounts quietly skip: every live product gets a verdict every month, and the board carries the products that lost. The playbooks, the kill memos and the real numbers land in the operator log first.