I recently published the most detailed thing I’ve written about how I run Eterna Clarity as a solo founder. It started as an attempt to explain a strange result: in less than three months, I had built an operating environment that could coordinate a surprising amount of real company work from one ordinary Windows PC.

The part I think is most useful is not the amount of automation. It is the boundary I eventually had to draw around the AI itself. Eterna can use different frontier AI providers, a qualified local model, ordinary software, connected services and browser tools, but the company does not live inside any of them. Current work, files, decisions, research and operating state have durable owners outside the conversation.

I learned that distinction through failure. At one point I had systems confidently report that work was complete when the real result was wrong. I had a local-model candidate score 24 out of 24 on its target benchmark and still rejected it because it failed an older adversarial test. I built orchestration that looked sophisticated and then measured it against a simpler approach and found that the simpler system was substantially better.

Those experiences changed what I optimize for. I care much less now about how autonomous an AI looks and much more about whether its work can be verified, whether another model could continue the same job tomorrow, whether a failure leaves the company in a recoverable state, and whether human judgment is still being applied where it actually matters.

The same thinking now reaches into creative work at Eterna. Studio has explicit standards for AI-assisted writing and design, including evidence before claims, accurate attribution, human voice and rejection of output that looks polished but does not survive closer review. I am also constantly testing new models and figuring out where they actually belong in production workflows rather than assuming that a new capability is automatically useful.

I wrote the full blueprint partly because I think other small founders and independent operators can build versions of this without copying Eterna’s complexity. The article includes what worked, what failed, what I would never build again, and a much smaller progression someone could use to start with durable work and gradually add AI around it.

I’m especially curious what people here think because Contra has such an interesting mix of designers, independents, builders and people working directly with AI. If you read it, I’d genuinely value criticism. What feels useful? What looks overbuilt? What would you challenge if you were designing the system yourself?

https://eternaclarity.com/editorials/articles/highest-leverage-small-sized-operating-system-blueprint/