The line and the loop: what an AI-native SDLC looks like on a storefront
Anthropic recently published a playbook called the AI-native SDLC. We have been adopting it heavily for HI Engine development.
(full size, opens in a new tab)Release notes
The idea is simple. Code is no longer the bottleneck, so you rebuild the process around the parts that still run at human speed. Intent gets written down. Claude turns it into a spec, then a plan, then the change. Every PR is reviewed against rules we wrote once. Humans stop reading lines and start judging intent and risk.
It is working out nicely. We even have Arttu shipping code.
What we realized quickly is that HI Engine already runs this same loop, in principle, for conversion rate optimization (CRO).
Intent comes from the data. Agents pull GA4, Shopify and Click Context per store and emit findings that turn into CRO ideas. The merchant reviews the proposed change, the hypothesis, the evidence and a mockup, not a backlog ticket.
When the merchant likes the idea, the build runs from a queue into a draft theme, under rules that sit in the repo next to the rules for our own codebase. A verification gate screenshots every state on mobile and desktop before a human sees anything. If it fails, the draft is deleted and the merchant never knows it existed.
When the build is approved, the change is launched as an A/B test. Then the data decides. Revenue per visitor against a control. Winner baked into the theme, loser rolled back. The outcome is written down so the next discovery pass reads it. A store ends up with its own knowledge base: copy changes are dead here, imagery and price presentation move.
Same loop, one difference. A codebase has tests. A store has customers. So the merchant side has a stricter gate and a stricter definition of done. A shipped feature is judged by whether it works. A shipped storefront change is judged by whether it made money against a control.
In that one respect, the merchant side of the loop is the more mature one.