AI website failures rarely begin with invalid HTML. They begin with unmade decisions. The agent fills each gap locally, and the site slowly accumulates competing patterns, duplicated routes, fragile metadata, and more JavaScript than the page needs.
This library starts with the failure. Each guide explains how it appears, why a capable agent can still produce it, what to inspect, and which controls belong in source rather than in a one-time prompt.
Begin with the symptom you can observe. The production-gap guide audits the whole release. The SEO, performance, and design-consistency guides isolate narrower failure modes. When the cause is clear, move into the technical guides and change the build system that allowed it.
Constraints beat cleanup
A cleanup prompt can repair the current output. It cannot guarantee the next change will preserve the repair. Durable quality comes from scoped instructions, components that encode design decisions, build-time validation, and a review process that treats claims as carefully as code.
The documented site example shows how those layers appear in this repository without presenting the result as external customer proof.