A prototype that becomes the site
Conventional process places a translation step in the middle, and translation always loses something. A designer produces an artifact describing intent. A developer interprets that artifact. Interpretation introduces drift. Review catches some of it, but budget runs out, and the rest ships.
"The design is not a description of the site. It is the site’s front end."
Composing instead of assembling
Production is where design teams lose their weeks, so this is the part worth reading twice. A page gets built by composing it. Type a slash, and a command palette opens. Place a header treatment, an article body, a card grid, an FAQ, a pull quote, a contact bar. The order of the chips is the order of the page. Beneath the composer, a live legend lists every available command, style and piece with descriptions, generated from the registries those commands draw on. The legend grows as capabilities ship rather than aging into fiction the way documentation always does. Reordering by dragging recomposes the live preview. The header chip stays locked, because a page can never ship headless.
Direction attached per component
Any chip can carry notes steering only that piece: tone, what to emphasize, what to include. A marker shows which chips carry notes, so art direction stops being one paragraph applied to an entire page and becomes specific to the component it governs. Product knowledge is inherited rather than retyped. Direction for how a family of products should be introduced, illustrated and explained lives as a shared object that pages inherit, so what is learned art-directing one category improves every page in it.
Rebuilds default to keeping what exists
Every part of an existing page starts on keep. A rebuild changes only what somebody deliberately flips. Remove a piece and its content stays stored, so putting it back restores the original exactly. Metadata is protected more firmly still, with a visible record of when it was last written. That default matters more than it sounds, because the fear of losing work is why teams stop iterating. When a mistake is cheap to undo, people try things.
Preview, publish, roll back
A draft renders through the real shell before anything goes live. History is kept, while rollback takes one click. What gets approved is what launches. Design confidence is much easier when mistakes are cheap, and the absence of a safe rollback is why so many sites freeze after launch. A team that cannot undo an edit stops making them, and a site nobody touches decays on schedule.
The picker only offers what can actually be filled
Patterns declare their requirements while the site reports its capabilities, so anything unsupported is simply absent from the menu. An information site never sees a product grid option. Nothing errors and nothing renders empty, which sounds like a small thing until you have watched somebody insert a component that produced a blank region and quietly lose faith in the whole system.
Review becomes a design conversation
With no interpretation layer to lose fidelity in, the whole exercise changes character. Instead of comparing a build against a file and cataloging discrepancies, review becomes an actual conversation about whether the thing is right. It also changes who can safely touch the site afterward. Copy edits, image swaps, menu changes and new pages composed from existing patterns become available to whoever runs the marketing, in the admin they already use, without a ticket. The designer gets called back for design decisions rather than for maintenance.
Where prototype work connects
The pattern library composition draws from is covered in design systems. The ingestion contract that produces it sits under the Model Context Protocol, and generated content arriving in the build is covered in AI-assisted design.
How the engagement runs
Composition work happens through screen shared sessions in the real environment rather than in a prototyping tool that will later have to be translated. If your last redesign involved a handoff document and a list of discrepancies, the translation step was the problem rather than anybody’s execution.