Two read-only agents that turn design QA into a specification.
One compares a live Shopify storefront with its design file and writes down every place they disagree. The other reads a live site with no usable design file and writes the style guide that should have existed. Neither may change anything it looks at.
What needed
solving.
A direct-to-consumer pet brand was rebuilding its Shopify storefront, and part of the homepage had been built against a design file that had since moved on. Someone had to reconcile every section in both directions: built but not designed, designed but not built, and built wrong. That work is slow, it is the first thing dropped under deadline, and the cost of dropping it arrives later as rework. Brands also arrive at a migration with no written design system.
What I decided,
and why.
- 01
The agent audits. It does not fix. It has read access, writes nothing, and ends its report by stating that nothing was written to the theme or repo. Letting one agent find and fix in the same pass removes the only point where a wrong finding can be stopped.
- 02
The agent has to say what it could not determine. Both prompts require a list of unresolved items with reasons, such as a disabled button state that only a sold-out product would show. An agent that fills those gaps with plausible values produces a document that reads as finished and is wrong in places nobody will check.
- 03
Naming the authoritative source is my call, not the agent’s. When the design link was stale and the canvas held two conflicting versions, the agent picked the later one so the work could continue, said so in writing, and flagged the other, so I could overrule the choice.
What shipped,
and what isn’t proven.
The storefront audit goes section by section (settings changes, sections to delete, missing sections, final order, things the theme cannot produce today, the ambiguous calls), and it went to me, not to the theme. The style guide covers color, type, buttons, form fields, spacing, grid, radius, shadow, motion, icons and imagery, with a matching token file and its own list of values it could not determine. Not proven: whether either agent finds more or fewer real defects than a person doing the same pass, and at what false-positive rate. I have no measurement of that.