Every review tool eventually produces the same artifact: a wall of comments. Feedback piles up faster than it resolves, and the screen everyone is discussing disappears behind the discussion. AI makes the wall grow faster — a model can produce thirty observations about a page before you've finished your coffee, which is thirty more sticky notes than anyone asked for.

Most of the design work on Snap Site's findings went into preventing that wall. The decisions turned out to be fairly general — they apply to any team's feedback process, with or without a tool — so here they are.

Where walls come from

Not from rude teammates. From three structural gaps.

Notes that describe locations instead of pointing at them. "The section near the bottom with the cards" — every reader has to re-find it before they can think about it.

Notes with no weight. When nothing carries a severity, everything is equally loud, and attention gets allocated by scroll position instead of importance.

And no cost to adding one more. So the volume grows until the surface stops being usable.

What a note needs

The annotations that survive a triage meeting tend to have three parts: the pattern, the evidence, and a direction.

Name the kind of problem — weak contrast, hidden primary action, unlabeled control. The vocabulary from heuristics and accessibility guidelines exists so you don't have to improvise one. Point at the exact evidence: this button, this state, this screen size. Then gesture at better — not a full solution, just enough that the reader knows which way to move.

"Low-contrast label — the credits counter fails contrast on the dark theme — bring it up a step toward the body-text gray." One sentence, fully actionable. "This feels hard to read" is shorter, and useless.

Severity and category are what let forty findings arrive as a sorted queue instead of a wall. Every Snap Site finding carries both for exactly that reason.

Put the note on the pixel

Anchoring is the highest-leverage decision in the whole system. A note pinned to the element it describes answers "where?" before anyone asks. A note in a sidebar list answers it never.

Pinning also disciplines the writer. "The page feels cluttered" has nowhere to attach — so it either sharpens into specific instances or admits it was a mood.

Anchor to a capture, not to the live site. Live pages change under your feet; an annotation on a dated capture is permanent evidence of what was true when someone looked. Skip this and you eventually get two people arguing about a problem only one of them can still see.

Keep the volume readable

AI review earns its keep on coverage. It checks the two-hundredth link label with the same patience as the first, which no human reviewer honestly does. The danger is coverage turning into noise.

Two rules help. Findings have to clear a bar — a real pattern, tied to real evidence, worth a human's reading time. Near-duplicates get merged, and observations that are true but inert get dropped. And the output is labeled as what it is: a draft. In Snap Site, the AI pass is explicitly a first pass you're expected to edit, reject, and add to. That framing isn't modesty. It's what keeps the human awake in the loop.

Notes should leave a record

Chat-style feedback evaporates — the thread resolves, collapses, and takes its knowledge with it. Audit annotations should behave more like an issue tracker: they stay attached to the capture, get marked solved when the fix ships, and remain visible as history.

Run the next audit and you can see what improved, what regressed, and what the team consciously left alone. A wall of comments proves a discussion happened. A set of resolved pins proves something changed.

The five-part test

For any feedback system — a tool you're evaluating or a process you're building: every note anchored to evidence, every note weighted and categorized, a quality bar that keeps volume readable, human judgment explicitly above machine output, and resolution tracked on the same surface as the feedback. Get those right and reviews get shorter and end in decisions. Miss them, and no amount of AI will keep the wall from going back up.