/visual-verify

Role. You are shipping something a human will look at: a page, a chart, a diagram, a deck slide, a 3D scene. Your usual proof — "it compiles, the tags balance, the script parses" — proves nothing about whether it looks right. This skill is the render-and-look loop that closes that gap.

Context. Three invariants drive everything below.

1. A visual deliverable is not done until you have seen it. Structural validation cannot detect overlapping text, a camera inside its own geometry, dots labelled with the wrong number, or a legend wrapping into an 8-character column. All of those ship green.
2. You can almost always render locally. A Chromium binary is usually cached on the machine even when Playwright itself is not installed. Software WebGL works.
3. Animation state is invisible to a timed screenshot. This is the trap that costs the most time — see The settle problem.

Task. Build → render → look → instrument what you cannot see → force the settled state → re-look → only then report. Never collapse this into build → validate → report. How much of this loop a given change requires — one shot vs. the full cartesian matrix — depends on the verification tier; §9 defines the three tiers and when each applies.

1. Render it

scripts/render.sh wraps the whole invocation. Manually, the shape that works: