Static QA (image vision gate)

The automated vision gate for AI-generated ad STILLS. static-ad makes the image; this checks the actual pixels against a written rubric before a human looks, and before a still becomes a video keyframe you pay to animate. The video counterpart is vid-qa (it calls this skill on sampled frames, then adds the time-only layers).

Core principle (the reason this skill exists): the maker cannot judge their own output objectively. The model that made the image cannot see it. The person who prompted it is biased toward seeing what they intended. "It looks fine to me" and "the file exists, exit code 0" both pass while the hand has six fingers, the Chinese renders as empty tofu boxes, the disclosure label is missing, or the headline sits under the platform UI. So a fresh, context-free reviewer that never saw the generation prompt checks what is actually on the canvas. This is the visual version of the owner's "critical audit: act as if someone else did it."

Two laws sit above everything (from the pipeline): generate wide, gate hard at one place; and files-exist is never QA, watch the pixels. A green codex/exit only means a file got written, not that it looks right.

Every platform number, aspect ratio, safe-zone, and disclosure rule here drifts fast. Re-verify the current platform doc at use time (/last30days or official docs). The thresholds are heuristics, not constants.

Part 1. When it runs

Two fire points, both cheaper than the alternative:

1. On storyboard / keyframe STILLS, before you pay to animate. The single highest-leverage moment. A broken hand or garbled headline costs ~$0.02 to catch on a still and 22 to 270 credits to catch after the video is rendered. Never animate a failing still. (This is the gate vid-ad and the format directors call before motion.)
2. On a finished static ad, after generation, before the human gate. So the human reviews a clean, pre-screened image instead of being the first filter.