Self-check before showing a render
Render to PNG, view the whole image, then zoomed crops (scripts/image_tools.py sheet for side-by-side candidates).
- Offer is the first thing you see; script headline does not touch it (≥ 16 px gap).
- No text collides with products, particles or edges (check descenders, swashes, diacritics).
- No large empty zone; nothing crammed against the canvas edge (≥ 50 px margin for text).
- Every logo on a product: centred, fully visible, looks printed (curved, lit, not a flat sticker).
- One light direction everywhere; shadows agree; back items darker than front.
- Colours from the brand palette only; background fresh/light unless the brand is dark.
- Every text line passes
image_tools.py contrast(4.5:1; 3:1 only from 65 px, or 50 px bold); ≤ 3 text colours. - No visible seams where an image meets vector background (zoom the boundary).
- Fonts rendered (no fallback warning); every font passes
image_tools.py fontcheck; ≤ 2 families; stacked marks (Ặ Ầ) clear of the line above. - Placeholder copy listed in the reply; asset sources/licences noted; credits spent reported.
- After any structural change (zoom/crop an image, add/remove a frame or clip): look at an app screenshot
(
ui_screenshot/render_previewin the GUI), not only the export — artboard-edge lines show only there. - Zoom 2× on every seam between an edited region and the untouched original (tone shift, cut line, a sagging thin line left half-fixed next to a kept box).
- Empty space measured, not guessed:
image_tools.py emptiness preview.png— every flagged cell is filled or justified in the reply. - A fix changes only the reported defect, evenly along its whole length; no fade, gradient or new effect the client did not ask for.
- Only one working file; preview rendered to a temp folder; no export unless approved or requested.
Source: skills/design-ai/references/checklist.md on GitHub ↗