gradient-screenshot-repair

name	gradient-screenshot-repair
description	Repair native-looking mobile screenshot backgrounds by erasing localized UI content and reconstructing nearby gradients from sampled pixels. Use when Codex must remove status-bar items, avatars, nicknames, labels, chips, or overlay text from app screenshots without visible seams, flat fills, or color mismatches. Also use for Chinese requests like 擦除顶部、补渐变、抹掉头像、抹掉昵称、保留原图底色、不要白块、不要假遮罩.

Gradient Screenshot Repair

Overview

Use this skill to clean a bounded region inside a mobile screenshot while keeping the original page surface believable. Prefer rebuilding the background from nearby sampled pixels instead of drawing flat color blocks, blur overlays, or fake replacement panels.

Workflow

1. Identify the exact region to remove. Use a slightly larger target rect than the visible dirty pixels so edge remnants are also covered.

2. Choose clean sample bands. Use these three samples whenever the damaged area sits on a mixed horizontal and vertical gradient:

• top-sample: a clean horizontal band just above or just below the target area; this captures left-to-right color drift
• right-top-sample: a clean block near the target's upper-right edge; this captures the upper vertical tone
• right-bottom-sample: a clean block near the target's lower-right edge; this captures the lower vertical tone

3. Build the patch with sampled gradients. Run scripts/sample_surface_patch.py to reconstruct the target area. The script combines a horizontal palette from the top band with a vertical delta from the clean side edge, anchored at the top-right reference color so the two gradients do not double-count.

4. Restore preserved foreground elements if needed. If a badge, chip, or icon should remain, paste that untouched source rect back after the patch is applied. This is often the cleanest way to hide a seam around small preserved UI.

5. Verify at native scale. Inspect the edge nearest the clean source samples first. If one side still shows a seam, move the sample rect closer to the seam before trying manual tint adjustments.

Rules

• Do not use flat fills unless the source area is actually flat.
• Do not add white masks, translucent overlays, or new background cards.
• Do not sample from already-damaged pixels.
• Do not trust a single point sample when the source surface is a gradient.
• Prefer replacing pixels in the reference image or asset itself, not layering a fake background above it at runtime.

Tuning

• If the repaired area is too light or too dark on one side, move the side samples closer to that side's seam.
• If left and right gradient direction looks reversed, the horizontal sample order is wrong or the wrong clean band was used.
• If the patch looks flat, widen the top sample or increase its height slightly.
• If a thin seam remains, expand the target rect by a few pixels past the visible artifact.
• If only vertical gradient matters, keep the same workflow but choose a top sample with minimal horizontal variation.

Resources

scripts/

• scripts/sample_surface_patch.py: paste a reconstructed patch into an image using one target rect, one horizontal sample band, two side samples, and optional restore rects.

Usage:

python3 scripts/sample_surface_patch.py \
  --input screenshot.png \
  --output repaired.png \
  --scale 3 \
  --target 14,76,190,74 \
  --top-sample 14,60,178,16 \
  --right-top-sample 220,82,40,18 \
  --right-bottom-sample 250,136,40,14 \
  --restore 192,110,74,24

references/

• references/gradient-repair-playbook.md: parameter selection rules, failure symptoms, and the proven OpenAiPay profile-header example.