Same Claude, same screenshot: blind vs checking itself. Here's the prompt and the script.

1 The prompt (the self-check arm)

Work in a loop: write index.html, then run ./compare.sh (it renders your page to shot.png, writes a red diff.png and prints the percent of pixels that differ). Read shot.png and diff.png with the Read tool, fix the biggest differences, and run ./compare.sh again. Stop when the mismatch is under 3 percent or after 8 rounds, whichever comes first.

The blind arm got the same first paragraph and one different instruction: write index.html once, do not render or check it.

2 compare.sh

#!/bin/sh
cd "$(dirname "$0")" && python3 compare.py

Save it as compare.sh next to compare.py and make it executable (chmod +x compare.sh).

3 compare.py

import sys, os, json, time, shutil
from playwright.sync_api import sync_playwright
from PIL import Image, ImageChops
import numpy as np
here=os.path.dirname(os.path.abspath(__file__))
exe=os.path.expanduser('~/Library/Caches/ms-playwright/chromium_headless_shell-1248/chrome-headless-shell-mac-arm64/chrome-headless-shell')
with sync_playwright() as p:
    b=p.chromium.launch(executable_path=exe)
    pg=b.new_context(viewport={'width':430,'height':int(Image.open(os.path.join(here,'target.png')).height)},device_scale_factor=1).new_page()
    pg.goto('file://'+os.path.join(here,'index.html'),wait_until='load'); pg.wait_for_timeout(500)
    pg.screenshot(path=os.path.join(here,'shot.png'),full_page=False); b.close()
t=Image.open(os.path.join(here,'target.png')).convert('RGB'); s=Image.open(os.path.join(here,'shot.png')).convert('RGB')
a=np.asarray(t).astype(int); c=np.asarray(s).astype(int)
d=np.abs(a-c).max(axis=2); bad=d>24
pct=100*bad.mean()
vis=np.asarray(s).copy()//3+170
vis[bad]=[255,40,40]
Image.fromarray(vis.astype('uint8')).save(os.path.join(here,'diff.png'))
rd=os.path.join(here,'rounds'); os.makedirs(rd,exist_ok=True)
n=len([f for f in os.listdir(rd) if f.endswith('-shot.png')])+1
shutil.copy(os.path.join(here,'shot.png'),os.path.join(rd,f'{n:02d}-shot.png'))
shutil.copy(os.path.join(here,'diff.png'),os.path.join(rd,f'{n:02d}-diff.png'))
open(os.path.join(here,'log.jsonl'),'a').write(json.dumps({'round':n,'t':time.time(),'mismatch_pct':round(pct,2)})+'\n')
print(f'round {n}: {pct:.2f}% of pixels differ (red in diff.png). Look at shot.png and diff.png.')

It renders your index.html at the screenshot's width and height, counts the pixels where any colour channel differs by more than 24 out of 255, and paints them red in diff.png. Change the Chromium path to your own install.

4 What we measured

Tested 24 Oct 2026 with Claude Code 2.1.309 on Sonnet 5.5, on a 430 x 1500 mobile screenshot of the Wikipedia article "Voxel" (text and images CC BY-SA 4.0). Blind: 17.07% and 12.73% of pixels differed (two runs, 18 s, about 7 cents API-equivalent). Self-check: 2.52% in 6 rounds and 64 s for 18 cents, and 3.51% in 5 rounds for 14 cents. The blind run put the Wikipedia wordmark in the first picture slot because it never opened the image files.

Fair to say: results vary by page. On the Wikipedia "Claude" article and python.org the loop helped about 1.5 to 2 times; on Hacker News it did not help at all (fonts and row heights it could not match). Not pixel perfect: 2.5% of pixels still differ.