Screenshot-annotated web edits tested with Ling-3.0: only 2 of 3 runs passed
alifcoder · x · 2026-09-24
The author tested Ant Group's Ling-3.0-flash-VL on editing an event signup page from annotated screenshots — red circles, arrows and boxes marking the changes, passed along with the source code. Across three runs of the same task, the first version misplaced the button every time; after feeding back screenshots of the actual rendered page, two runs passed all 17 acceptance checks in one round, for a final 2/3 success rate.
The checklist covered text preservation, dialogs, FAQ expansion, mobile overflow and console errors, not just layout. Switching the target to a different card still failed, so full-page edits can't yet be left unattended. Reusable workflow: annotate the screenshot, feed back rendered pages when the first pass fails, and explicitly state what must be preserved.
More from coding & agent
- Hugging Face Agent Traces now show tokens, cache hit rate and cost per run — victormustar · 2026-09-24
- Building a From-Scratch Computer That Runs a 107K-Parameter Language Model — tetsuoai · 2026-09-24
- Indie Dev Shows How Claude Opus 5.5 Built Pressure Wash Panic: Tripo, Blender and Water Physics — chongdashu · 2026-09-24
- One prompt, full go-to-market plan: a real one-shot demo on self-hosted agent Octo — SarahAnnabels · 2026-09-24
- Beyond the "hi" chat bubble: what happens when you hand an agent real work — SarahAnnabels · 2026-09-24
- Even doctors are 'Claude Maxing' now: AI-built sites become recognizable — gaganghotra_ · 2026-09-24