# AntiSlop

- **Event:** [Gemini 4 NYC Hackathon](https://cerebralvalley.ai/e/gemini-4-nyc)
- **When:** Sat, Nov 21 at 9:00 AM – 10:00 PM (EST)
- **Where:** New York, NY
- **Team:** [Shantanu Patil](https://cerebralvalley.ai/u/Shantanu_Patil)
- **GitHub:** https://github.com/REDFOX1899/antislop
- **Demo video:** https://youtu.be/BJo5pevn49w
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/78

You can often guess which AI built a website by looking at its gradients. Coding agents are good at implementing requirements, but visual taste is personal. Once a page is on screen, the easiest way to explain what feels off is to point at it.

AntiSlop lets you edit the website you’re looking at. Circle a section, say what you want changed, and show a reference image or motion clip. Gemini edits the underlying code, then AntiSlop renders the result on desktop and mobile and uses the Decisions API to catch problems and trigger targeted fixes.

In eight controlled trials, AntiSlop passed 5/5 visual evaluations. The text-only baseline passed 0/3. Try it at https://ui-ux.fun.                       


How it works:
1. Raw Multimodal Ingestion: Drop moodboards, capture any live production URL via headless Chromium (screenshot + computed CSS DNA in ~2.4s), hold up a physical magazine/object via webcam, upload motion videos, draw freehand vermilion circles directly onto a running preview, or hold a two-way real-time voice conversation with Gemini 3.8 Live (gemini-3.8-live) over WebSockets with barge-in interruption and canvas vision.
2. Multi-Model Taste Synthesis & Domain-Grounded Code Surgery: gemini-4-carbon (or argon in fast mode) distills a typed DesignManifest guided by our open-source Design Craft Skill (skills/antislop-design-craft/SKILL.md). Stack-aware generators (gemini-4-hydrogen & Best-of-N parallel gemini-4-neon) rewrite React/Tailwind or vanilla HTML/CSS/JS shader code with live HMR, while gemini-nano-banana-2.1 generates domain-grounded bespoke hero artwork.
3. 900-Probe Calibrated Decisions API & Breaking Points (Element 119 + 5 Sigma): Instead of fragile LLM-as-a-judge text, gemini-decision-1 runs zero-output-token calibrated probes (~250ms 12-probe critic gate + 900-probe single-pass Spatial X-Ray in ~1.6-2.3s; dual-key sharded to 1,800 probes to break the 1,000-probe single-call API limit) fused with deterministic Playwright DOM audits. Empirically on unseen inputs: Taste Index improves from 13.8 -> 99.5/100 (0/3 vs 3/3 pass) and Slop drops 9.09 -> 1.41.
4. Remote Ship via Managed Agents: Dispatches a repo-isolated Antigravity Managed Agent mounting our inline Skill into Google's remote Linux VM, enforcing npm run build exit-0 verification before opening a real GitHub PR.

## More from Gemini 4 NYC Hackathon

- [Team Carthij](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/75)
- [Team SceneTrace](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/76)
- [Sync](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/77)
- [Adaptive NPCs](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/79)
- [Sheetsight](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/80)
- [Chifles](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/81)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/78. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
