# "okay, but I want Gemini3 to perform 10x for my specific use case" - Here is how

Source: https://www.youtube.com/watch?v=UuyaeSLRTkE
Recap page: https://rapidrecap.app/video/UuyaeSLRTkE
Generated: 2025-11-23T11:02:34.987+00:00

---
## Quick Overview

The key to getting Gemini 3 to perform exceptionally well for specific use cases, such as generating high-quality Excalidraw wireframes or complex UI designs, is to structure guidance at the right altitude by identifying the model's convergent default behaviors (like using generic fonts or incorrect JSON schema) and providing explicit, concrete alternatives and rules within the system prompt, rather than relying on overly verbose or complex prompt engineering.

**Key Points:**
- Gemini 3 requires concise prompts and is highly sensitive and steerable, often defaulting to generic aesthetics (like Inter font, purple gradients) if not explicitly guided.
- The video demonstrates fixing Excalidraw JSON output errors by specifying rules in the system prompt, such as requiring explicit width/height for text elements and using correct element types (e.g., "ellipse" instead of "circle").
- The three-step process to improve model output quality involves: 1. Identify convergent defaults, 2. Find the root cause of those defaults, and 3. Structure guidance with concrete alternatives.
- By adding detailed design constraints in the system prompt, the model generates superior outputs, such as a highly stylized music player UI with neon aesthetics, compared to the basic output from a minimal prompt.
- The speaker highlights that for tools like Excalidraw, specific JSON format rules must be enforced in the system prompt, explicitly stating which properties to include/exclude (e.g., excluding metadata like 'seed' or 'versionNonce').
- The concept of 'structuring guidance at the right altitude' means providing domain-specific constraints (like typography rules or interaction patterns) exactly when needed, avoiding permanent context overhead for unrelated tasks.
- The effectiveness of this prompting methodology is demonstrated by achieving significant improvements in UI generation quality across different models and tasks, including web design and finance dashboards.

![Screenshot at 00:54: The three-step process for improving model output quality: 1. Identify convergent defaults, 2. Find the root cause, and 3. Structure guidance with concrete alternatives, is introduced as the core methodology.](https://ss.rapidrecap.app/screens/UuyaeSLRTkE/00-00-54.png)

**Context:** The video explains advanced prompting techniques to improve the output quality of large language models (LLMs) like Gemini 3, specifically focusing on steering the model away from its generic, high-probability training data outputs towards highly specific, creative, and functionally correct results, particularly in the context of front-end design generation using tools like Excalidraw and Superdesign.

## Detailed Analysis

The presenter argues that LLMs, when given minimal guidance, converge on generic, high-probability outputs—a phenomenon called distributional convergence—which undermines brand identity in customer-facing products. To combat this, the video proposes a three-step process for achieving higher quality, domain-specific results: 1. Identify convergent defaults (like using generic fonts or incorrect JSON structures), 2. Find the root cause of that default behavior, and 3. Structure guidance at the right altitude with concrete alternatives. The demonstration focuses heavily on generating Excalidraw wireframes. Initially, a basic prompt yields a generic, poorly aligned wireframe (00:52:57). By adding detailed constraints in the system prompt, such as rules for text elements (requiring explicit width/height and using textAlign) and eliminating undesirable metadata fields (like 'seed' or 'versionNonce'), the model produces significantly better, more stylized, and functionally correct JSON output for Excalidraw (00:59:54). The video also shows a successful application of this method to Claude using its 'frontend-design' skill, resulting in high-fidelity UI mockups for a music player (01:47:18) and a MetaMask-style crypto wallet (12:03:17), proving that structured, high-altitude guidance, rather than low-altitude hardcoded logic, maximizes model performance.

### Gemini 3 Prompting Differences

- Gemini 3 requires concise prompts and is less verbose than older models
- It may over-analyze verbose or complex prompt engineering techniques
- Output verbosity is lower by default, preferring direct, efficient answers

### The Three-Step Improvement Process

- 1. Identify convergent defaults
- 2. Find the root cause
- 3. Structure guidance with concrete alternatives
- The goal is to articulate reasoning behind desired behavior to avoid default issues

### Excalidraw JSON Schema Fixes

- Model defaults to width: 0 for text elements causing invisibility/layout issues
- Model uses incorrect element types (e.g., "circle" instead of "ellipse")
- Model includes unnecessary metadata fields (seed, versionNonce) that should be excluded

### Structuring Guidance

- Guidance should specify properties for each element type (e.g., for 'text' include 'colorStrike', 'width', 'heights')
- Crucially, NEVER output properties that only impact styling/metadata like 'seed', 'version', or 'isDeleted'

### Impact on Design Generation

- Applying these principles leads to much higher quality outputs across domains, like generating a neon-themed music player UI (06:00:00) or a Brutalist-style MetaMask UI (12:03:00), moving beyond generic LLM aesthetics.

![Screenshot at 00:04: The initial screen showing the Gemini 3 introduction on the Google DeepMind website.](https://ss.rapidrecap.app/screens/UuyaeSLRTkE/00-00-04.png)
![Screenshot at 00:07: A demonstration of an early prompt resulting in a basic music player wireframe resembling a Sony Walkman.](https://ss.rapidrecap.app/screens/UuyaeSLRTkE/00-00-07.png)
![Screenshot at 00:25: The Google AI Studio playground interface where the user is testing Gemini 3 Pro with system instructions and a prompt to build a 'Hello World' page.](https://ss.rapidrecap.app/screens/UuyaeSLRTkE/00-00-25.png)
![Screenshot at 01:48: An example of Claude's high-quality output using the 'frontend-design' skill to create a detailed vinyl player UI.](https://ss.rapidrecap.app/screens/UuyaeSLRTkE/00-01-48.png)
![Screenshot at 03:44: A slide illustrating the concept of prompt specificity on a scale from 'Too specific' \(red\) to 'Just right' \(green\), emphasizing providing contextually relevant guidance.](https://ss.rapidrecap.app/screens/UuyaeSLRTkE/00-03-44.png)