Skip to content
1st Place

DefaultTaste

Agent Taste Profiling as a Service!!

Code Video Cerebral Valley

Core trick

Samples an agent hundreds of times across prompts and clusters recurring colors, fonts, and layouts into a profile that yields a correction prompt.

Tags

dev toolscreative/music

Description

Agent Taste Profiling as a Service!! - It works by probing an agent repeatedly across many generations, then measuring the patterns it keeps returning to, such as colors, fonts, layouts, genres, tempo, and mood. - The result is a clear **taste profile** plus a **correction prompt** that helps users override those defaults. - The problem it solves is that most people judge AI from a single output, which hides the model’s underlying bias and stylistic tendencies. - In practice, this means creators often ship outputs shaped more by the model’s default taste than their own intent. - DefaultTaste makes those defaults visible and actionable, so developers, designers, and creators can steer AI outputs toward something more original, intentional, and differentiated.
ilhamfp/DefaultTaste★ 52 forksHTMLMITCreated Mar 7, 2026 · last push Mar 8, 2026

README excerpt

<p align="center">
  <img src="img.png" alt="DefaultTaste — Point us at an agent. Reveal the taste it defaults to." width="100%" />
</p>

<p align="center">
  DefaultTaste reveals the hidden aesthetic preferences baked into AI agents.<br/>
  We probe a model hundreds of times, map the patterns it defaults to,<br/>
  and generate a correction prompt to break out of them.
</p>

---

## The Problem

Ask an AI to "make me a website" and you'll get React, Tailwind, Inter font, purple gradients, and three feature cards. Ask it to compose music and you'll get 120 BPM pop in the key of C. Do it again — same thing. And again.

These aren't creative choices. They're **defaults** — aesthetic biases absorbed from training data that silently shape every output. A [March 2025 study](https://arxiv.org/abs/2503.01633) found LLMs pick Python 90–97% of the time. A [December 2025 paper](https://arxiv.org/abs/2412.13153) showed image generators converge to just 12 visual motifs.

Every AI model has taste. DefaultTaste makes it visible.

## How It Works

```
1. PROBE      Send the same vague prompt hundreds of times
2. ANALYZE    Extract dimensions: colors, fonts, frameworks, BPM, key, genre...
3. PROFILE    Aggregate into a taste fingerprint with frequency distributions
4. CORRECT    Generate a negation prompt that inverts the defaults
```

The result is a **taste profile** — a data-driven portrait of what an AI reaches for when you don't tell it what to do — paired with a **correction prompt**

Team