Prompt Complexity Quantification

A vibe-research project exploring how constraint density in LLM prompts affects output complexity, readability, and model behavior.

This repository was created with Sahil Ohe using the Gemini CLI.

Overview

We study the relationship between prompt specificity and LLM response quality using quantitative metrics such as the Automated Readability Index (ARI) and lexical density. The work introduces a Three-Zone model:

1. The Parrot Zone — underspecified prompts yield clichéd, low-complexity outputs
2. The Sweet Spot — optimal constraint density produces sophisticated synthesis
3. Stochastic Collapse — extreme or contradictory constraints trigger refusal or logical auditing

Contents

File	Description
paper.md	Research write-up (markdown)
paper.html	Rendered version of the paper
wiki.html	Supplementary wiki-style notes
analyze_results.py	Python script to compute ARI and related metrics
concept.mmd	Mermaid diagram of the constraint-complexity model

Setup

python3 -m venv venv
source venv/bin/activate
pip install textstat numpy

Usage

Run the analysis script against the sample model outputs:

python analyze_results.py

Authors

Gemini CLI & Sahil Ohe — May 2026