Rebellis.ai
2024
Ai Recommendation system
Designing an AI Recommendation System to Improve Prompt Quality
Overview
Rebellis AI enables users to generate 3D character animations from text prompts. Research showed that many first-time users struggled to write prompts that the AI could interpret effectively. Short, vague inputs often resulted in disappointing animations, not because of the model itself, but because they lacked the descriptive context needed for high-quality generation. To address this, I designed an AI recommendation system that guided users toward writing richer, more structured prompts, improving both prompt quality and the overall animation results.
Problem
Why the simple Prompt Wasn't Working
In a generative AI product, the prompt is the entire interface, it's the only lever users have to tell the model what they want. Rebellis AI's prompt field, though, was a single empty text box with no support around it.
Users across the board were consistently blaming the model for poor animation quality.
Looking closer, through Microsoft Clarity recordings, Discord conversations, community feedback, live workshops, and a direct review of real prompts, the actual issue came into focus: the model wasn't the problem. People simply didn't know how to write a prompt it could work with.
This led to three main challenges:
Friction in Prompt Quality: Users routinely typed inputs as sparse as “walking,” with no style, pace, or emotional detail, and had no way to know what a stronger prompt looked like before hitting generate.
Missed Teaching Opportunity: The prompt field was the single most-used surface in the product, yet it offered no examples, defaults, or contextual hints, every user was expected to already know prompt engineering.
Lack of a Scalable Pattern: We needed one system that worked equally well for a one-word prompt and a ten-word prompt, without leaning on a one-time onboarding tutorial that most people skip or forget.
Research
Exploring Solutions: Prompt Research and Pattern Analysis
o get past anecdotes, we combined behavioral and qualitative research, Clarity session recordings, Discord threads, workshops, and support conversations, with a direct audit of real usage: a sample of 93 user-submitted prompts, reviewed line by line.
15% of prompts (14 of 93) were only one or two words, too short and ambiguous for the model to interpret.
9% of prompts (8 of 93) were resubmitted multiple times in the same session, a sign the attempt hadn't landed.
52% of prompts (48 of 93) were six words or longer, averaging 6.8 words, and without meaning and with our detail and description.
We also grouped prompts by intent to see what people were actually trying to create. Walk/run motions formed the largest identifiable category at 26%, followed by fight/weapon (12%), pose/idle (9%), dance/performance (6%), and emotion/interaction (6%). The single biggest group, though, 41% of the sample, was too vague to categorize at all, the clearest signal that the input problem cut across every use case, not just one.
This research pointed to two testable directions:
Contextual Recommendations: Surface stronger prompt alternatives directly beneath the input, tailored to what the user already typed.
Progressive Validation: Check every submitted prompt for missing context , style, pace, emotion , before it ever reaches the model.
Ideation
Designing the Recommendation System
When building the high-fidelity flow, my main goal was to make good prompting the path of least resistance for every user, not an extra step reserved for onboarding.
The product now surfaces an AI Prompt Recommendation panel directly beneath the input field.
For a prompt like “a person walk,” it suggests fuller alternatives — such as “a person walk slowly and move the hands” — each paired with a short preview and a style filter (e.g., Retro Style), so a user can adopt a stronger prompt in one tap instead of guessing what to add. Rather than judging a prompt only after generation, the system checks for missing context the moment a prompt is submitted, returns suggested rewrites, and lets the user accept one as-is or edit it further before it moves on to the model.
Beyond the input field itself, we layered in lightweight prompting challenges, short tutorials, and real community examples, so writing better prompts became part of using the product for everyone — not a separate skill new users had to learn before they could get good results.
Designs
By turning a blank field into a guided, data-informed input, we improved generation quality without touching the model itself. To measure the redesign, we looked at prompt quality, generation success, and feature adoption:
Stronger Prompt Quality: Prompts across the platform grew more descriptive, with fewer one- and two-word submissions than in our original 93-prompt audit.
Fewer Retries: With missing context caught before generation, fewer prompts needed a second or third attempt to produce a usable result. More Exploration: The recommendation panel gave users a low-effort way to explore styles they hadn't typed themselves, instead of relying purely on trial and error.
Platform Growth: The product grew to 3.1K active users, and the recommendation logic became the foundation for later onboarding and discovery features
