I needed a way to refine my AI prompts. So I built one.
Pearl Prompt Lab is a multi-persona simulation engine for testing and refining conversational AI.
Multi-Persona Testing
Five simulated personalities. One prompt.
Run your prompt against a Talkative storyteller, a Guarded skeptic, a Tech-Anxious novice, and more. See how your AI handles the full spectrum of real human behavior before it ever talks to a real person.
PEARL PROMPT LAB
TEST PERSONA
SeniorsMiddle-Aged
Talkative Storyteller
Dorothy, 82 - Loves to talk, goes on tangents, very warm
Reserved / Needs Coaxing
Harold, 78 - Short answers, opens up with patience
Auto-Refine
Automated iteration until convergence.
Set it running and walk away. Pearl auto-refines your prompt through successive iterations, detecting when scores plateau and stopping when further changes would just be noise.
AUTO-REFINE
Iteration 7 of 10Converged
8.4
Score
0.12
Variance
Evaluation Engine
Eight dimensions. One weighted score.
Every response is evaluated across emotional awareness, naturalness, conversation quality, story extraction, format compliance, and more. Weighted scoring surfaces exactly where your prompt excels and where it falls short.
SCORE BREAKDOWN
EA
8.7
NAT
9.2
CONV
7.8
STORY
8.4
FMT
9.5
MEM
6.3
Pairwise Comparison
A/B test your prompts with data, not gut feel.
Run two prompt versions against the same persona and see which one performs better across every dimension. No more guessing whether your edit actually improved things.
Prompt A
7.6
Overall
vs
Prompt B
8.4
Overall
+0.8
Guardrails
Refine without drifting.
Define a constitution of non-negotiable behaviors. Pearl ensures your prompt gets better without losing what already works. Optimization with guardrails, not optimization at any cost.
CONSTITUTION
✓Always acknowledge emotions before asking follow-up questions
✓Never suggest the user's feelings are invalid
✓Maintain warm, conversational tone throughout
How it works
1
Configure your prompt
Paste your system prompt, choose a persona to test against, and set your constitution of non-negotiable behaviors.
2
Run simulated conversations
Pearl simulates a full conversation with your AI, playing the role of different personality types and evaluating every response.
3
See what works
Get a detailed score breakdown, identify weak spots, and let auto-refine iterate your prompt until it converges on the best version.
Built for builders.
Pearl Prompt Lab is an ongoing experiment by Takisaki Strategy. Follow along as we build in public.