Show HN: A Genetic algorithm that red-teams your copy with 100 LLM personas
TL;DR
An indie developer built a system that scores and iteratively improves landing page copy using over 100 distinct AI personas.
Key Points
- The approach draws inspiration from Google DeepMind's AlphaEvolve, applying evolutionary algorithm principles to copywriting.
- Each persona carries its own expertise, personality, and aesthetic grounded in real-world data – bypassing the generic single-model output problem.
- A genetic algorithm selects top text variants and recombines them, replacing the typical one-shot prompt workflow.
- The tool is live as CrashTestCopy, launched on HN with the developer openly acknowledging the product-pitch risk.
Nauti's Take
The technical approach is more substantive than the usual 'AI writes your website' noise: evolutionary selection across diverse personas is conceptually sound and mirrors ensemble methods from classical machine learning. Whether 100 personas built on 'real-world data' actually produce genuine diversity or just perform diversity theater is the critical open question.
The AlphaEvolve reference is well-chosen, though worth noting that AlphaEvolve optimizes against verifiable correctness while 'good copy' remains a fuzzy target. Still – anyone who has ever tried to coax ChatGPT into genuinely original marketing text immediately understands the exact problem being solved here.
Sources
