The CLI workflow is straightforward:
matraix init --project my-onboarding-flow
matraix persona add --name "skeptical_dev" --traits "technical,impatient,detail-oriented"
matraix persona add --name "casual_mobile" --traits "non-technical,distracted,short-sessions"
matraix run --prompt-file prompts/v3_onboarding.yaml --personas all --iterations 50
Output includes per-persona completion rates, drop-off points, sentiment drift, and a divergence score showing where your prompt fails for specific archetypes. The web dashboard visualizes conversation trees so you can spot exactly where "skeptical_dev" bails versus "casual_mobile."
What's useful: the personas aren't just prompt prefixes — they maintain memory across turns, simulate typing delays, and can be seeded with prior context (e.g., "user saw pricing page but didn't convert"). You can also inject adversarial behaviors: hallucination triggers, prompt injection attempts, multi-language code-switching.
Still rough edges: no native CI/CD integration yet (GitHub Action is on the roadmap), and the persona library is small — you'll spend time crafting your own. Export format is JSONL only; would love Parquet for heavier analysis.
If you're iterating on LLM-driven onboarding, support bots, or survey logic, this beats manual testing by a mile. Free tier covers 500 simulations/month.
Next CISA just dropped an advisory about AI-generated malware →
an AI side-hustle playbook, with plenty of directly applicable cases.