> ## Content Index > Fetch the complete content index at: https://intelligence.artificialrealitymedia.com/llms.txt > Use this file to discover other available public pages before exploring further. # How Dr. Nikki Siso tests her own AI without a developer - URL: https://intelligence.artificialrealitymedia.com/how-dr-nikki-siso-tests-her-own-ai/ - Published: 2026-09-02T12:10:00.000Z - Updated: 2026-09-02T12:10:00.000Z - Description: She tells Claude to pretend to be five very difficult clients and walk the whole flow. The bugs turn up. Here is the method, step by step. - Author: Galina Fendikevich - Tags: Health & wellness, #For AI Buyers, #What's working, #Import 2026-09-03 05:20 Dr. Nikki Siso, a certified holistic health practitioner with no technical training, tests her own AI by telling [Claude](https://claude.ai/?ref=intelligence.artificialrealitymedia.com) to role-play five very difficult clients and walk the entire flow. It found bugs in her fifteen-skill assessment that she had previously been passing to a hired developer one at a time. The method costs nothing and needs no tooling. ## What exactly does she tell the model? Two prompts, one simple and one adversarial. The simple version, in her words: > "You can tell Claude: hey, pretend like you're a user and go through and test this chatbot. And on the back end, it will go through and do a whole run and, like — yeah, got stuck here. Let me fix the code. Does great job." The version she offers as a tip is the one worth copying: > "Pretend like you're five very difficult clients. Don't answer this fruitfully. Answer almost like not wanting to answer, or give half answers — just be difficult. And it did. It ran through five different types of clients that could come with all sorts of different health conditions, and we found the bugs." The instruction is doing two things at once: covering the range of people who actually arrive, and covering the way they actually behave when they arrive. ## What was breaking? Failures at the end of long sessions, which is the most expensive place for a conversational product to fail. Siso's assessment, TOX, is a total toxic load assessment built on Claude, and a client can spend forty-five minutes working through it. > "It would stop midway and just stop, and we're like: what happened." > "Apparently we ran out of tokens." The worst version came at the final step, when the system was attempting to deliver a ten-page report, a practitioner briefing and structured data in a single response. > "It had three instructions at the end and then it would just freeze. And then it was freaky, because we thought: this person just took 45 minutes to go through this full assessment, and where's the report?" ## Why did finding her own bugs matter so much? Because it removed a dependency. Before, every bug she found became a ticket for someone else; after, most of them became something she could fix the same day. > "Once I figured that out ... I was doing it on my end, because otherwise I was finding a bug, but my developer had to fix it. And now I could do it myself." That is the practical case for the technique for any operator running a system they specified but did not code. The bottleneck in a build like this is rarely the fixing. It is the waiting. ## What could she still not fix? Anything in the orchestration layer between agents. Her fifteen skills were hers; the wiring between them was not. > "I couldn't fix what he set up. I couldn't fix if the conductor and the health assessor weren't talking together. That was on the back end. I could only fix the skills." Worth knowing before you assume self-testing removes the developer entirely. It does not. It changes what you need the developer for. ## How do you run this on your own system this week? Six steps, the first four from Siso's method and the last two added by Artificial Reality: - **Instruct the model to be a user, not to review the design.** Ask it to walk the flow as a person, not to critique the prompt. - **Make the personas uncooperative.** Half answers, reluctance, evasion. Cooperative test users find almost nothing. - **Cover the range of people who arrive.** Siso ran five client types with different conditions, not five variations of the same person. - **Run the complete flow to the final output.** Her failures were at the end, after the long part, which is exactly where a spot check does not look. - **Re-run it after every change.** A change to one skill can break a path nobody re-walks. - **Judge tone, not just completion.** For a coaching product, an answer that arrives cold at the end is a failure even though the system technically finished. Spend an hour on this before you evaluate any testing platform. If an hour of adversarial role-play finds nothing, you have learned something worth knowing. If it finds three bugs, as it did here, you have your answer about what to do next. ## Where this comes from S1E7: How a non-tech founder built a 15-agent AI business on Claude — the full interview with Dr. Nikki Siso. Listen or watch: [YouTube](https://www.youtube.com/watch?v=R%5FzWAgAEIMw&ref=intelligence.artificialrealitymedia.com), [Spotify](https://open.spotify.com/episode/08s9u7LSBTtGGx94b4z2MZ?ref=intelligence.artificialrealitymedia.com) or [Apple Podcasts](https://podcasts.apple.com/us/podcast/how-a-non-tech-founder-built-a-15-agent-ai/id6780699025?i=1000777283240&ref=intelligence.artificialrealitymedia.com).