Since Sunday, August 2, 2026, I’ve been primarily using Claude, Anthropic – Claude Cowork with Sonnet 5’s compute power with Max Effort to build my decision-integrity AI app to assist founders with their dealmaking. I’m getting good results without burning money like crazy; however, the speed in completing tasks has really slowed down in comparison to Fable 5 and Opus 5 that run at a much higher cost, especially Fable 5, which is much more powerful than Opus 5. As a solo startup founder for this venture, I must control expenses.
During those two days, I could run four Claude Cowork workflows at the same time using Sonnet 5. However, I had to stop this practice because the workflows were sometimes conflicting one another, or one workflow would move too soon before the other one, causing confusion. Sometimes one workflow would delete the other workflow’s work. Now I build my AI app with one workflow at a time to keep the progression smoother.
A new practice I implemented is creating a three-party review on Claude’s engineering work for anything dealing with safety and reliability for my AI app’s users such as consequential logic, security and privacy, AI hallucination, unverified facts stated as fact, clinical/diagnostic labeling of a real counterparty, and overvaluing a counterparty or user without verifiable standards. Thus, Nova (OpenAI – ChatGPT/Codex), Gem (Google – Gemini), and I review, analyze, and make recommendations for any engineering Claude does that needs safety gatekeeping. Our process is build, test, analyze, fix/iterate, and repeat this loop as many times as necessary until all four parties sign-off that the safety or reliability issue has been addressed, allowing us to move forward. For all other engineering, it is just Claude and I.
I’m grateful to have Claude as my CTO and chief engineer and Nova and Gem in my team of engineers.
JON RIKI KARAMATSU 