A conversational task planner I built to learn how you build a probabilistic product.
Tudu is a conversational planner. You give it a goal in plain language, and it breaks the goal into weekly and daily tasks so you make real progress toward it. I have designed and built several products end to end, but never a probabilistic one. I wanted to learn how you make one behave reliably.
Learning to evaluate
Getting a probabilistic system to perform reliably comes down to evaluation. I stood up a throwaway chat interface and ran the planner through a large number of real conversations to generate traces. I read through them, did error analysis from the bottom up, and identified failure modes from actual traces. From that taxonomy I built an LLM-as-judge, binary pass or fail with a written critique, following the method Hamel Husain and Shreya Shankar teach, and used it to score changes to the prompts and the planning logic.
A design language from scratch
I also built the product's design language in Figma from the ground up. A token set for color, type and spacing, the focus-area model, and the full conversational planning flow, all wired into the component layer of the app. I had not designed anything since Let's, the last company I built, and I wanted to reacquaint myself with the craft and explore Figma.
Tudu was always a way to learn. Once it had taught me what I came for, how to build a probabilistic system that performs, I set it aside.
You describe a goal in chat, and it comes back as a multi-week plan, broken into daily tasks you refine and approve before they land.
Today’s tasks gathered into a single list, each tagged with the focus area it belongs to.
Focus areas are the parts of life you choose to prioritize. Tudu fits every task into one of them, and marks anything that matches none as Unfocused, so you can see what is falling outside the areas you set.
You tell Tudu about yourself, how you like tasks and plans structured, and any standing instructions, and it plans around them.
