I built a personal AI agent in WhatsApp to learn how agents work. The mistakes taught more than the build.
If you want to learn AI, start with something you care about whether it works. Not a lesson. A project.
I wanted to understand personal agents like Meta's Muse, which launched September 8. So I built a light-touch personal trainer in WhatsApp. Every hour from 8am to 8pm it sends one bodyweight movement I can do in two minutes. I reply "done" or "skip." Done earns a new movement. Skip or silence repeats it.
The tech is simple: scheduled messages, a chat thread, a setup file, and a small state record. The learning came from real use.
The first movement arrived and I did not recognize it. I fixed that by adding a "pic" command that sends an annotated graphic plus written steps. Muse confirmed the change. At the next movement I replied "pic," and WhatsApp asked what kind of picture I wanted.
That failure was the lesson. Muse confirming a change proved it understood me. It did not prove the live jobs had changed. Only a test where the feature actually runs could prove that. One "sync trainer" command from WhatsApp fixed it and gave me a repeatable path for every change since: design it, write it to the setup file, sync, verify.
A week in, every feature came from the same loop. Ask for an outcome. Observe what happens. Record it. Adapt the next instruction. Report progress. Repeat until it creates value. F.A.I.L. means First Attempt In Learning. The first attempt is an instrument, not a verdict.
Strip out the fitness and the pattern is formative feedback. Ask a student to attempt something, observe, record, adapt, report. The agent does not change the pedagogy. It shortens the distance between the steps. The trainer is the use case. The loop is the asset.
Two documents carry the full story and the build. Read both.
The field guide: the week of iteration, what broke, why, and what transfers to the classroom.
The build guide: copy-and-paste prompts, checkpoints, and the deployment path, with every week-one fix already installed.
Before you start: Muse is US only and 18+ and this project consumes tokens on hourly jobs, but you get 100 million tokens a week and my week of building and using only cost me 7% of my weeks quota. Keep district accounts and student data out of it. Both guides are current as of September 2026 and will drift.
Build the smallest complete loop. Put it in your life. Let real use edit it.
Below are two PDF files, one is the full story of the experience and the other is the Step-by-Step how to Guide.
Project Overview
View or download
Step by Step Guide
View or download