After the web side of the smart-water platform went live, the field engineers raised a very practical problem: they are out on site and cannot sit in front of a computer all day. That is how the companion mini program came about.
From signed-off requirements to a working release, it took about 3 weeks — roughly 60% shorter than a typical outsourced build.
Decide what not to build first
I went through everything the field scenario seemed to need, and kept only five modules:
- Overview: how many alerts today, how many devices online
- Live monitoring: current water level and device status
- History: look back over a period of data
- Alerts: view and handle alerts
- Settings
Maps and reports were cut on purpose. They are not useless — they are just not what you reach for first, on a phone, standing on site. The web version already has them, and the mini program does not need to repeat it.
Where AI really saved time
On this project I used Claude for the whole journey from product design to release. The places where it felt fastest were these:
- Breaking down requirements: turning "I want to see alerts on my phone" into pages, fields and states
- Prototyping: producing high-fidelity pages you can click, instead of static wireframes
- Revisions: changing one field no longer means redrawing three screens
Once the prototype was realistic enough, reviews were about "is this flow right" rather than "where does this button go", and rework dropped naturally.
What was left for later
The MVP ran the full flow on simulated data first. The roadmap from there:
- Connect real sensor APIs
- Add a map view of station locations
- Push alerts through WeChat messages
What I took from it
AI made "building it" cheaper, but it did not make "what to build" any easier. What really decided those three weeks was the discussion in week one that cut the scope down to five modules.