I run a team of assistant bots now. One of them, Bill, tracks money. For a while, Bill was working out of a Google Sheet.
That worked until it didn't. A spreadsheet is fine for me. It is a bad interface for an agent. There is no contract, no clear shape for the data, and nothing that tells the bot what it is allowed to do.
The bot asked for the feature
At some point Bill effectively asked for a better way to track household money items. So I built it.
adamaoc.me is my personal admin site. I added a "money watches" feature there: a household money open-items API at /api/money/watches. It replaced the Google Sheet. I filmed the whole thing and shipped it on camera with Grok Build.
The video is called "Grok Bot asked for a feature. Grok Build built it." That title is pretty literal. One agent found the gap, another agent helped close it, and I sat in the middle making the calls.
Video: https://youtu.be/YqJQWZnANnc
Then I did it for everything else
Once money had a real API, the rest of my data looked wrong sitting behind a login screen. So I added bearer-token APIs for notes, projects, and journal. That merged on September 15, 2026.
Now my assistant bots can read and write my data with a machine token. No scraping. No pretending to be me in a browser. No copying rows between tabs.
The part I didn't expect to matter as much: public "skill" docs. Each API has a doc that tells an agent how to use it. What the endpoints are, what the data looks like, what it should and shouldn't do. The bot reads the skill, then calls the API.
The pattern
This is the thing I keep coming back to.
Give agents real APIs plus a skill doc. Not spreadsheets. Not screenshots of a dashboard. Not a long prompt describing how your data is organized.
The API gives the agent a stable contract. The skill doc gives it the instructions. Together they turn a chat bot into something that can actually do work on your stuff.
The friction is real. You have to build the endpoints, think about tokens, and write the docs. But that work pays off every time a bot does the right thing without me babysitting it.
What's next
Next up is Money Watches v2. That's issue #66 on adamaoc.me: card balances, a daily server-side re-check, and an attention endpoint so the bots can ask "what needs me right now?"
I plan to build this one myself. Grok 4.7 finally launched on September 21, 2026, after repeated delays, so it's a good time to put it to work.
The lesson
Projects come and go. The money sheet is gone. The first version of watches will get replaced too.
The projects die. The lessons don't. And the lesson here is simple: if you want agents to help, give them a real door to walk through.
It's pretty crazy how much changes when you do.