Write the rules down before you automate more

I spent six weeks building a multi-agent assistant that pulled nine apps into one place. Then it broke for two weeks and I didn't notice. When I came back, I didn't rewrite the code. I wrote the filing rules on a single page.

Same models. Same stack. Usage roughly doubled. The useful lesson wasn't technical.

The failure mode nobody demos

Most AI write-ups stop at "it works." Mine failed quietly. An expired token. Two automations pointed at a folder I'd deleted. Nothing crashed — it just stopped being useful, and I stopped opening it without deciding to.

What "the rules" actually are

Not a second brain manifesto. A short list of where things go, in language a tired human and a literal agent can both follow.

• What lands in tasks vs notes vs reference files
• Which chat / inbox means "file this" vs "answer me now"
• What never gets auto-filed (money, people, irreversible sends)
• What "done" looks like so you can audit it later

How to write yours in an afternoon

• List the last 20 things you captured across apps. Group them into 4–6 buckets max.
• Name one home for each bucket. If two homes feel right, pick the worse one on purpose — ambiguity is the bug.
• Write the routing rule as a sentence: "If it's a next action with a date, it goes to Todoist. Everything else that isn't a file goes to Obsidian inbox."
• Put that page somewhere both you and your agents read on every run. If it's only in your head, you don't have a system.

What to measure instead of vanity

I tracked cost and message counts because they were free. I should have tracked whether I could find what I'd saved. That's the only number that justified the project.

If you build something this week, make the rules page first. Automate second.