What worked, what failed, and what we got wrong the first three times. Published when a piece is ready rather than on a schedule, because a weekly slot is how writing gets thin.
Real things people built or tried with AI, traced to the original source, with what you can take from it tomorrow morning.
An agency's flashy research agent made a startup almost nothing, while a plain text and voice agent for a realtor's 60,000 old leads became its best paying client.
A hospitality consultant replaced open ended consulting with a fixed $199 report where AI drafts the analysis but a human checks it first.
A Reddit poster built a personal Monarch-style finance dashboard using Opus 5.5 for under $20 in Claude usage.
A builder shadowed a primary care clinic's referral staff for two days before automating outbound referrals, then rolled it out slowly under supervision.
A solo operator swapped Buffer for his own LinkedIn script, and for a few days two systems could have posted the same thing twice.
A small agency let an AI agent send invoices on its own, until one had a wrong charge and one had a mistyped IBAN.
A marketing employee with no software background used AI coding tools to build an internal finance platform and a facial verification system for employee attendance.
These are still being written, so you can see the shape of what is coming.
Crossing from the side that specifies to the side that ships. The things that turned out to be easy, the things nobody warns you about, and why the first four automations we built were all solving the wrong problem.
In the first meeting everyone likes it. In the second, someone has to give up something real. What we learned watching hundreds of companies decide whether to actually change how they work, and how it applies to shipping AI inside a team.
The filter itself, written out. What it keeps, what it throws away, and the correction loop that makes it sharper every week. You can point the same thing at your own sources.
Two years of Obsidian, most of it wasted collecting. What changed when capture stopped being free and every note had to survive a weekly pass to stay.
The tips page is already live and moves faster. Start there.