top of page

What I Actually Built to Get Running by Day One

Writer: Amy Westlake
Amy Westlake
Aug 18
4 min read

Situation

Last week I wrote about the moment someone said a name I didn't recognize in a meeting, and I already had a profile on them. That post was about what it felt like. This one is about what's actually running underneath it, because a few people asked, and because "I connected some tools" isn't a setup anyone else could actually copy.


The AI Move

There are two scheduled tasks running in Claude Cowork, one every morning and one every evening. No custom app, no code I wrote by hand. Just two sets of plain language instructions that run on a timer.


The morning job pulls today's calendar. For every real meeting, nothing marked out of office or focus time, nothing I've declined, it builds a note in the notes app I keep in Obsidian. Each note gets who's organizing it, who's on it, and a short prep section pulled from a handful of places: the meeting's own description if there's a doc attached, my open task list if something on it matches this meeting's people or topic, a quick look at recent email and Slack for anything directly relevant, and a check of my own past notes in case this exact thing already came up with these same people before. Then, because Zoom doesn't hand over a transcript the second a meeting ends, the job also schedules three follow-up checks, thirty, sixty, and ninety minutes after each meeting, to go back and grab the transcript and summary once they exist and drop them into that same note.


The evening job is the cleanup pass. It goes back through whatever meeting notes got created that day, pulls out anything that reads like an open question or an action item, and checks it against my existing task list before adding anything new. It also builds and updates a short profile on anyone I met with that day. A few consistent sections: personal background, work history, what they're actively focused on right now, and how they communicate, split between how they talk in meetings and how they write in email and Slack. Simple one on one meetings get filed into their own folder automatically. Notes about a specific program get filed under that program. Everything else stays put. Then it does one more pass through Slack for anything that never showed up in a meeting at all.



A few things made this actually hold up past the first two weeks.

  • Give it real boundaries, not just permission. The prep searches each run once, with a cap on how many results come back, and no retrying an empty search hoping for something better. Without that, it would happily spend ten minutes digging through everything it has access to for a meeting that just doesn't have much prep material, and I'd never see it actually finish.

  • Never let it overwrite what I wrote myself. Every note has one section that's mine alone, and the instructions are explicit that nothing automated ever touches it, only adds around it.

  • Pull the full history before deciding what still counts as open. The evening job checks my entire task list, closed items included, not just the active ones. Otherwise an old note phrased as a question could resurrect something I'd already answered weeks earlier.

  • Design for the day it doesn't run. The evening job also looks at yesterday's meetings, not just today's, in case my computer was off or the job just didn't fire. That one small choice means a missed run doesn't turn into a missed day.


The Shift

None of this showed up at once. The first version tracked tasks in a plain text file, and within the first couple of days I moved that over to my actual task tracker instead, because a text file living somewhere I had to remember to open wasn't going to survive a real week. About a week and a half in, I rewrote how the prep section matches meetings to open tasks, because it kept surfacing things that had nothing to do with the meeting I was about to walk into. Most recently, I added Slack as a source it checks before building morning prep, and moved where everything's stored so it lives somewhere I can reach from any machine, not just the one I set it up on.


I didn't plan any of that in advance. Each change came from actually using it for a week and noticing exactly where it was wrong.


Here's the part I didn't expect. When I sat down to write this post, I couldn't remember the order any of that happened in. So I asked Claude to look back through our recent chats and reconstruct it: what changed, roughly when, and why. It came back with a clean list, in order, without me reconstructing two weeks of edits from memory. I hadn't been keeping a changelog. I didn't need one.


The Pattern

I don't think of this as something I built once and now maintain. It's more like giving direction to someone new: correcting them in plain language every time something comes out slightly wrong, and treating that correction as the real process rather than a sign the first attempt missed. Each version came from noticing exactly where the last one fell short, one gap at a time, after actually living with it for a week.


The Implication

If you want to try something like this, don't start with two jobs running twice a day. Start with the smallest piece of your own routine that repeats often enough to be worth the effort: one meeting type, one recurring report, one folder you keep meaning to clean up. Describe it in plain language to whatever tool you have connector access to. Tell it exactly what it's allowed to touch and what it's never allowed to touch. Then use it for a week before you change anything.


And if you lose track of what you've changed along the way, don't try to remember it. Ask whatever you've been using to look back through your own conversation history and tell you. Mine did, and it was right.


What I'm Testing Next

I'm not doing anything yet with the written communication profiles beyond letting them build quietly in the background. The next thing I want to test is actually using one before a difficult email instead of just letting it accumulate.

Comments


bottom of page