Three messages a day
25 October 2026
My phone buzzes three times a day with a message from my own system. 08:43, 13:43, 19:43. A morning briefing, a midday check, an evening review. I call them gm, mc and ge, because those are the words I type when I want one early.
The whole system is described in Every correction becomes a rule. This piece is about the three messages, plus one monthly rhythm. Most of what I learned is about what to leave out.
The first version talked too much
The first version ran on my laptop every hour from 8 am to 8 pm. A light cycle every hour, a full cycle every three hours, briefings by iMessage. It started failing on 9 June and I switched it off the next day.
When it came back in August on a home server, I cut it from ten runs a day to three and moved it from iMessage to Telegram. One channel instead of two.
Morning: fifteen seconds
On 7 October I rewrote the morning briefing around one number. I read it in 15 seconds. Anything that needs longer belongs somewhere else.
An illustrative one, translated from German:
Tue 03.11. gm
TOP 3
1. Send the revised proposal to a client
2. Check that the VAT return went out
3. Scan the paper mail for the tax advisor
MEETINGS
11:00 Client call
18:00 Training
ONE MORE THING
A friend has been waiting 4 days for an answer
Reply with numbers: 1 do · 2 later · 3 dropThe top 3 are actions with a verb, and a project name alone does not count. They have to be urgent and important. Urgent but unimportant things the agent does itself and never mentions. At most three meetings, the most important ones, whatever their order. A workout is never flagged as a meeting that needs preparation. If promises I made are due, up to three show up, plus at most two questions of mine nobody answered.
“One more thing” is exactly one line, the first match from a fixed order: a mail waiting for my reply for more than three days, a birthday today or tomorrow, a task overdue by three days or more, five or more untriaged tasks in my inbox.
The whole message has twelve lines. No project tables, no email lists, no social media calendar, no list of what the agent itself is working on. The last line is always the reply line, with nothing after it. No “have a great day”.
If I want the rest, I write “mehr”.
Midday: correcting the morning
The morning plan is made before the day happens. The midday check is there to correct it.
It shows what is left of the calendar, which mails came in, which threads still need a reply (oldest first, with their age in days), and which tasks are still open. Then the part that matters: three things to do this afternoon, written as actions, plus anything that will slip if nobody touches it today. Its budget is a 30-second scan, double the morning’s.
The midday run also does a sweep in the background. It compares the meetings that already happened against my meeting notes, finds transcripts I never filed, project briefs older than 14 days and overdue markers that expired. That part only reports.
It also keeps my project dashboard honest. A row with no new evidence now says “no activity since” and the date of its last entry. Back in August, the weekly refresh got an explicit ban on inventing progress where nothing was logged.
Evening: short in chat, long on disk
The evening review follows the same rule as the morning, with fourteen lines. One sentence on what happened today, one win, tomorrow’s three most important meetings and top 3, one more thing, three journal questions.
The detail goes into files. A note for tomorrow with the full schedule, what each meeting needs and the tasks that carry over. A journal page with three prompts: one about something that happened today, one from a rotation of questions, one about how I feel. Then it asks once whether I made decisions today that belong in the decisions log. Three at most. On Fridays it brings back two or three ideas from my idea inbox and drops them the moment I say “not now”.
Once a month: the sweep
The slowest rhythm has nothing to do with admin. On the first Saturday of each month a recurring task tells me to run the music sweep for my DJ crates. Research agents dig through sources per lane, every candidate is checked against my library, and the result is capped at 80 tracks with a top 10. It arrives as an email and a Spotify playlist for one focused pre-listen of about 45 minutes.
Its gate is one question per record: which record I already love does this sit next to, and would I be excited to drop it? The test went in on 16 September at 13:26, with my centre written down as peak-and-release. At 13:30 it was corrected. A centre is no fence, and in my words, “the vibe has a broad range of acceptance.”
What the corrections changed
Most corrections to the check-ins say what a message may not do.
On 13 August a 19:16 run told me all was clear while the mail tool could not even answer. Now an unreachable source is reported as unreachable, and never as “inbox clear”.
On 14 August the model wrote a good report and skipped writing the day file. Since then code writes that file and the model supplies judgment. The same day showed that skipped runs left no trace in the log: eleven silent runs looked like a dead agent.
From 25 to 29 August every slot crashed and I took the silence for a quiet week. Now a fallback without AI still sends calendar and tasks when the model fails.
On 7 October the briefings got radically short. Two weeks later they got numbers I can answer.
How many lines does your morning get?