The Morning Briefing Pattern: What My AI Running Coach Taught Me About Reporting

Marathon Tracker is software I built for an audience of one: it syncs my Garmin watch, folds in sleep and the weather each run was actually run in, and judges my marathon training every morning. Selfish software, but it forced me to solve two problems every business has: reports nobody reads and AI nobody should trust. The solutions transfer.

Dashboards are where numbers go to be ignored

Every business has dashboards. Almost nobody looks at them on purpose. A dashboard is pull reporting: it waits for you to remember it exists, navigate to it, and know which of its 40 numbers matters today. The half-life of dashboard attention is about 2 weeks after launch.

Marathon Tracker's answer is the morning briefing: the first time I open it each day, it has already written three sentences about today. Long run day: 8 miles, cap 147. Dew point hits the humid band by 8am, so expect drift. Sleep was short; don't chase pace. I didn't query anything. It knew what today was, looked at what changed, and led with what matters.

That's push reporting, and the pattern works for any operation: the Monday email that says "two jobs are trending over budget and here's which," the morning note that says "yesterday's orders: 41, one flagged; the Hendersons' address didn't validate." The report that arrives with a verdict gets read. The dashboard that waits gets archived.

The test for any recurring report: does it lead with a sentence a busy person can act on? If it leads with a chart, it's homework.

The rule that keeps the AI grounded

The briefing is written by AI, and that should worry you: AI writes fluent nonsense with total confidence. So Marathon Tracker draws a hard line I now consider non-negotiable in client work:

Code calculates. AI explains.

My goal-risk score (am I on track for the race?) is deterministic arithmetic: five weighted components, same inputs, same answer, every time. The AI is allowed to explain the score and recommend what to do about it. It is not allowed to invent it. The verdict can never contradict the math, because the verdict is the math.

And when the assistant answers a question, it shows its work ("looked at: your runs, your sleep") or admits "answered without looking anything up." That one line changes the relationship: you always know whether you're reading analysis or vibes.

Put those together and you get AI reporting a business can lean on: numbers computed by code you can audit, narrated by AI you can check, delivered before you asked. Skip either half and you get the failure modes: trustworthy numbers nobody reads, or readable prose nobody should trust.

The quiet prerequisite: your data has to show up by itself

None of this works if a human has to export CSVs first. Marathon Tracker's pipeline pulls from my watch automatically, down to per-second heart rate, and enriches every run with the weather at that GPS point and hour. The briefing is only as good as the plumbing under it, which is why most "reporting projects" are really integration projects wearing a nice shirt, and why I usually start there.

If your Monday numbers take half a day to assemble, or arrive as a dashboard nobody opens, the briefing pattern is buildable on top of the systems you already run. Tell me what your ideal three sentences would say, and I'll tell you what it takes to make them write themselves.