← SELECTED WORK
PERSONAL SYSTEMS / CASE STUDY

CHIEF OF STAFF.

A personal AI system for figuring out what needs my attention, before the day gets away from me.

Three email accounts and two calendars are an effective way to learn that an unread count has nothing to do with what matters. I built a morning brief that notices what changed, remembers what is still open, and gives me a short list to react to.

01 / THE PROBLEM

The important thing is rarely the loudest thing.

A deadline is easy to spot if it announces itself. A school form halfway through a thread is less cooperative. A commitment can disappear from the inbox without being finished at all. I wanted the brief to care about that last category especially.

An unread count doesn't tell me any of that. An inbox summary leaves me with the same problem I started with: what actually needs me?

02 / THE IDEA

Don't summarize the noise. Find the signal.

I wanted the day to collapse into a smaller set of things worth thinking about. The output needed to be a starting point for decisions, not another feed to manage.

What changed?What's open?What am I waiting on?What am I committed to?What needs me?
03 / WHY TWO BRAINS

One model gave me an answer. I wanted another set of eyes.

I gave Claude and ChatGPT the same shared context and compared their readings. In my early runs, one tended to wait for stronger evidence while the other surfaced more possibilities. That tension helped the brief distinguish what it knew, what it inferred, and what I still had to decide. The different inbox connections are a practical constraint; the comparison happens after the facts are shared.

ONE READING
Claude

In early runs, tended to wait for stronger evidence before calling something urgent.

ANOTHER READING
ChatGPT

In early runs, tended to surface more possibilities from the same facts.

Angel and devil on the same shoulder, working the same problem.

04 / THE SYSTEM

What needs me today?

Nothing about who reads which inbox was a strategy choice. It's a connector limit. What both assistants do with the shared context in Notion, that's the part I actually built.

05 / FRESH INPUT. PERSISTENT CONTEXT.

The email gets reread. The commitment doesn't disappear.

Email is today’s evidence. The open-items ledger is memory. A message can disappear below the fold; the thing I promised to do cannot disappear with it. Notion holds what is still open, rather than a copy of the inbox.

EVERY RUNReread.

Current messages and calendars are reassessed against today's context.

ACROSS DAYSPersist.

Commitments, waiting items and intentions remain available until their state changes.

06 / THE BRIEF

Less noise. Clearer next moves.

Everything collapses into something I can use without managing another system. I don't need twenty summaries or a dashboard of unread counts. Unfinished items sit untouched until I deal with them. Completed items get marked in Notion and become the shared truth for the next run.

ILLUSTRATIVE BRIEF · ANONYMIZED
Act today.
Follow up.
Monitor.
Act todayComplete a time-sensitive school form.
Follow upSign the tenant contract still waiting on me.
MonitorKeep an open intention visible without inventing urgency.
07 / THE HUMAN

The system doesn't run my day. I do.

It can surface.

It gathers signals, interprets what changed, connects context across tools, and recommends what may deserve attention.

I decide.

I review the brief, disagree when it gets something wrong, change priorities, and choose what happens next.

The goal isn't autonomy. It's better judgment.

08 / WHAT BROKE

The first version wasn't the system. It was the test.

01 / NOISE

Everything showed up twice.

An old forwarding rule duplicated every email. I killed the forward before asking the system to judge what mattered.

02 / ACCESS

The unattended run couldn't see.

Screen control couldn't reach a locked machine. I moved the unattended digest to a browser session.

03 / SCOPE

Too much at once hid the failure.

I tried to build the whole workflow in one pass. I started small, isolated each failure, and added one piece at a time.

09 / WHAT IT CAUGHT

The harder one isn't obligations. It's intentions.

CAUGHT

An unsent client response.

I thought I'd replied. The response was still a draft, and the commitment was still open.

CARRIED FORWARD

A tenant contract.

Stayed on the list without a new email to push it up.

REMEMBERED

My own intentions.

Projects stay visible when louder things take over.

10 / WHAT I LEARNED

The hard part wasn't connecting AI to my inbox.

The useful wins have been ordinary: a reply I thought I sent that was still a draft, a contract with no new email to remind me, and projects that mattered to me but were losing a volume contest to my inbox. I do not need another system to tell me I have a meeting in twenty minutes. I need help remembering what I said I would do after the meeting.

Automation handles the mechanics.
AI helps interpret the ambiguity.
I keep the judgment.
STILL A WORK IN PROGRESS.
NEXT: Selected Work →