📬 AI Controls My Inbox: I Had to Select One Trusted System

🔥

The results across the three systems weren't consistent.

ChatGPT Scheduled Tasks, Cowork, and Gmail's own AI Inbox each caught different things. Together, they covered everything. Separately, none of them did.

So I found myself doing something I hadn't planned on: bouncing between all three, cross-checking one against another, instead of trusting any single one to just handle it.

That's not sustainable.

That's a person doing the job the AI was supposed to do.

Around the same time, on an unrelated but related project, I started building out my AIOS — what I've been calling my second brain. Getting that set up required real, sustained effort inside Cowork.

That's where I actually learned how Cowork's scheduled tasks work. Not the surface version. The real mechanics — task files, hard constraints, a run that reads a fresh spec every time instead of carrying memory forward.

It was clear to me that Cowork was the best choice for mission critical triage at this point, and therefore the scheduled task is much more robust.

📝 What we actually built in Cowork

A daily scheduled task that runs the inbox triage automatically, no manual trigger
A broad Gmail search across the full inbox, not just "unread" — misclassified emails don't show up if you only look at unread
Every email sorted into one of three buckets: meetings, business development / prospects, or needs a look
Every meeting request cross-checked against the calendar for conflicts before anything gets touched
One narrow auto-accept rule for a specific type of meeting invite — all other meeting notices get flagged for review, not guessed on
Replies created as drafts only — nothing ever sent automatically
Existing Gmail labels reused, never invented on the fly
One consolidated report at the end of each run: Meetings, Business Dev, Unsorted — nothing dropped silently

That's the skeleton. Here's what happened once it actually ran.

The Cowork layer is a written task file. It gets read fresh every run. No memory of the last one. Nothing to slowly drift.

📝 Lesson one: hard rules beat clever wording.

The task file doesn't say "be careful."

It says:

Never send an email. Draft only.
Never delete a calendar event, even an obvious-looking duplicate.
Never delete anything without asking me first, live.
One narrow auto-accept rule, for meetings from a specific client. No other meeting requests get accepted automatically.

📝 Lesson two: the first real failure wasn't a missed email. It was a silent one.

On an early run, the triage tried to tag two email threads with existing Gmail labels, so I could find them faster later. The Gmail connector didn't have write access at that moment. It could read my inbox and search it fine. It just couldn't apply the labels.

It didn't pretend the labels went on.

It told me, in the report, that the write failed.

A system that fails loudly is a system I can actually trust.

📝 Lesson three: working correctly and looking correct are not the same thing.

Remember that narrow auto-accept rule from Lesson one — meetings from a specific client get accepted automatically, no review needed. After one run using it, I said: I still see meeting requests in my inbox.

Turns out every invite had been accepted correctly. But accepting a meeting with Cowork doesn't automatically mark the email as unread, that requires two different prompt instructions. Therefore, the messages just sat there, unread, looking exactly like something got missed.

Nothing was broken.

It just looked broken.

If I hadn't asked, I would have assumed the system dropped the ball.

The fix wasn't a better explanation.

It was one line added to the task file: when you auto-accept, mark the email read too.

Now every future run marks a meeting as accepted and email read without me asking.

📝 Where this leaves the three-system setup.

I'm not dropping ChatGPT or Gmail's AI Inbox, just yet. Each one still catches things the others miss, and that redundancy is still doing real work.

But Cowork is now the layer I trust to triage my email and calendar without obsessively checking my inbox.

That's what the second brain project actually handed me. Not a better prompt. A better foundation to build the triage on.

10 comments

📬 AI Controls My Inbox: I Had to Select One Trusted System