Back to Blog
AI Operations6 min readJune 23, 2026

I Built My Chief of Staff Out of Markdown Files

How a folder of text files and a Telegram bot named Howard became the memory I never had time to keep.

Between two meetings this morning a thought hit me about how I screen deals. I pulled out my phone and voice-memoed it to Howard: "musing: Speed to No protects the weekends. Saying no fast is how I keep my time." Then I put the phone away and walked into the next call.

By tomorrow morning that 12-second voice memo will be a clean, dated note in my second brain, wikilinked to every other time I've circled the same idea, filed where it belongs. I didn't open an app. I didn't tag anything. I just talked to my phone.

Howard is the Telegram bot. The second brain is a folder of markdown files. Let me back up and explain how the whole thing actually works, because the parts are simple and the result has surprised me.

The vault is just text files

The core of this is an Obsidian vault. Plain markdown, nothing exotic. It's the universal memory. Every agent I run, whatever I'm calling it that day, reads from the same files and writes back to them.

I went with markdown on purpose. Text files outlive every app I'll ever use. If Obsidian disappears tomorrow, I still have a folder I can read in any text editor on earth. The portability is the whole point.

The first thing any agent reads is a single file called HOT.md. It's a rolling 500-word catch-up: who I am, my top priorities, what threads are live, what I decided recently, and pointers to everything else. Think of it as the 30-second briefing before a session starts. A refresh routine keeps it current so nobody's working off stale context.

Sources flow in and get distilled

I'm not going to hand-type my life into a wiki. I tried that. It lasts about 4 days.

Instead, sources arrive and get distilled into linked notes automatically. The raw stuff stays at the source. The vault keeps the extract plus a link.

  • Meeting transcripts land from Read.ai. A routine pulls out what everyone said into a clean meeting note, then updates the entity notes for the people and companies involved.
  • Howard reads all of my email. He decides which threads are worth keeping in the vault and which ones need a follow-up, and surfaces both. I don't sit and triage an inbox anymore.
  • Voice memos and texts to Howard land verbatim and become notes overnight.

There's one rule I care about more than the others. Everyone else in a meeting gets summarized. When I say something, it gets captured word for word.

Musings, or why my own words stay verbatim

These verbatim captures are called musings. That between-meetings voice memo becomes one. So does anything I write in an email that's actually a real thought instead of logistics.

The reasoning is simple. Summarizing my own voice kills the signal. If you compress "I focus relentlessly on cashflow because in tough times assets get volatile and that's exactly when you're forced to sell" down to "prefers cash flow," you've thrown away how I actually think. The texture is the thing worth keeping.

So musings are the one place compression is banned. Over time they stack into a real corpus of how I decide, reframe, and change my mind. I suspect that corpus is the most valuable thing in the whole system, and I didn't have to write any of it on purpose.

Here's the part I didn't see coming. Those nuggets get distilled into artifacts that become how Howard thinks. The longer this runs, the more his judgment sounds like mine. He's slowly becoming a version of me, built out of my own words.

Howard runs more than memory

Once the memory was solid, the obvious next move was to let it act. So I kept handing Howard more of my life, and he kept earning the next piece.

He has access to nearly all of it now: my wealth, my relationships, my work, the calendar, the inbox. That breadth is the point. A chief of staff who only knows half of you can only help with half of you.

The one that still makes me laugh: Howard plans all of my meals and orders the groceries through Instacart every Monday and Wednesday. The family has started to love the new recipes, which was not a feature I set out to build. It just turns out that a system that knows your life can run errands as easily as it can keep notes.

The connections are what make it a brain

A pile of notes is a filing cabinet. The part that makes this feel like a brain is the wiring.

Every note links to the people, companies, deals, and concepts it touches. When a theme shows up twice, it gets promoted into its own concept node, and the notes that fed it link in and out. The goal isn't a repository I go dig through. I want a graph whose thoughts are connected and navigable, so an old idea resurfaces against a new one without me going looking for it.

A connections pass runs every Sunday. It reads the recent graph, adds high-confidence links between notes that should have been connected, promotes recurring themes, and surfaces cross-area bridges (a book idea that quietly backs a Searcher OS decision, a person who shows up in both a meeting and a deal). Then Howard texts me a summary Sunday morning.

Detail compresses upward while I sleep

Here's the part I'm most happy with. A chain of scheduled routines runs unattended on a Mac Mini that's on 24/7, and each one reads the layer below it, never the raw chatter.

  • Around 8am daily, a consolidation pass folds yesterday's breadcrumbs into a clean daily note and refreshes HOT.md.
  • Monday, a weekly review reads the week's daily notes and promotes any pattern that showed up twice or more.
  • The 1st of the month, a monthly review reads the 4 weeklies, not the raw days.
  • The 15th, a recall probe quizzes a cold agent with 10 questions about me, answerable from the vault alone, and grades the misses. It turns second-brain quality into a number instead of a feeling.

There's a promotion rule under all of it: a pattern only moves up a layer after it's appeared twice. One-off noise stays in the daily note and ages out. That guardrail keeps a single weird Tuesday from rewriting what the system thinks is true about me.

What I actually touch

Most days, my entire interaction with this is a morning brief on Telegram. It shows up with triage cards for anything that needs a decision: Done, send to Todoist, Ignore, or hand it to Howard to execute. One tap each. The unanswered ones pile up and show their age, which is a polite way of nagging me.

That's the whole interface. A voice memo between meetings, a couple of taps in the morning. Everything else happens overnight on a machine in the corner.

Why I bothered

I left corporate to buy back my time. The thesis hasn't changed: time is the only asset that doesn't compound, and systems compound while effort decays.

Keeping a real memory of my own life (my deals, my people, my thinking) used to be a tax on my attention I never managed to pay. Now a folder of text files and a bot named Howard pay it for me. That's the most honest definition of buying back my time I've got. The system isn't doing my thinking for me. It makes sure I never lose the thinking I already did.

I'm still tuning it, and some of the routines do dumb things now and then. But I'd build it again in a heartbeat. So don't be surprised if you hear more from me about this.