Day 1: Teaching Sully My Business

Yesterday I gave my AI assistant a personality. Today I gave him a job.

The Integration Grind

The fun part of building an AI assistant is the personality, the first conversations, the “holy crap this is cool” moments. The useful part is the boring stuff — connecting it to the systems where your actual work happens.

I started with Toggl. I do consulting through App Vitals, and every hour needs to be tracked. I’ve got clients across multiple projects with different rates. Missing even an hour of tracking is real money left on the table.

Getting Sully connected to the Toggl API took about 30 minutes. API token in a secrets file, workspace ID configured, and suddenly Sully could see all my time entries. But I didn’t want a passive integration — I wanted an active one.

So I set up an end-of-day report. Every weekday at 5 PM Pacific, Sully sends me a summary on Telegram: hours tracked, which clients, billable total. One important rule: don’t include App Vitals internal work in the billable report. That’s our own stuff, not client work.

It sounds simple. It is simple. But it’s the kind of thing that saves me from the “did I log everything today?” anxiety that hits at 9 PM.

Memory and Context

Here’s what makes OpenClaw different from just having a ChatGPT window open: memory files.

Sully has a daily memory file (memory/YYYY-MM-DD.md) where everything notable gets logged. Decisions, discoveries, things Dave mentioned, work completed. There’s also a long-term memory file (MEMORY.md) where Sully curates the important stuff — my clients, my preferences, my business context.

I spent a good chunk of today filling in that context. Teaching Sully about App Vitals, our goals, our positioning. The USER.md file got updated with my full background — 20+ years of DevOps, the companies I’ve worked at, my skills.

This is the part most people skip. They get an AI tool, ask it a question, get a mediocre answer, and say “AI isn’t that useful.” Of course it isn’t — you gave it zero context! It’s like hiring someone and never telling them what the company does.

The Compound Effect

What I’m already noticing — just two days in — is the compound effect of persistent context. When I message Sully on Telegram, he already knows:

• What I’m working on today
• Who my clients are
• My billable rate and tracking preferences
• My timezone and schedule preferences
• That I prefer voice responses over text
• That I appreciate Boston humor and ball-busting

Every interaction builds on the last. It’s not starting from zero each time. It’s accumulating.

This is the difference between a tool and an assistant. A tool does what you tell it. An assistant anticipates what you need because it knows you.

Calendar and Email (Almost)

I also started working on Google Workspace integration — calendar and email access. This one’s more involved. Google’s OAuth flow is notoriously painful, and getting it working on a Raspberry Pi that isn’t exposed to the internet adds extra fun.

I wrote the instructions, got the groundwork laid, but didn’t fully connect it today. That’s tomorrow’s battle.

The Bigger Picture

Two days in and I can already feel the shape of what this becomes. Right now Sully is a smart Telegram bot that knows my business. In a week, he’ll be checking my email, managing my calendar, tracking my time, and helping me brainstorm ideas. In a month? Building products with me.

The key insight is that usefulness compounds. Each integration makes every other integration more valuable. Time tracking + calendar = automatic time entry suggestions. Email + client context = priority inbox. Memory + everything = an assistant that actually assists.

Takeaway: The gap between a “cool AI demo” and a useful AI assistant is context. Spend the time teaching it your business, your preferences, your systems. It’s boring work, but it’s the foundation everything else builds on.