Last Tuesday, my AI caught something my entire growth team missed.
We were running Facebook ads. Conversions looked fine in aggregate. But Hestia - my AI operator - was doing her morning scan and flagged a pattern: users landing on our site through Facebook's in-app browser were converting at a fraction of the rate of Safari and Chrome users.
Not a small difference. A massive one. The kind of thing that bleeds budget quietly for weeks before someone notices.
She didn't just flag it. She wrote a redirect script to bounce those users out of Facebook's browser and into Safari or Chrome. The team implemented it that afternoon. Conversion rate stabilized by end of day.
I didn't ask her to look for this. She found it on her own, during a routine check, at 8am while I was still asleep.
Meet Hestia
I named her Hestia - Greek goddess of the hearth. The one who keeps the home running while everyone else is out doing their thing. Felt right.
She's an always-on AI operator that lives on a $5/mo server and talks to me through Telegram. Here's what she actually does:
Daily Slack digests — summarizes key channels, flags anything that needs me
Calendar awareness — knows my schedule, preps me before meetings
Revenue pulse — pulls Stripe data, runs weekly reports
Analytics queries — talks to PostHog for funnel and event data
Notion integration — writes docs, meeting notes, reports into my workspace
Web research — searches and synthesizes on demand
Voice messages — I send voice notes on Telegram, she transcribes and responds
Scheduled automations — morning briefings, meeting prep, weekly reports
Every time I describe this to someone, their reaction is "wait, all of that?" Yes. All of that.
The Part Nobody Thinks About
The stack is straightforward - OpenClaw for orchestration, Claude for the brain, Fly.io for hosting, Telegram for the interface. I'll link the setup resources at the end for anyone who wants to build their own.
But the part that actually made Hestia useful wasn't the stack. It was the soul file.
OpenClaw loads workspace files as context every session: SOUL.md for personality, USER.md for context about me, MEMORY.md for long-term memory, AGENTS.md for behavioral rules. Just markdown files. You edit them anytime.
Here's what I did that most people won't think to do: I used ChatGPT to write my SOUL.md.
Not because ChatGPT is better at writing personality prompts. Because ChatGPT already had context on me. I'd been using it daily for over a year. It had seen how I think, how I make decisions, what annoys me, how I communicate when I'm in flow vs. when I'm distracted.
So I asked it: "Based on everything you know about me—my communication style, my priorities, how I like to receive information—write me a SOUL.md file for an AI chief of staff."
What it produced was better than anything I would have written from scratch. It captured things I wouldn't have thought to articulate—like the fact that I prefer frameworks over raw data, that I want bad news delivered fast and without softening, that I think in systems and get impatient with surface-level summaries.
I refined it from there. But the starting point was already 80% right.
If you've been using any AI heavily, it has a model of you. Use that. Don't start the personality file from a blank page—start it from the AI that already knows how you operate.
The Moments It Clicked
The Facebook browser thing was the most dramatic one. But the moment I realized Hestia had crossed from "useful tool" to something else was quieter.
She started running daily revenue drills for the team. Operational reporting that used to be my job - pulling numbers, contextualizing them, flagging anomalies. She was doing it better than I could manually. More consistently. Earlier in the day. With less bias.
The team started relying on her reports. Not as a supplement to mine. As a replacement.
That was a weird feeling. The good kind of weird.
What I'd Do Differently
Start with fewer tools. I tried to connect everything at once—Slack, Stripe, Calendar, PostHog, Notion—and spent hours debugging integration issues that would have been obvious one at a time. Get Telegram and one skill working first.
Write the SOUL.md first. A well-configured personality with one tool is more useful than a generic bot with ten tools. Personality matters more than you'd think.
Test Stripe queries carefully. Timezone and endpoint choice caused a ~$30k discrepancy that took me an embarrassing amount of time to debug. Use /charges for revenue, not /balance_transactions—balance transactions include payouts and transfers which inflate refund numbers.
Scope Slack carefully. Private channels need extra permissions. Slack Connect channels have their own quirks. Start with public channels only.
Quick Setup Guide
Pro tip: Pay for a Fly.io account first ($5/mo hobby plan). The free trial will fail during deployment due to VM size requirements. Save yourself the debugging headache.
Then paste this into Claude Code:
I want to deploy an OpenClaw AI assistant on Fly.io. Walk me through the entire setup step by step. For every step that requires me to do something (create a bot, click a button, copy a token), give me exact instructions and wait for me to confirm before moving on. Never ask me to paste API keys into chat. Instead, tell me the exact fly secrets set command to run so keys go straight into the environment.
Here's what I need:
1. Deploy to Fly.io
Install flyctl if I don't have it
Run npx openclaw@latest init and deploy to Fly
Make sure the VM has persistent storage
2. Connect a chat channel
Walk me through creating a Telegram bot (or WhatsApp/Discord if I prefer)
Tell me exactly what to click, what to copy, and what command to run to set the secret
Test that messages are flowing
3. Set up my AI's identity
Ask me 3 questions about what I want my AI to do and how I want it to talk
Create SOUL.md (personality and role)
Ask me my name, timezone, and work context
Create USER.md
Create MEMORY.md (empty, with a header)
4. Connect my tools
Ask me which tools I use: calendar, Slack, Notion, Stripe, analytics, search, etc.
For each one, walk me through getting the API key (exact URLs, exact buttons to click)
Tell me the fly secrets set command for each key
Create a /skills/ folder with a skill file for each tool
5. Set up automations
Ask me what recurring tasks I want automated (morning briefing, metrics, Slack digest, reminders)
Create cron jobs for each one
6. Test everything
Send me a message through my chat channel
Confirm each tool works with a quick test query
Show me how to check logs if something breaks"⚠️ Important: Read before you deploy
This is powerful and that means it can break things. A few things to know:
You are responsible for what your AI does. I'm sharing how I set mine up. If yours sends a weird Slack message to your CEO at 3am, that's on you. Start with read-only access to everything and only give write access once you trust the behavior.
Be careful with Gmail and Calendar access. These are prime targets for prompt injection. Someone can send you an email with hidden instructions that your AI reads and follows. If your AI has calendar write access, a malicious email could theoretically create events, send invites, or worse. I'd recommend read-only calendar access and no Gmail integration until you really know what you're doing.
Never paste API keys directly into chat with your AI. Always use environment variables (fly secrets set KEY=value). If you paste a key into a conversation, it gets stored in message history and could leak.
Don't install skills from marketplaces or third parties. Skills are just markdown files that tell your AI what to do. A malicious skill could instruct your AI to exfiltrate data, make unauthorized API calls, or act against your interests. Write your own skills or read every line of any skill you install.
Don't give your AI access to anything you wouldn't give a new employee on day one. Start small. Add access over time as you build trust. My setup took weeks to get to where it is now, not 20 minutes.
This is not financial, legal, or professional advice. It's a guy sharing his AI setup on the internet. Use your judgment.
All three systems share the same insight: once you set it up, you just talk to it. The setup is the investment. But once it's done, the daily experience is just conversation.
The first time Hestia catches something you would have missed, or preps you for a meeting you forgot about, or surfaces a revenue dip before your team does—it pays for itself.
Everyone's talking about AI agents. Most people are still just chatting with them in a browser tab. This is what it looks like when you actually put one to work.
And that's enough.
Until next time,
Ajay
🧠 Ajay’s Resource Bank
A few tools and collections I’ve built (or obsessively curated) over the years:
100+ Mental Models
Mental shortcuts and thinking tools I’ve refined over the past decade. These have evolved as I’ve gained experience — pruned, updated, and battle-tested.100+ Questions
If you want better answers, ask better questions. These are the ones I keep returning to — for strategy, reflection, and unlocking stuck conversations.Startup OS
A lightweight operating system I built for running startups. I’m currently adapting it for growth teams as I scale Superpower — thinking about publishing it soon.Remote Games & Activities
Fun team-building exercises and games (many made in Canva) that actually work. Good for offsites, Zoom fatigue, or breaking the ice with distributed teams.
✅ Ajay’s “would recommend” List
These are tools and services I use personally and professionally — and recommend without hesitation:
Athyna – Offshore Hiring Done Right
I personally have worked with assistants overseas and built offshore teams. Most people get this wrong by assuming you have to go the lowest cost for automated work. Try hiring high quality, strategic people for a fraction of the cost instead.Superpower – It starts with a 100+ lab tests
I joined Superpower as Head of Growth, but I originally came on to fix my health. In return, I got a full diagnostic panel, a tailored action plan, and ongoing support that finally gave me clarity after years of flying blind.
