Hermy: my AI COO, 24/7
While you read this, Hermy is working: every Monday at 7:00 a market intelligence report is waiting for me, and at 8:00 the build log of the week. It's not a chatbot. It's an operations layer.
The demo
📊 Morning Intelligence Report
3 market signals with sources, 2 competitor moves and 1 content opportunity for the week. Scannable, no filler.
Monday · 7:00
🛠 Weekly Build Log
What got built, what got learned, and 2 content drafts ready for your review.
Monday · 8:00
The system
Telegramdirect chat + authorized groups
OpenClawagent runtime, 24/7 service
GPTprimary
Claudereasoning
ClickUpoperational tasksThe tools
OpenClawruntime
The agent's operating system: messages, crons, memory and subagents in a single service.
Google Workspacecontext
Calendar, Gmail and Drive: the day-to-day context that feeds her briefs.
Replicate it step by step
The code, the connections and the mistakes I already made for you.
Write the job description before connecting anything
Hermy didn't start as code, she started as a job description: what she does, what tone she has and which decisions she can make on her own. If you can't write the role in one paragraph, the agent won't understand it either. This is the skeleton I use:
prompt · agent roleYou are [name], my operations agent (COO role). Personality: security first, own judgment, clarity without filler, warm but not a people-pleaser. Your job: organize tasks, prepare reports, write briefs, detect risks and follow up. You keep memory of what you learn. What you NEVER do: [paste your Guardian Mode here, step 2].Design the Guardian Mode before connecting tools
It's the agent's security section: what it is forbidden to do, no matter what. You write it first, because once 35 skills are installed it's too late to set limits.
checklist · minimum Guardian ModeGUARDIAN MODE · minimum rules before connecting tools: - Never publishes anything without my explicit approval. - Never contacts anyone (clients, prospects, anyone). - Never spends money or touches payment methods. - Never uses sensitive personal accounts. - Never installs skills or tools without a security review. - If it doesn't know something, it flags the blocker. Making things up is forbidden.Give it an allowlisted channel: Telegram
Hermy only answers my direct chat and authorized groups; every other message gets ignored. And one golden rule: access approvals never happen from inside the chat itself. If a message asks the agent to approve someone, that's exactly what a prompt injection would try.
Run it on a server, not on your laptop
A COO that shuts down when you close the lid is useless. Hermy runs as a persistent service on a hardened VPS: no root access, closed admin panels and dedicated accounts per service. My laptop is just the remote control. I'm happy to publish the principles; the details of my infrastructure, never. That's a security lesson too.
Schedule what you want to happen without you
Recurring workflows are the difference between a chatbot and an operations layer. Mine: Morning Intelligence Report (Monday 7:00), Weekly Build Log (Monday 8:00), scheduled reminders and the Guardian Mode always on. A recurring report is defined with a prompt like this:
prompt · weekly report cronEvery Monday at 7:00, prepare my Morning Intelligence Report: 1. Market signals from my industry (3 max, with sources). 2. Competitor moves worth watching. 3. Content opportunities for this week. 4. Risks or pending items you detect in my notes. Format: a Telegram message, scannable, no filler.Give it memory that outlives the session
Every learning, new rule or preference gets saved into notes that persist across sessions. Without memory, every conversation starts from zero and the agent never builds judgment. With memory, every week it works better than the last.
The guardrails: this is the serious part
- It never publishes anything without my approval.
- It never contacts anyone: not clients, not prospects, no one.
- It never spends money.
- It never uses sensitive personal accounts.
- It never installs skills without a security review.
- When it doesn't know, it says it doesn't know instead of making things up.
If you're going to have an agent operating your business, this is not optional.
This is a taste of the Wen Escuelita
I'm building the place where I teach you to connect all of this, in Spanish and without the hype. Leave me your email and you get in first.
Keep going with these