Roman Slysh 🔋

CEO of AI agents 🤖, Builder, Investor, Trader

[Roman Slysh 🔋 @rslysh]

@danmercede Yes, I noticed a collision between Hermes and Claude Code. I’ve added a shared-lock rule. I also once asked Codex to work with two Chrome sessions on my desktop at the same time, and it handled the task perfectly. Different setup: one agent controlling both sessions.

[Roman Slysh 🔋 @rslysh]

I use Hermes Agent every day. Mine is called Jarvis. It coordinates coding agents, works in my logged-in browser, finds flights and keeps my sales board up to date. Here’s the setup and where I keep human approval. 1/7

[Roman Slysh 🔋 @rslysh]

@aleckloss Per task for research and board updates; per send for client emails. Jarvis can prepare the work without asking about every click. I still approve the exact outbound message. Fewer interruptions during prep, but every send still needs me.

[Roman Slysh 🔋 @rslysh]

@mario_vfai Good point. “Maybe next week” should stay tentative. Keeping the original note beside the next step would make it easier to catch an invented deadline before acting on it.

[Roman Slysh 🔋 @rslysh]

Before giving an agent access to email or money, I’d define what it can do alone and what needs human approval. That boundary matters in daily use. What useful task does your agent handle regularly? 7/7

[Roman Slysh 🔋 @rslysh]

Writing: a dedicated style skill cuts stock phrases and brings email drafts closer to my voice. The caveman skill keeps replies to me short: fewer introductions, less padding, less to read. 6/7

[Roman Slysh 🔋 @rslysh]

@bitangel84 SPACE's E2B wrapper has a useful caveat: compatibility is tested against specific Python SDK versions, while the JS SDK isn't covered yet. The API shape is converging; switching backends still comes with a compatibility boundary.

[Roman Slysh 🔋 @rslysh]

@bentossell Five polished versions of the wrong screen still leave someone with the job of choosing what to change. The ability to spot what feels wrong is getting more valuable.

[Roman Slysh 🔋 @rslysh]

@thdxr Nearly 2M daily users means updates now land in a huge number of ongoing projects. Keeping those projects working through the next release seems like a bigger challenge than getting people to try the tool.

[Roman Slysh 🔋 @rslysh]

@bentossell For email triage, I’d log each run outside the chat: scheduled time, actual start, messages processed and failures. Alert if no run completes within the hour. The agent’s answer should link to that record. A plausible timestamp is too easy to invent.

[Roman Slysh 🔋 @rslysh]

Claude Code. One skill. One prompt.

20 minutes on an old Mac mini. Video below.

Voice, frame sync and rendering: all local. No video APIs. Gemini Flash, not Opus.

Want a promo video like this for your business? Leave a comment.

[Roman Slysh 🔋 @rslysh]

SwitchFix v0.0.10 is out 🚀

My free, open-source Mac utility fixes text typed in the wrong keyboard layout.

Improved corrections, short-word detection and permission setup.
Added a visual installation guide.

Try it: https://t.co/qlBahaRkvM

[Roman Slysh 🔋 @rslysh]

I built offer-design for my AI agent Jarvis: 13 methods from Alex Hormozi's $100M Offers to review offers, pricing and customer value.

Want to try it? Skip the "skill" comment. Reply with a concrete problem and I'll run it through the skill and share the result.

[Roman Slysh 🔋 @rslysh]

Before adding another AI tool, check which parts of your workflow create rework and cost.

If your project has outgrown your ChatGPT-and-agents setup, DM me. We can review the bottlenecks and work out a practical plan.

[Roman Slysh 🔋 @rslysh]

Just spent two hours helping a product manager untangle their AI coding workflow.

We reviewed ChatGPT, GitHub and CI/CD, found where the growing project was straining the process, and wrote a new plan. The redesign also removed the need for several paid subscriptions.

[Roman Slysh 🔋 @rslysh]

I also showed my own journal project. I've used a local version for two weeks; my entries are stored locally.

An agent has spent 18 hours implementing the next MVP from a written spec, flagging blockers for me to resolve. Work is ongoing. I plan to open-source it.