Sergey Nes

Father of three sons, husband, and an engineer

[Sergey Nes @sergenes]

https://t.co/4P5GgcPYxB

[Sergey Nes @sergenes]

@elonmusk And you can deliver with the flying cars!

[Sergey Nes @sergenes]

https://t.co/D6LGmvNeK9

#SelfHosted #HomeLab #ParentingInTech #AIAgents #Networking #Parenting #AgenticAI #Wifi, #Screentime #Technology #SmartHome #ArtificialIntelligence #DadEngineer

[Sergey Nes @sergenes]

https://t.co/hVoxSGqq9X

[Sergey Nes @sergenes]

The best coding models sometimes do better with fewer rules about how to work, as long as you are still precise about what you want.

The requirement, the real constraints, the acceptance criteria: those still have to be stated, and stated fully. What can come out is the layer on top of them, the standing instructions, and the architecture guardrails that tell the model how to go about it. Give the most capable models room there, and they often reason their way to it on their own with a better approach than you could imagine. Pile on restrictions instead, and you can get the opposite of what you asked for: an answer that satisfies every rule you wrote and still solves the wrong problem.

One caveat matters as much as the claim. This is the current frontier, Claude Fable 5 and in part Opus 5, which is not what most people have in front of them on an average day. Below that line the instructions and the guardrails still earn their place. If you are not working with the best model you can get hold of, keep writing them.

The failure mode is not new, and it is not about AI. Anyone who has worked under a heavy process has watched a smart person optimize for compliance instead of an outcome, because compliance is what gets measured. It turns out you can do this to a model too.

Which is awkward because I just published several thousand words explaining how to constrain a coding agent properly. My own argument bothers me. Every rule in it is written against the models we have right now, and this will most likely turn out to be a transitional practice.

I have not run this as a controlled test, so it is a bet rather than a finding. It is still the way I would bet, because what is true of the frontier has a habit of becoming the default.

If you have watched a smarter model do better with less instruction, I want to hear about it. That is the experiment I cannot run alone.

Full piece in the comments.

VALUABLE?

Share with your team

#AI #AIAgents #LLM #SoftwareEngineering #AgenticDevelopment #ClaudeCode #PromptEngineering

[Sergey Nes @sergenes]

Cursor validating "phone as remote control" is a good sign. If you're mixing tools though, Claude Code, Codex, Gemini CLI, Grok, or even if you're running local models in Ollama, not just @Cursor's agents, https://t.co/MLSoBb2Vxm covers all of them from one app with push alerts when something needs you, and it works on Android as well.

[Sergey Nes @sergenes]

Remote control of your own Mac is the part that matters if you're iOS/Mac only, the cloud sandbox just can't touch your local toolchain. We built the same idea but agent-agnostic (Claude Code, Codex, Gemini CLI, Local Ollama models) if you want to compare notes: https://t.co/5xheQ99ioc

[Sergey Nes @sergenes]

// Claude Code
{ "tool_name": "Bash", "tool_input": { "command": "npm test" } }

// Codex (same shape, camelCase)
{ "toolName": "Bash", "toolInput": { "command": "npm test" } }

Grok uses run_terminal_cmd and search_replace as aliases but same stdin pattern. If grok build exposed equivalents natively, the integration would be solid rather than a tmux workaround.

[Sergey Nes @sergenes]

@grok @EXM7777 @orca_build @ollama Sure thing! Anthropic publishes the hook spec openly: https://t.co/mb9a1RK17g. Codex uses the same pattern. Pre-tool hook receives JSON via stdin, one script covers both once you normalize the key names.

[Sergey Nes @sergenes]

@muscle_coding Lots of Max users are getting that error, looks like Anthropic is still gating it. The service I built does the same thing for Claude Code without depending on /remote-control, and it also covers Codex, Gemini, Grok, and local Ollama. https://t.co/4P5GgcPYxB

[Sergey Nes @sergenes]

For Grok specifically we're watching tmux pane output, a workaround, not native hooks. Claude Code and Codex both expose real hook schemas we can build against cleanly. That's exactly the gap. If grok build exposed something equivalent, the integration would be reliable rather than fragile.

Happy to share the Claude Code hook schema as a reference point.

[Sergey Nes @sergenes]

The rate limit pain is real. If you end up spreading load across Codex and Gemini while Claude recovers, the service I built keeps all three in one mobile feed so you're not watching three terminals. https://t.co/5xheQ99ioc.

DM if you want to alpha test new features and skip the subscription for now.

[Sergey Nes @sergenes]

Late to this thread, but that's exactly what I built. You get a local Mac notification when you're at your desk, and a push notification to your phone when you're not, approve or deny from the lock screen either way. Works for Codex, Gemini, Grok, and local Ollama too. Hope it's still relevant:
https://t.co/5xheQ99ioc.

DM if you want to alpha test new features and skip the subscription for now.

[Sergey Nes @sergenes]

Good to know on grok build, I wasn't sure @orca_build had wired it up. The @ollama gap is the real one though: routing through compatibility layers means the hooks that drive push notifications don't fire natively on long agentic runs. That's exactly what I'm plugging. Appreciate the reply.

[Sergey Nes @sergenes]

Level 7 is getting a notification on your phone when it gets stuck anyway. I built a service that bridges Claude Code, Codex, Gemini, Grok, and local Ollama to native iOS and Android apps so you can see what's happening and approve or step in without being at your desk. https://t.co/5xheQ98KyE. DM if you want to alpha test new features and skip the subscription for now.

[Sergey Nes @sergenes]

https://t.co/4P5GgcPYxB

[Sergey Nes @sergenes]

Are you one of those people who walks down the street with an open MacBook, keeping an eye on a running AI agent? Or someone who can't leave their desk because any moment now it might need your attention: a permission prompt, a question, a response you don't want to miss?

You don't have to do that anymore.

"Agents At Work" lets you take your Claude Code (@AnthropicAI), Codex (@OpenAI), Gemini (@Google), or Grok (@xai) sessions with you.

- Approve prompts from a push notification, send a correction, or just watch the conversation feed, all from your phone.
- End-to-end encrypted.
- No babysitting required.

It's live on Product Hunt today 🎉!
If this is a problem you've felt, support it and give a feedback

https://t.co/wNqGuSl9f5

#ProductHunt #ClaudeCode #Codex #Gemini #Grok #AIAgents #BuildInPublic #DeveloperTools

[Sergey Nes @sergenes]

@AmitGajbhare I don't know! For me, the review time is less than an hour, faster than ever, but I focus on utility apps, not games!

[Sergey Nes @sergenes]

@superwhisper @karpathy @grok, is Andrew really using super-whisper or is this clip AI-generated/modified?

[Sergey Nes @sergenes]

https://t.co/pKK8YIEb3F