Vlad Haltakov

Vlad Haltakov

@haltakov.net

Building products in the age of AI. Automating everything. Ex-VP Eng (cybersecurity), ex-BMW (self-driving).

195 followers 342 following1.2K posts

Vlad Haltakov reposted

If I was a mathematician... I'd try to get early access to one of the advanced math models used to generate the 700+ articles, use it to understand the latest advancements in my field, and then use it to advance the field much faster than I could before. What am I missing?

If I was a mathematician... I'd try to get early access to one of the advanced math models used to generate the 700+ articles, use it to understand the latest advancements in my field, and then use it to advance the field much faster than I could before. What am I missing?

Vlad Haltakov reposted

Grok @bot is actually really good. It's super fast, has a ton of connectors that just work, and gives me live access to X without paying for the API. I'm ditching OpenClaw as my default bot. Sorry, OpenClaw.

Grok @bot is actually really good. It's super fast, has a ton of connectors that just work, and gives me live access to X without paying for the API. I'm ditching OpenClaw as my default bot. Sorry, OpenClaw.

And since I'm in the AI-godforsaken land called the EU, I don't have access to Muse or Dots.

Vlad Haltakov reposted

AI agents killed my flow state. I've found a way to get it back. Mixing tasks that need different kinds of attention and letting agents hold the details helped me get back into the zone. Thread ↓

AI agents killed my flow state. I've found a way to get it back. Mixing tasks that need different kinds of attention and letting agents hold the details helped me get back into the zone. Thread ↓

View full thread

I still give instructions and correct the agent, and I trust my judgment more on decisions that depend on the broader context. Its job is to give me enough information to decide without making me reconstruct everything that led there.

I want the speed of AI agents and the feeling of being completely lost in the work. I'm convinced I can have both. How has working with agents changed your flow?

AI agents killed my flow state. I've found a way to get it back. Mixing tasks that need different kinds of attention and letting agents hold the details helped me get back into the zone. Thread ↓

View full thread

Letting agents hold the details frees up my attention for reviewing work and making decisions.

For a deployment, I agree on a plan, then ask the agent to track progress and guide me. When it needs my input, I want a short explanation of the situation, clear tradeoffs, and the decision it needs. A wall of text forces me to reload the whole task.

AI agents killed my flow state. I've found a way to get it back. Mixing tasks that need different kinds of attention and letting agents hold the details helped me get back into the zone. Thread ↓

View full thread

The switching feels ordered, and I can get into the zone again.

The mistake is trying to hold every task's full context in my head. My brain has limited RAM. Overload it, and every switch feels like swapping to disk: remembering where I left off and loading the problem again.

AI agents killed my flow state. I've found a way to get it back. Mixing tasks that need different kinds of attention and letting agents hold the details helped me get back into the zone. Thread ↓

View full thread

Coming back meant reconstructing what I was doing. I got a lot done, but after a few hours I was mentally exhausted.

What works for me is mixing tasks that need different kinds of attention. A substantial feature where I think deeply and review code. Deployment or configuration work where the agent guides me through the steps. Something lighter and visual. While one agent works, I move to the next task.

AI agents killed my flow state. I've found a way to get it back. Mixing tasks that need different kinds of attention and letting agents hold the details helped me get back into the zone. Thread ↓

Before AI, one of my favorite parts of programming was getting into the zone. Good coffee, some background music, and hours inside one problem. Agents broke that rhythm. I'd prompt one, check another, review some code, then switch to a different feature.

Vlad Haltakov reposted

I know how to code, but how do I teach my son? AI is changing the rules of programming, so I’m trying two things with my 8-year-old.

I know how to code, but how do I teach my son? AI is changing the rules of programming, so I’m trying two things with my 8-year-old.

We sit down together and use AI to vibe-code his little personal website. While the AI is doing its thing, we switch over to Scratch and build a simple game ourselves, step by step.

I want him to get comfortable building with AI, but also learn the basics so he understands how things actually work. I don’t know if this is the right approach. How are you teaching your kids?

Vlad Haltakov reposted

Sometimes I do some complex shell command myself instead of asking my agent... just to see if I still got it. I still do :)

Goodbye Lexware 👋 I only needed the basic plan, but you forced me onto the most expensive one just to create invoices in English. So I vibe-coded my own invoicing app instead. Subscription cancelled. I’m free now 😄

Sometimes I do some complex shell command myself instead of asking my agent... just to see if I still got it. I still do :)

Vlad Haltakov reposted

Plugin extensions in ChatGPT are pretty cool. You can go from discussing an idea to working on it right inside the conversation. Less copying between tabs, less explaining the same thing again. Having the tools where the context already is makes so much sense.

Plugin extensions in ChatGPT are pretty cool. You can go from discussing an idea to working on it right inside the conversation. Less copying between tabs, less explaining the same thing again. Having the tools where the context already is makes so much sense.

Vlad Haltakov reposted

I’ve got OpenClaw running on my VPS, a Grok bot, and I’m waiting for Muse in the EU. But right now, I’m mostly just excited to see what OpenAI releases today. Feels like the AI assistant space is about to get a lot more interesting.

Vlad Haltakov reposted

You won’t read the code. One thing holding developers back with AI is still believing they’ll eventually read it all. Let that go.

I’ve got OpenClaw running on my VPS, a Grok bot, and I’m waiting for Muse in the EU. But right now, I’m mostly just excited to see what OpenAI releases today. Feels like the AI assistant space is about to get a lot more interesting.

You won’t read the code. One thing holding developers back with AI is still believing they’ll eventually read it all. Let that go.

Vlad Haltakov reposted

This was my office the other day. I was on vacation with my family, drinking coffee in the mountains, when a SimplePost user hit an error.

This was my office the other day. I was on vacation with my family, drinking coffee in the mountains, when a SimplePost user hit an error.

View full thread

Then the user got a free month, and Codex drafted the explanation email in Gmail for me to review and send.

This is what modern customer support looks like: I get to enjoy my vacation, and the user gets a fix minutes after something breaks.

This was my office the other day. I was on vacation with my family, drinking coffee in the mountains, when a SimplePost user hit an error.

I got a Telegram alert. Codex checked the logs, found the issue, implemented a fix, and explained the change. I reviewed and approved it.

Vlad Haltakov reposted

On the school run to pick up my son, testing ChatGPT Live. Now I can use SimplePost in the car to check my social media schedule, move things around, or schedule new posts - all with voice. https://x.com/OpenAI/status/2102808325742322002

Vlad Haltakov reposted

Stop rebuilding social media integrations. Every platform has different text limits, media requirements, upload flows and edge cases.

Vlad Haltakov reposted

Check out my submissions for the WebMCP challenge. All projects are live so you can try WebMCP in each one directly. Make sure you use the ChatGPT internal browser or enable WebMCP in Chrome https://webmcp.devpost.com/submissions/search?utf8=%E2%9C%93&terms=haltakov&sort=recent

On the school run to pick up my son, testing ChatGPT Live. Now I can use SimplePost in the car to check my social media schedule, move things around, or schedule new posts - all with voice. https://x.com/OpenAI/status/2102808325742322002

Vlad Haltakov reposted

Today GPT-6 Astra fixed a UI I couldn’t get right with Opus 5.5. That sentence would’ve sounded ridiculous a few months ago. What a time to be alive.

Stop rebuilding social media integrations. Every platform has different text limits, media requirements, upload flows and edge cases.

View full thread

With MCP, your AI agent can adapt the content for each platform while SimplePost handles validation, media adaptation and publishing.

The SDK, MCP server, CLI, and scheduling web app are all open source under a permissive MIT license. https://github.com/simple-post/core

Stop rebuilding social media integrations. Every platform has different text limits, media requirements, upload flows and edge cases.

SimplePost validates posts against each platform before publishing, converts images when needed, and keeps the integration logic in one place as APIs and requirements change.

Check out my submissions for the WebMCP challenge. All projects are live so you can try WebMCP in each one directly. Make sure you use the ChatGPT internal browser or enable WebMCP in Chrome https://webmcp.devpost.com/submissions/search?utf8=%E2%9C%93&terms=haltakov&sort=recent

Today GPT-6 Astra fixed a UI I couldn’t get right with Opus 5.5. That sentence would’ve sounded ridiculous a few months ago. What a time to be alive.

Vlad Haltakov reposted

You’re integrating AI backwards. Instead of putting another AI assistant inside your app, give the assistant your customers already use access to your tools. When someone already pays for ChatGPT or Claude, why charge them for another writing assistant?

You’re integrating AI backwards. Instead of putting another AI assistant inside your app, give the assistant your customers already use access to your tools. When someone already pays for ChatGPT or Claude, why charge them for another writing assistant?

View full thread

That’s how I’m building SimplePost. Your assistant does the writing and reasoning; SimplePost provides the tools to preview, validate, schedule and publish. The chat where you plan a launch becomes the place you prepare its announcement, without copying everything into another tool.

I focus on making publishing work instead of reselling model access, which keeps SimplePost inexpensive. Your AI subscription’s limits still apply, but you’re not paying me for AI a second time. https://simplepost.social

You’re integrating AI backwards. Instead of putting another AI assistant inside your app, give the assistant your customers already use access to your tools. When someone already pays for ChatGPT or Claude, why charge them for another writing assistant?

Embedding AI means paying for inference, then charging more, limiting usage, or choosing cheaper models to protect your margins. Let customers use their existing subscription instead and keep the context from the conversation they’re already having.