Rok Benko

I went from being rejected 500x on the job market in 2024 to my first 10 paying customers in 2025. Here's my story and what I learned. 👇 In 2022 I made my first OpenAI API call. Not because I saw the future of AI in 2025. Not because I had a strategy. I was just amazed. So I started sharing knowledge on Stack Overflow. For the next 3 years I lived in the AI community. I answered questions 24/7/365. I remember one New Year's Eve hitting the Enter button to submit an answer on Stack Overflow before running out the door to a party. What happened next? • Became the #1 OpenAI API contributor on Stack Overflow. • Became the #1 CrewAI contributor on the official forum. • Reached 2.4 million developers just on Stack Overflow. • Opened a YouTube channel and put in 40+ hours for one coding tutorial. • Open-sourced 27 AI projects, gaining 300+ GitHub stars. • Got feedback my answers were better than official docs. Unfortunately, none of that landed me a DevRel job. In 2024 I was rejected 500x! DevRel? No. AI Engineer? No. Anything in between? No. Finally, I got a role at a startup. After 12 months of searching. Guess what? I was fired after 1.5 months! That CRUSHED me! No exaggeration, I was sitting on a bed and crying. Eventually, after months, I came to a realization. • "Startups are risky." • "The fact is that 99% of startups fail." But here’s the thing… • Can I fail more than being rejected 500x? • Can I risk more than job hunting for 12 months only to be fired in 1.5 months? So in 2025, I went all in: Solopreneur, building in public. For the first time in my life, I saw a Stripe graph that wasn’t flat. Before that? I never shipped ANYTHING in production! Not a single landing page was deployed. Nothing was ever good enough. The term "MVP" was not in my vocabulary. Today, I still consider myself a solopreneur rookie, but... I think the chances of succeeding as a solopreneur have NEVER been higher in the history of the world! Startups are probably still carrying the same amount of risk. But iteration cycles and costs are RIDICULOUSLY low. To put it simply: Your startups will still fail (until you find a gold mine). But today, you’ll know in 1 month for $100. Just 3 years ago, it would take 1 year and $10,000. THINK ABOUT THIS! Here's what I learned on my way to the first 10 paying customers: ✅ Pick a real-world pain you can solve 10x faster and/or cheaper with AI. ✅ Test your startup idea with an MVP that can be built in 1 week. ✅ Maximize AI tools to cut time and cost while you work. ✅ Validate your idea on the market, not among FFF (family, friends, fools). ✅ Make decisions based on data, not your emotions. 🔄 Repeat! Because you're not that 1% that will hit a gold mine with your first startup.

@boltdotnew @Lovable @anything Done, LFG!

@rokbenko @Lovable @anything Let's do it 😎 shoot us a DM

Hey, @Lovable, @boltdotnew, @anything! 👋 I used to work at an agency. Simple website? Not under $1.5K! Simple e-commerce? Not under $3.5K! Recently, I met up with a childhood friend who wanted a website for his business. The reason he still doesn't have one? He kept getting quoted crazy amounts for a pretty simple website! As someone who's been vibecoding since the GPT-3 days in 2022, I knew I could build it in about an hour. But then I thought: How do you charge a friend $750 for one hour of work? And more importantly… What happens after launch? Because every developer knows the classic: “Can we make one quick change?” So instead of building it FOR him, I decided to build it WITH him using a no-code AI platform like @Lovable, @boltdotnew, or @anything. That way he gets: • Website. • Ability to edit it himself. • Introduction to the world of vibecoding. I'm genuinely convinced this is going to open his eyes. He already has ideas for apps and side projects. So here's my question: Would any of you be willing to sponsor some credits for him? You'd basically be helping onboard a future long-term customer. Also, I'll be sharing our whole build journey publicly tomorrow. Anyone in? @antonosika @FabianHedin @ElenaVerna @EricSimons @iamalbertpai @GarrettServ @dhruvamin @marcus_lowe @zariazinn @leojrr

The easier something looks, the more mastery is probably behind it. 👇 A few days ago, I filmed and edited my first-ever aftermovie for an event. It was for Bevog Brewery in Bad Radkersburg, Austria. You might watch a short video and think: “You just record clips and edit them together.” Yeah... Not even close. I did record a few GoPro videos with friends before. But this was different: • First time using a proper Sony camera. • First time filming strangers. • First time filming a real event. • First time shooting across 2 days. • First time trying to tell a story through footage. The day before the event, I already went there to scout the location. Where should I stand? Where will people gather? Which spots will look crowded? Which shots could look cinematic? Where will the sunrise and sunset create the best lighting? Then came the actual event day. IPA Day. Sweating the whole time. Kneeling on the floor. Getting my knees dirty. Trying to capture the SAME moment from 5 different angles. Because once you start editing, you realize how much footage gets thrown away. Probably 25% instantly unusable: • Bad focus. • Awkward movement. • Someone walks into frame. • Bad lighting. • Shaky footage. Then from the remaining clips, maybe 50% never even make the final video. And even after all that, you still need to: • Sync clips to music. • Create transitions. • Build pacing. • Fix colors. The raw footage looked super orange/red. Without color grading, the whole thing looked amateur. But if you only watch the final video, it looks “easy.” That’s the point. The better someone is at something, the easier they make it look from the outside. Until you try doing it yourself, you have NO IDEA how much effort is hidden underneath. I’ll never watch aftermovies the same way again. The next time I look at something and think: “Easy money.” I’ll remember this day. PS: X seems to distort the video. The original looks fine.

Working on something exciting... 👀

It took me 1 year to realize I was sending cold outreach messages WRONG. Here's what I screwed up and what to do instead. 👇 Why do we reach out to people? To sell our product or service, right? Yes, eventually... But NOT in the first message. In the past year I reached out to 100s of people for my 7 failed startups. The playbook was always the same: try to sell in message #1. I never saw outreach as a conversation. More like a one-message-or-nothing process. My logic? "If I don't sell in the first message, they don't need it." How did I figure out I was wrong? I copy-pasted my first message into Claude and asked it to rate my outreach and build me a cold DM playbook. Long story short? My message was too much of a pitch right off the bat. Which makes 101% sense since this is what I thought was the name of the game. Claude flipped the script. Cold outreach should be a CONVERSATION between me and my potential customer, where: • Phase 1: Permission (messages 1-2) → Goal: Get a one-word reply. • Phase 2: Pitch (messages 3-5, after they reply "yes" / "sure" / "tell me more") → Goal: Demonstrate value and create curiosity to book. • Phase 3: Meeting (messages 6+) → Goal: Book a(n) (online) call. Breaking cold outreach into a sequence of messages creates a curiosity gap. Which I already figured out plays a HUGE role in bumping up conversion rates. Let's see how this affects my reply rate.

Hey, @Lovable, @boltdotnew, @Replit, @anything, @rork, @vibecodeapp_, @cursor_ai, @windsurf, @Base44, @emergentlabs! 👋 I've been coding with LLMs since GPT-3 in 2022. First in line for every vibecoding tool. Last year I shipped 7 startups. All vibecoded. All failed. Here's the gap I see in your space... 99% of your content is about how to build. Better prompts. Better stacks. Better workflows. But building is 50% of the story. What kills most startups today is distribution. And your users figure that part out alone. They glue 10+ AI tools together to find a niche, get leads, write cold emails, and run ads. Your platform is one tool in a big stack. But your content pretends building is the whole game. Plenty of YouTubers show the path to MRR. No vibecoding platform does. Why would they? Nobody wants to tell users "you also need Claude Cowork for outreach on top of our tool." I want to do something like that with you. Whoever's in. The idea: 5 builders a month. We film their path from empty repo to first MRR. Daily or weekly. Not just the building. Also, how they pick a niche, find leads, do outreach, and test marketing. The wins and the flops. DM me if you're in. PS: I'll handle production. Years in content and AI communities (reached 3 million people). I shoot on Sony and edit in Premiere Pro. You cover tokens for the builders and a prize for whoever hits the highest MRR. @antonosika @FabianHedin @ElenaVerna @EricSimons @iamalbertpai @GarrettServ @amasad @dhruvamin @marcus_lowe @rileybrown @anshnanda @mntruell @sualehasif996 @ArVID220u @amanrsanger @mukundjha @madhavjha @maorshlomo

Do you think that mostly depends on the LLM being used? The stricter SCOPE.md constraints are partly there because weaker models tend to break layouts or deployments once you allow broader structural edits. My sense is that more capable models, like GPT 5.5 or Opus 4.7, can handle larger refactors much more reliably.

@rokbenko the SCOPE.md restriction to copy, button styles, and above-fold reorder is the right call. structural changes are where these tools usually go off the rails

I open-sourced an agent that continuously improves your landing page conversion rate. Every time you build a landing page, you wonder: • “Should this CTA be changed?” • “Is this layout confusing for visitors?” • “Does this section belong higher on the page?” Answering this = manual A/B tests + guesswork. This agent automates the loop: 1. Reads your landing page code. 2. Generates a hypothesis. 3. Rewrites the UI. 4. Pushes to GitHub. 5. Waits for real visitors. 6. Measures the impact via PostHog or Plausible data. 7. Keeps the change if it worked or reverts it if it didn't. Then repeats forever. Think of Karpathy's autoresearch but for CRO. SCOPE.md is the system prompt for the AI. It tells the LLM exactly what it's allowed to change and what's off limits. You can loosen or tighten these SCOPE.md rules to match your comfort level. The default SCOPE.md is intentionally restrictive. The agent can only change copy, tweak button styles, and reorder above-the-fold elements but cannot touch navigation, pricing, auth flows, state management, API calls, or anything structural. github.com/rokbenko/CROto…

Let’s test cold outreach on Facebook this round. 📨

I sourced 1.2M leads, enriched with email, with Claude Cowork in ≈2 hours. If this post gets 100 comments in the next 48 hours, I'm going to reveal the playbook with the prompt that you can copy-paste into Claude Cowork COMPLETELY for free! Comment "Cowork" below. Share with your friends so we can hit the goal. 👇

In the past month, I locked myself in. Building like a maniac with Claude Code. Here’s what happened to my MRR. 👇 I subscribed to Claude Max 5x in March. I thought I’d run out of tokens fast. I didn’t. I couldn’t even max it out. Best I did: A few 100% session limit hits, 44% of the weekly limit. So, what did I build? • 1 full app • 1 MVP app • 1 npm package • 1 landing page • 4 websites (1 personal + 3 to sell) What happened with my MRR was... ...exactly what I expected. Nothing! If I hadn’t failed 7 startups this past year, I’d be losing my mind right now. Thinking my MRR would hit 5 figures. No one gives a fu€k about your startup. Not until you leave the vibecoding rabbit hole. So this time, I expected this. Building is not the full story. Last 30 days: Building. Building. Building. Next 30 days: Selling. Selling. Selling. Code doesn’t pay. Customers do. LFG!

@Cloudflare just released a free tool to check if your website is ready for AI agents. 🚀 PS: Find the link in the comment below.