Jarosław Michalik

Steal this AI coding workflow 👇 open.substack.com/pub/theimpl/p/…

The paradigm shift. You're leveraging your strategic thinking. AI is leveraging its ability to execute.

If this hit home, Like this thread to boost its reach. Share it with a founder fighting the same battle. And if you're ready to actually solve these problems instead of just putting out fires, DM me "SCALED". Let's talk about building a truly robust app.

Founders get caught thinking "more users = more revenue = problem solved." Wrong. More users on an unscalable no-code/AI backend means slower load times, dropped sessions, and angry customers. Revenue without performance is just a bigger headache.

Your vibe coded app will hit a performance ceiling. It's not a question of IF, but WHEN. That project vibed over the weekend can quickly become a bottleneck as users scale. It's inevitable. Here are 5 simple strategies to DELAY and MINIMIZE the damage 👇

You'll get the best feedback from your users. But what if you don’t have any? (or you’re too scared to talk to them)? > Record a screencast of your app > Upload it to Gemini 2.5 (via AI Studio) > Add a prompt: “act as an unhinged user, explain what you see in the app and write a critique” Optionally, you can include a short description of your ICP (in my case, personal trainer/fitness studio owner). I deliberately don’t add more context about what the app is or what it’s supposed to do. The result? A cold bucket of water on my UX. Gemini said (quote): “This is a CATASTROPHIC FAILURE! My grandma could code a more stable app on a Nokia 3310” It hurt. But it also noticed that the action buttons were hidden. It pointed out that it’s unclear which client is assigned to which trainer. Valuable insight from a studio owner’s perspective. Raw feedback from a fictional user who doesn’t give a damn about my feelings. If you need quick UX feedback for your SaaS—give it a try. It’s worth it.

When I start working on a new feature... > fire up my assistant (Claude, in my case) > prompt: “act as an event storming coach (...) ask me questions (...) create a process map in Mermaid” > discuss, iterate > you get a process map - it goes into docs and provides context for Cursor > Profit Why is this awesome? First off, it saves a ton of time. Instead of spending hours in Miro wrestling with finding the right building blocks, you get a draft, cohesive process map in just a few minutes. You’ve got a solid starting point for further work. You can quickly validate it with the team. It’s also a great way to document your assumptions and ensure everyone’s on the same page about the process. But… It’s just a supporting tool. The AI assistant gives you a strong foundation, but it won’t replace your intuition. Always: - verify - discuss - refine Designing a meaningful flow takes more time than you think, especially with more complex processes. It’s still an iterative process.

Building EngageUp.io day 37. User acquisition path (experiment): - I’m promoting the "roast your reel" page - user gets a few metrics and basic analysis - "want more? Create an account!" After creating an account, all metrics are visible + the ability to add your brand context to the analysis. I launched Meta campaign targeting Baltic countries. What will the CAC be? I guess we'll find out in the coming days.

@acm0x It’s amazing how „human” psychology applies to LLMs

@JarekOnTech This reminds me of a psychological trick when your brain completely ignores "not" in whatever you say. E.g. "don't do" Vs "avoid" have very different perception

LLMs have STRONG confirmation bias. When asked "is this good?" they start looking for reasons why something IS GOOD. But if you ask... "Find ALL problems in this code. Be RUTHLESS." Suddenly the same model, in the same conversation, will spit out a list of 15 potential bugs, inefficiencies, and holes that it previously "forgot" to mention. Test from yesterday. Changes in my TrainingPro app. • Prompt 1: "Is this code good?" → "Yes, it's well-written, readable, and..., comments could be added" • Prompt 2: "Point out all the errors" → "Here we're making unnecessary database queries..." THE SAME CODE. Two completely different results. Models respond according to the context we give them. Are you looking for a pat on the back, or are you looking for problems? You'll get exactly what you ask for. And if you really care about solid code, don't just trust machines. Sometimes you need a human eye to see that your beautifully generated app will break in production

Side Project Lifecycle: Day 1 - This will be the next Facebook! 👉 Just finished your fifth rewatch of The Social Network. 👉 Buy a domain (so cool, so cheap!) 👉 Plan to quit your job 👉 Check how to set up a company in Cyprus Excitement lvl 10^6. Day 7 - "Maybe I should do MVP..." 👉 Struggled with the architecture 👉 0 people on the waitlist ("must be perfect before I share!") 👉 Add one more feature Feeling both SUPER productive and not. Day 14 - Crisis. 👉 Stack is "a bit" overkill 👉 Kubernetes was too much 👉 3 users (friends logged in once) But the CI/CD pipeline works perfectly. Day 30 - "I'll finish it someday" 👉 Project sits in the "my_app" folder 👉 "I'm busy now" 👉 Consider going no-code Feel free to loop back to Day 1 and buy a new domain for a new project. A year later. 👉 Domain renewal notification arrives 👉 "Oh, that was fun, I'll dig it up!" 👉 "But first, I'll rewrite it in new NextJS" 👉 "The old code was just a prototype anyway"

A data science agent popped up in Google Colab. 👉 Upload your CSV 👉 Agent reads the file 👉 Explores data 👉 Conducts analysis 👉 Creates reports with charts All within Jupyter Notebook—view and tweak every step. Accelerated video: March Stripe data analysis took about 3 minutes. Want to try it? Click the "Gemini" icon in Google Colab and LFG. Enjoy.

thank you for nothing

Coding with @cursor_ai feels like magic.