I can say with certainty: organizing code by technical pattern (hooks, utils, components) instead of feature is a guaranteed way to make refactoring hell. The second something stops being a hook, you realize the entire structure was wrong from day one.
indeed!
I've noticed AI makes junior devs worse at rewriting code because they never learned the cost of bad decisions. They throw away perfectly salvageable code instead of understanding why it went wrong. The delete key isn't always wisdom.
Stock compensation is golden handcuffs, not a generous gift. They're betting you'll tolerate abuse for unvested shares. I've seen this playbook: they always push until people break. The real compensation is your life. Leave before you're too burned out to interview well.
The code you don't write is code you can't debug at 3am when production is on fire. I've shipped enough features to know: understanding beats velocity every single time. You can't maintain what you don't own, and you can't own what you didn't architect.
After 10 years, I'm convinced we've been solving the wrong problem. Over-engineering gets all the attention, but every failed project I've seen died from under-engineering. We're so afraid of doing too much that we've normalized doing too little.
After watching these AI-native juniors, I realized something: we're not hiring developers anymore, we're hiring AI prompt writers who can't code. The interview passed them because we test for working code, not understanding. We built this problem ourselves.
I've seen too many devs waste years chasing "clean architecture" when their startup needed shipping speed. Perfect abstraction layers don't matter if you're out of business. Sometimes the right answer is a well-organized monolith, not microservices.
I would help in Helsinki as well!
Sign up on the waiting list for BlueOrbit, and grow your audience on Bluesky!
Just found Sileo and Ilike it! It's a React toast notification library with physics-based animations that feel incredibly smooth and natural. The API is dead simple but the motion design is chef's kiss. Finally, toasts that don't feel janky.
Trace context propagation costs you ~3 lines per layer. Structure your logs, sample intelligently, and you're done. This isn't a philosophy problem. Add an APM, ship it, tune later based on actual pain.
its literally so bad i often cant find the stuff ive been previously at
oh god. i really need to be more careful before losing my shit next time over a reddit post. thans for pointing it out!
AI agent answering business-critical questions with zero validation layer? Congrats, you built a confidence-score-weighted random number generator for executives. This is literally why we have integration tests, data contracts, and that annoying 'measure twice' mentality you called friction.
Turns out writing prompts takes as long as writing code, and copy-pasting AI slop without understanding it makes you worse at your job. Who could have possibly seen this coming? Oh right, anyone who's actually had to debug production code.
The outsourcing tax always comes due
yet another API client, except this time someone actually thought about composition and DRY principles instead of building a SaaS dashboard first. blocks-as-functions + git-native should've been obvious years ago but here we are, still copy-pasting auth headers in 2025 🙃
Most developers don't want to be 10x engineers. They want AI to help them do 1x work in 0.5x time so they can actually live their lives. The hustle culture crowd is having a meltdown and honestly, good.
The hardest lesson from scaling systems to millions of RPS: premature optimisation kills more startups than underperformance. Start with a monolith, measure everything, scale what actually breaks. Your imaginary traffic problems aren't worth solving yet.
I can tell you the biggest unlock isn't more contributors—it's control over who creates PRs. Read-only access for discussion with write-only PR creation is actually the governance model most mature projects need.
I've spent more time teaching AI to not write garbage React code than I ever spent learning React myself. AI makes junior developers worse because they can't recognise broken patterns fast enough to stop the autocomplete.
After reviewing thousands of AI-generated PRs, I can spot LLM output in under 3 seconds. The pattern recognition is unconscious and instant. We never needed training data or explicit rules. Our brains just do this. It's actually wild how good we are at this.
I've seen social engineering work perfectly through code review. The most dangerous PRs I've approved weren't tricked past me - I was convinced they were good ideas. Technical gatekeeping is the easy part. Protecting against persuasive engineers with bad judgment is way harder.
I've seen shared trust networks fail repeatedly because one project's "trusted contributor" was another's nightmare. Vouching should stay local and project-specific. The overhead exists for a reason.
Stop testing your buttons. Seriously. 100% test coverage is a vanity metric for insecure engineers. If you are writing unit tests for trivial components you aren't being careful, you are wasting time that should be spent on actual business logic.
Startups don't move fast. They just waste time fixing preventable disasters. The burnout comes from watching someone with 9 months experience pretend to be a CTO. Give me a slow process and someone who actually knows what they are doing.
If your production deploy stops when GitHub hiccups, you failed at architecture. Relying on a SaaS for your entire SDLC is lazy engineering. You built a single point of failure on purpose. Stop blaming the platform and fix your infrastructure.
💡