Denis Perov

AI-augmented SWE eager for challenges

[Denis Perov @denismajus]

Ever wonder what HTTP stands for? Or can you list all the network protocol layers? How about naming some big ECMAScript updates?

These are just a few interview questions folks shared with me after my last post about LLMs in code reviews.

But one story really stuck with me.

"I had a take-home task from a company. I am more comfortable with TypeScript, but went with Python because they seemed to use it a lot. They penalized my application for not using TypeScript."

This guy's been in IT for 20 years. If he's facing this at Senior Lead level, it must be way tougher for Juniors.

He got the vibe right. He showed he could adapt. Still lost out.

Here's what's messed up with hiring right now:

1. Interview tasks that set you up for failure based on the very answers they guide you to.
2. Processes meant for new grads but used on folks who've launched real systems across three tech eras.
3. Trivia that checks your memory, not your skills. Just knowing what HTTP stands for doesn't mean you can design a solid API.
4. Feedback that totally contradicts the requirements they gave you at the start.

Plus, the whole setup is flawed. Candidates are stressed out. Interviewers have all the control. Even a senior engineer who confidently ships stuff daily can choke on a question they know the answer to. The format tests how you handle stress, not your actual skills.

Here's what those companies just don't get:

Interview benchmarks often reflect the interviewer's own learning curve from way back, not what the job actually needs now. What they struggled with a decade ago ends up being the gate every candidate must pass through. It's not about testing skills. It's a mirror.

Developers who can adapt to whatever tools the team uses are showing exactly why senior engineers get hired. Seeing the situation and making the right call? That's not a flaw. That's the job.

The right company will:

1/ Look at what you've built, not just whether you picked the right answer under stress.
2/ See adaptability as a strength, not a sign you're lacking skills.
3/ Know the difference between "this isn't how I'd do it" and "this is wrong."

If a company's process punishes adaptability, imagine the chaos in their codebase when requirements change.

The interview is broken. Not the developer.

What's your story? Still collecting them.

[Denis Perov @denismajus]

I went through this remote tech interview last month.

The task was small graph problem in TypeScript. Used Claude Code. It worked. Passed all tests. Names clean, helpers small, README explained approach in three paragraphs.

Still didn't get the job.

The feedback was honest: "Solid code, but not how I would've done it. Recursive approach when iteration would've been cleaner. Names are too generic. We wanted someone who puts more personality in the code."

That line's been on my mind for weeks.

I wasn't rejected because my code was wrong. It was because the senior reviewing it had an ego.

Pattern I keep seeing everywhere now:

We tend to overestimate the quality of our own code compared to what an LLM churns out.

Hard to notice because no one shares bad code. We post the clever refactor, the rewrite that survived production. We don't share the function that flopped four times last Tuesday.

Interviews are where the truth seeps out. I've seen take-home assignments solved cleaner by an LLM than by the senior reviewing them. And instead of owning it, the senior picks some stylistic issue and moves on.

That's not technical critique. That's an EGO.

Think about the math:

- You've got ten years of experience and your gut instincts are based on those ten years. That's your context window.

- The LLM has read everything: popular blog posts, Stack Overflow answers, every well-documented open-source project from the last decade. When you and the model clash on style, your "experience" is going up against the entire posted history of human programming.

You'll lose more often than your ego is ready to admit.

Three patterns I keep seeing in human code:

1. Personal formatting the linter doesn't even need.
2. Choosing one technical approach over another just because of preference.
3. Over-engineering for "future cases" that aren't in the requirements.

Never seen an LLM do any of these. It does exactly what you ask. Nothing more, nothing less. Doesn't push for its favorite paradigm. Won't insist on some abstraction that makes future-you feel clever in a review.

In your career you've probably written same code three different ways. Each time it was "the right way." The LLM doesn't care which version it produces.

LLM just produces. Cleanly and cheaply.

The LLM nailed it. In ten percent of the time. The senior who turned me down probably couldn't crank out code that quality without a long grind and ten cups of coffee.

If the LLM is doing it cheaper, cleaner, faster, and good enough for the customer.

What value are you really adding? This question freaked me out when I first asked it myself.

Ever turned someone down because of "style"? If you came across that same code today, would you still say no?

[Denis Perov @denismajus]

I noticed this context window breakdown appared in Claude Code Web a few weeks ago, but it was gone as quickly as it appeared.

Would really like to get it back @claudeai

[Denis Perov @denismajus]

This looks helpful, need to try it.

[AfterQuery @AfterQuery]

Introducing IDE-Bench!

A multi-language, full-stack benchmark evaluating LLMs acting as autonomous IDE agents

IDE-Bench assesses agents' ability to navigate, reason, and modify complex repositories using the same tools available in modern AI-native IDEs like Cursor

Models tested from @AnthropicAI, @OpenAI, @Alibaba_Qwen, @GoogleDeepMind, @xai, @deepseek_ai, @Meta, and @cohere

Check out the full results at https://t.co/MdqB7SL5JY!

[Denis Perov @denismajus]

@uttam_singhk Diversity is good for the economy

[Denis Perov @denismajus]

This went largerly unnoticed, but it's huge.

Antrophic is raising the limits to their pre-cutting levels 🍾

[Claude @claudeai]

Effective today, we are:

1) Doubling Claude Code’s 5-hour rate limits for Pro, Max, and Team plans;

2) Removing the peak hours limit reduction on Claude Code for Pro and Max plans; and

3) Substantially raising our API rate limits for Opus models.

[Denis Perov @denismajus]

Giving out more gift codes for a free week of Claude Code.

Share what you're building.

[Denis Perov @denismajus]

So inspiring to see a first Claude Code subscription experience 😊

[Denis Perov @denismajus]

$18 in fees for bridging ¢10 via @debridge?
Please, tell me that's just a bug 🫣

[Replit ⠕ @Replit]

Join us for a 24-hour buildathon tomorrow at: https://t.co/obDpoRayXF

[Denis Perov @denismajus]

@SeekerClaw @nvidia @solanamobile @solana @toly Is Minimax M2.7 the best for agentic tasks? What other would you recommend to try?
Also, they are deprecating many of them soon 🫤

[Denis Perov @denismajus]

@agentcashdev Didn't work well with @NanoClaw_AI out of the box 🥲 Because they don't keep anything outside whitelisted folders.

Is there any way to restore it or should I just say 'buy' to my bonus? 😮‍💨

[Denis Perov @denismajus]

Almost forgot what a proper E-mail looks like until I got this from @jonathanfishner @ OneCLI. Short, direct, plain text.

I actually enjoyed reading this one.

Everyone should be doing this in 2026+ IMO 📧✊

[Denis Perov @denismajus]

@agentcashdev offers up to $25 for your AI agents + a great interactive onboarding flow.

Their screening agent will analyse your profile in real-time and assign 0-1000 score. Mine was assigned ~600 points.

Sign up and see what yours gets 🤑 https://t.co/pUITJ9gywQ

[Denis Perov @denismajus]

@iamknownasfesal @encrypt_xyz Excited to try all that out 💪☕🚀

[Denis Perov @denismajus]

@d3h3d_ @colosseum @ikadotxyz @encrypt_xyz @iamknownasfesal Noticed anything missing on the diagram? (a) payment facilitator (b) access keys storare. That's the "better".

@iamknownasfesal could you DM me please?

[Denis Perov @denismajus]

Everyone is crazy about new Claude Code AI buddies, except me. Mine looks odd, smells funny and doesn't seem to be very helpful, tbh 😐 #claudecodebuddy

[Denis Perov @denismajus]

I minted an amazing @poapxyz!
https://t.co/xALn4iJELo

[Denis Perov @denismajus]

We built reputation bridge for HackMoney by @ETHGlobal 🌉
Waiting to see if judges want to cross it: https://t.co/v27MYNKQNH