Vibe of the code is too good

Stay happy with your current subscription and whatever workflow works best for you. Avoid the hype of running countless agents. They have the resources and revenue for it, but you don't need that to be successful..

Once interviewed at a startup building digital addresses so people wouldn’t have to share their real ones. They kept saying how private and secure it was, so I asked a few questions about how it actually worked. I answered every question they asked me, but when I questioned their product, the interviewer got pissed off. Funny how that works.

How the heck one supposed to review this

I’m happy with my Pi thanks

Should’ve launched this when the model was actually GA. You’ve done the “tease now, maybe release later” thing with previous Pro models, and some of them we never even got to try. What am I supposed to do with benchmark charts for a model I can’t actually use?

AI made writing code cheap. It didn't make knowing what the code should do cheap. That part was always the actual job.

I’ve been vibe coding apps and websites that I would actually use myself. Usually, the first version is full of broken things. This time I changed the process. Instead of throwing a prompt directly at the coding agent, I first used ChatGPT to write a proper design spec. Then I had Claude Opus 5 implement the frontend. Once I was happy with the direction, I spent around 50 minutes with ChatGPT Web defining exactly how I wanted the app to be built: • project structure • component boundaries • where files should live • how APIs should be called • linting rules • quality gates • verification before marking work complete • mandatory TDD • independent code review And one more important rule: After all of that passed, a separate QA agent had to open the app in a real browser and verify the feature that was just implemented. The difference in quality was ridiculous. The app worked. Roughly 90% of the functionality worked as expected on the first proper run. The project structure matched what I had specified. Components were where they were supposed to be. The architecture, tests, verification and QA process were actually followed. And here’s the interesting part. The frontend was built with Opus 5, but most of the rest was Composer 2.5 Fast. I burned through both my X Premium+ and Cursor Start limits 😂 Composer got roughly 70% of the project done. The remaining work was finished with DeepSeek V4.1 Flash through OpenCode Go. The biggest lesson for me: The model matters, but the process around the model matters a lot more. Give an agent a vague prompt and you get vibe-coded software. Give it a spec, architecture, tests, quality gates, code review and browser-based QA, and suddenly it starts looking a lot more like software engineering.

That clever email validation regex copied from a forum can pin a CPU core on a single request. Nested quantifiers like (a+)+ backtrack exponentially on a long string that almost matches. Every unit test uses valid or obviously invalid emails, so it never shows up. Cap input length before the regex runs. Then fuzz it with 50k characters.

Green CI means the tests you wrote still pass. It says nothing about the flows nobody thought to write a test for. And that's usually where the bug report comes from.

it’s useless for all kind of projects

A unique index on (email, deleted_at) does not stop duplicate active users in Postgres. NULL is never equal to NULL, so two rows with the same email and deleted_at = NULL both pass the constraint. Use a partial unique index on email WHERE deleted_at IS NULL, or NULLS NOT DISTINCT on PG15+.

After using claude 5x I am tempted to upgrade to 20x but been hearing its not actually 20x in true 20xPro. Should I get another 5x or get 20x?

I’ve a feeling that Anthropic is going to cut the Opus 5.5 limits and would push you to use Sonnet 5.5 more often.

I think Opus 5.5 is the model we’ve been scared of.

Search built on LIKE '%term%' feels instant in staging with 5k rows. A leading wildcard means a normal btree index can't be used, so production does a full scan on every keystroke. Trigram index or real full-text search. Decide before the table grows, not after.

Don’t install anything from Meta. Absolutely anything avoid them as you would avoid E-Corp

Generating tests that click through an app is the easy part. The hard part is knowing what the page was supposed to do. A test with no real expectation only proves the app didn't crash, and that bar is lower than most people think their suite is clearing.

"SELECT to check the coupon is unused, then UPDATE to mark it used" works perfectly in every test. Send the same request twice in parallel and both pass the check before either writes. Coupon redeemed twice. The check and the write have to be one statement, or one lock.

Rate limiting by client IP behind a load balancer is a trap in both directions. Read the socket address and every user shares one IP, the balancer's. Trust X-Forwarded-For blindly and an attacker just sends a new value per request. Only take the hop your own proxy appended.