I work in fintech and I still review code to some extent. If PR doesn't touch payments or sensitive areas then it's mostly LLMs doing reviews. Otherwise LLMs do a first few passes in a loop and I look over it before merging. Maybe that will change in future, I just can't ship a PR knowing it deals with payment processing, cardholder data, etc without at least skimming through the PR. I certainly don't read every single line of code though.
@naicodes@PovilasKorop I respectfully disagree. Judgement and telling good from bad, noticing AI going down the wrong path and steering it in the right direction is something you'd have if you learned how to code and had experienced production issues and have had to make decisions based on trade offs.
I use AI for 99% of my code now and I'm not going back. But I'm seeing a lot of posts saying you don't need to read the code at all, and I don't think that scales or applies to everything.
For apps with few users, no PII, payments, or health data its probably fine. I vibe code all kinds of stuff without reading a single line. But that changes once a bug can cost someone money, their data, or their health.
My review effort scales with blast radius. Automate everything, then put human eyes where mistakes have real consequences. Does that mean I review every line? No. It means guardrails, deterministic checks, meaningful tests, and spot checks on PRs that touch PII, payments, security, and compliance.
I'll go in more detail in the new series I'm working on.
Crazy to think about that even after 5 years people find these videos helpful, especially in today's reality with AI.
Also it's only been 5 years but with how fast we're advancing it feels like this was more than a decade ago.
Docker sandboxes are pretty cool. Covering this as the first step of setting up agentic workflow to build an app in my series.
Should I call my series "Vibe the right way"? π«£
I'm planning a 3 part series where I build a budgeting app for myself. Not a product, just the features I actually need, built with AI agents inside guardrails so they don't do something silly.
Part 1 is gonna be the intro & planning. Part 2 is about the architecture, ADRs, specs, setup, skills, guardrails, so the agents have the reasoning and follow rules/patterns. Part 3 is the build, using the Threejs, React, Inertia, Laravel with AI SDK, MCP, Reverb, Boost and other good stuff.
I'm more excited about part 2 because that's where I spend most of my time day to day at work.
Miss making videos. I've lost some of the motivation over time and I'm hoping to find the free time, and the thrill, to do it again.
I'm planning a 3 part series where I build a budgeting app for myself. Not a product, just the features I actually need, built with AI agents inside guardrails so they don't do something silly.
Part 1 is gonna be the intro & planning. Part 2 is about the architecture, ADRs, specs, setup, skills, guardrails, so the agents have the reasoning and follow rules/patterns. Part 3 is the build, using the Threejs, React, Inertia, Laravel with AI SDK, MCP, Reverb, Boost and other good stuff.
I'm more excited about part 2 because that's where I spend most of my time day to day at work.
Miss making videos. I've lost some of the motivation over time and I'm hoping to find the free time, and the thrill, to do it again.
@traversymedia Its funny and cute first few times and then becomes annoying. It tells me to defer some tasks to next quarter, or it should be done separately when I'm just trying to brainstorm and ideate on stuff. I have a Claude rule for this but it tends to ignore it sometimes.
Almost five years later, I just used @GioDev8 's "DI container with reflection API autowiring" video from the "PHP the right way" series to refresh my memory while building a DI container for a Java project I'm working on.
#stillgold
@aap_twak@PovilasKorop@laravellivedk I've automated plenty with opus and sonnet. Haven't had issues. Not every automation needs fable. But everyone has their workflows and if Sol works for you that's great