Angehefteter Tweet
davert 🇺🇦
18.2K posts
davert 🇺🇦
@davert
Prince of Persia on day, developer at night. Author of @CodeceptJS and @Codeception testing frameworks. Currently working on @testomatio to make tests visible!
Kyiv. Ukraine Beigetreten Kasım 2008
898 Folgt3.1K Follower
@neogoose_btw The only benchmark if WTF/s
The metrics is based on how many times I stop agent with the all caps fuck what you are fucking doing.
So far opus is true leader, comparing to codex and Cursor. I almost not get annoyed using it
English
Internal closed source benchmarks make ABSOLUTELY ZERO sense.
It's hard to imagine that everyone is loosing to Cursor in the "CursorBench" scoring system.
Cursor@cursor_ai
Composer 2 is now available in Cursor.
English
@nullvoxpopuli I ask to find 5 most critical issues
I don't want to check on everything, 5 is enough for a developer to start fixing
Then it will pick next 5
English
Can it review my giant prs tho?
Every AI review tool I've tried gives up too soon.
I don't want these tools to give up when a human what give up.
Read the whole thing
Claude@claudeai
Introducing Code Review, a new feature for Claude Code. When a PR opens, Claude dispatches a team of agents to hunt for bugs.
English
@neogoose_btw But your issue was with rails 8 defaults
Without vibecoding I would do the same. I would choose the framework I know good and use it for my hobby project.
(I also don't know where did he made all of Ruby LOC and what they do, as Rails is very compact language)
English
Next time I will play 2 lies 1 truth I’ll say that CEO of ycombinator publicly called me smartass on a million views
And it’s gonna be a lie cause this shit will never get a million views
Garry Tan@garrytan
All these smartasses think it’s vibe coding I made Posterous (one of the biggest Rails sites in the world at the time) and Bookface (a social network used by half of YC alums daily) If you can fly, you fly to where you were and go from there Vibe code haters, I eat your hate for breakfast
English
@evilmartians I would be really afraid to give AI to code in something other than Rails for backend. It already knows architecture, principles, ecosystem. So much less chances to create slop if it just follows standards.
English
Seems like every other week there's a new study confirming Ruby is the most cost-efficient language for agentic coding. And every time, it resurfaces the "Ruby is dead" conversation.
We're confident there's future in Ruby. Not just because we've seen it outperform other languages, but because of the companies actively building with Rails.
This discourse inspired @inazarova's keynote at @rubyconfth. She interviewed 50+ startup founders, including recent YC grads building AI, fintech, and dev tools. The takeaway: "Rails is the highest-productivity framework for startups, and the AI era only makes that advantage sharper."
Watch the full talk: evilmartians.com/events/rubycon…
Yusuke Endoh@mametter
I benchmarked which language works best with Claude Code. Ruby, Python, and JavaScript came out cheapest, fastest, and most stable! See my article in detail: dev.to/mame/which-pro…
English
@Karmedge I also think it is a scam
I replaced @CodeAntAI with 1 opus prompt and result is so much better I can't say. No trashy noise, no stupid diagrams, only thibgs that matters
Opus with a simple prompt performs better than all their AI tooling
English
Can someone explain to me code review startups valued at 1b$
when its literally 1 command in claude code?
English
@ContextQa @CodeceptJS As for healing in codeceptjs we will have 2 strategies
Fast (bad locator) which was made in 2024 and worked good with llama and mistral models by that time
Complete when we pass test trace to AI agent and it will be able to run test step by step to find an issue
English
@ContextQa @CodeceptJS I am in process of something much bigger like self writing tests.
However, my aim is to focus on cheapest and fastest models like gpt-oss-20b
This is my new project on it
github.com/testomatio/exp…
English
AI healing of flaky tests is a reality!
In @CodeceptJS 3.6 you can connect various models to allow them revive failing tests.
I tried 7 #AI models to heal my browser tests and they worked surprisingly well.
But which one is the best?
dev.to/davert/unleash…
English
@neogoose_btw Hey, its ok, to confess
All yiur friends are already vibe coding!
English
Query parsing feature for fff.nvim is in. Absolute game changer.
Now you can just type as much keywords about your file in space + special tokens and the fastest in the world glob powered matcher (written by me as well)
English
This is actually huge
dax@thdxr
we've taken the first steps in eliminating opencode's ripgrep dependency :)
English
@neogoose_btw I think we are out if days where people shared experience with other people
Now:
- I have 5 agents running
- And I have 10
- I have 50
- Opus opus opus opus
English
With llms now you can not let yourself spend a week tickling some really annoying bug/issue.
What we are going to make conference talks abotu now? Endless agentic madness?
English
@nullvoxpopuli Thanks!
Btw, is there a way agent can access component tree as in Ember devtools? Just curious how can i close the loop
English
I should make a site for the ember skills i made
Available here
github.com/NullVoxPopuli/…
Alfredo Perez 🦋 @alfredo-perez.bsky.social@alfrodo_perez
Inspired by @vercel react-best-practices I just shipped @angular Best Practices—curated, human-reviewed rules that teach your AI coding agent how to write better Angular code. 19 categories. Every rule under 50 lines. Incorrect vs correct patterns and avoiding code when rule is self explanatory ...saving context tokens 🪙 Separate skills per library — install only what you use: Core Angular rules NgRx SignalStore TanStack Query Want a rule added? add it in the website Hate a rule? submit a GitHub issue Works with Claude Code, Cursor, Copilot, and 30+ more agents. `npx skills add alfredoperez/angular-best-practices` angular-best-practices-web.vercel.app
English
@DmytroKrasun @forgebitz People say vibe coding aa solution ro everything like it costs nothing
No, tokens are expensive on their own, infra is expensive too, hiring clawdbot to be sysadmin is also $
So I think saas will work, and can even benefit from AI agents
Like having skills for screenshotone
English
Talked with @forgebitz about how they use ScreenshotOne in Promptwatch and why they decided to buy instead of build.
I don't know how long that equation will last. Maybe, at some point building with AI will be a no-brainer versus buying, but once again our conversation reminded me that to win you need to focus on your customers and your core competitive advantage; everything else can be bought/outsourced.
screenshotone.com/blog/promptwat…
English
@vercel_dev Yesterday was looking into repo and surprised there is no website 😅
Does bash tools provide file reading limits and suggest using jq/grep/tail for big files?
English
𝚓𝚞𝚜𝚝-𝚋𝚊𝚜𝚑 gives agents the power of filesystems and shell commands without the need to spin up a full sandbox.
justbash.dev ↓
Malte Ubl@cramforce
just-bash has a website now. It's just bash justbash.dev
English
@nateberkopec Well, there was 0 progress in vibe testing niche in recent years
But lets imagine what happens if all industry goes that way...
But in the end of a day some one should be blamed for bugs, so yes QAs are needed 😅
English
If anything, human QA becomes even more important in an LLM coded world. Probably the job will move up a level to product experience: humans ensuring “does this feel good?” Call it Taste Assurance
DHH@dhh
@scottharveyco We certainly are! Everything I wrote back in 2017 still applies. signalvnoise.com/svn3/the-value…
English
@nateberkopec I think the question here should we drop LLM into everything or use it as a regular function inside a well designed application. Functions nay have side effects, llms too. But good designed apps put boundaries to ensure they won't affect the system
English
What Clawdbot shows us is that LLMs are not authentication/authorization schemes.
If they CAN do it through tool calls, they WILL do it given a particular series of tokens as input. The only way forward is restricting tool calls (or the environment those tools access).
English
While Codex building an MVP of a side project and testing it with Playwright without me touching anything...
I realized maybe I could deliver food. Or open a car wash shop...
Have you already checked open jobs offline?
English
@rodgetech @DmytroKrasun Codeant which I tested in parallel
It made valid points to look. Not always correct in suggesting changes, however its insights are worth taking into account
While coderabbit posted jokes, ascii art and lymrics like ChatGpT in early days
English
A counterpoint to people having fun at roasting CodeRabbit:
Harjot is a human. And we all make mistakes.
If you didn't make mistakes, even in PR, you probably were never a founder or didn't build anything people use.
I am not affiliated and don't use CodeRabbit, but people seem way over exaggerate over that reply, instead of focusing on the code review tools.
Harjot Gill@harjotsgill
Aiden - you clearly have no idea about what you are talking about. We have more users than everyone you just mentioned (combined). See aitooltracker.dev Still, I am curious how we can make it simpler and fool proof for indie developers. Maybe simpler controls and a lot of handholding can help.
English