Principal Frontend Engineer @atlassian • Leading Jira’s frontend • GDE for Web • CSS&JS • ☕️ , Sci-Fi, Comics • owner @htmlmag
Nuremberg, Germany
Joined May 2008
- Tweets9.4K
- Following674
- Followers5.8K
- Likes3.1K
Pinned Tweet
My slides from today's talk. I talked about Micro Applications. I shared the idea of distributing the complexity of building SPAs to different levels for creating independent teams and improving the developer experience. #WeAreDevelopers2022
speakerdeck.com/bcinarli/ind…
Check it out and let me know if it works for you; it improved my workflow.
For the last couple of years, I've been actively using AI to help with my work. Then around this time last year, I promised myself I would become AI-first and stop writing code manually.
Since then, I haven't written a single line of code, or even a single keystroke, including git operations. Yet I have had more than 1500 PRs in different products.
On top of this, I created multiple products that became my daily tools.
I built tools to improve my workflow and created an agentic SDLC that my colleagues highly appreciate. Now my agents write tests, code, run CI, and review code. I still actively work on spec/plan creation and accept or revise the end product/results. However, I am involved very little in between, aside from hardening/improving the guardrails.
It is amazing how far it has come.
Thanks to Codex and continues usage reset from @thsottiaux , my hobby became my kids’ wish list deliveries.
“Hey codex can you prepare 3D model for this photo and print it in my printer”
ALT 3D printed NinjaGo golden lightning nunchucks
Meet @PanePeek , a file browser with an integrated code editor and markup renderer.
Since last year, AI agents have changed our way of working a lot. I don't recall when I stopped checking the written code and started to focus on reviews and implementation guardrails. Maybe it was 1000 PRs ago or even more.
ALT PanePeek one window file browser with HTML rendered preview
ALT A markdown filed rendered with syntax highlighted front-matter
ALT Edit the code in the file browse with syntax highlighting
ALT Dual pane windows
Nowadays, I mostly need to browse a directory, quickly check documents, make small text/code fixes in files, and copy the absolute path of the file/directory into the agent chat.
For these needs, TextEdit was too limited; launching an IDE or text editor felt like a waste of time. Then the idea of PanePeek came up. A file browser with an integrated code editor and a markup renderer that shows rendered Markdown for easy reading.
I built the app while travelling to my team meeting, hiking in the woods, and using my phone. Over the weekend, I fine-tuned the app and made a few improvements.
I also open-sourced the code: github.com/bcinarli/pane-pee….
In the repo, I have a directory called "agent-paper-trail" where you can see how I built the app, from idea/design to release, including decisions, specs, draft designs, and changes of mind during development.
On the way to the team meeting at the airport. In the meantime, I'm on my phone, continuing to develop a personal project with Codex on my Mac at home.
I fear that when my wife hears the voice-over tests of the app and task-finish pings of Codex, first she will freak out, then turn of my device 🫣
When I show my work to friends other developers, they usually tell, use Opus, use Sol, use Gemini etc.
All of them tells this according to their experience and struggles of explain task to different models.
However I passed that way back, built myself a harness manager that helps me to switch between Codex to Claude to Cursor or another Agent/model at ease.
Nowadays, rather than the model and agent, I’m more curious about the reasoning.
While I was cleaning up long “to read” tabs, I came along Pulse 140 from @GergelyOrosz published in July 2025.
I’m curious after more than 1 year, whether this research is still valid
ALT Screen shot from Pragmantic Engineer Pulse 140 newsletter, focusing on the section mentioned a research takeaway stating “Surprisingly, we find that when developers use AI tools, they take 19% longer than without. AI makes them slower.”
+1 to what Fatih said. I had the same work type but more like 4pm to midnight for sometime and I missed all good things with my family and kids.
Luckily in my current company we don’t have restrictions and I mostly prefer 6am to 2pm work since I’m a morning person
This doesn’t work very well. I tried it when I had no kids and started working remotely for a U.S. startup in 2012/2013.
At first it looks great on paper, but over time you realize it’s pretty bad socially. It can work if you’re single and don’t have many people around you, but if you have a family, kids, and a large network of friends, your life revolves around them. And surprise, their whole life is basically 9 to 5 :)
You also don’t just go to bed after 10 PM every night. Your body and habits change, and suddenly you end up going to sleep at 2/3/4 AM, which means waking up very late too.
If you’re working remotely, I think the best approach is to mostly live and work in local time, but keep some pockets during the day for working out, groceries, errands, and so on. Then be flexible and have meetings with your coworkers in the evening on some days, even if that means meetings at 8/9/10 PM.
That gives you the best of both worlds. You’re still connected with your team and can make sure everything is covered, while still adjusting work around your actual life.
Türkiye’ye her geldiğimde yaşadığım en büyük sıkıntılardan biri, özellikle 10+ yıl evlerde ve eski binalarda, odaların sadece kapısı yanında 1-2 adet priz olması. Hemen hemen hiç bir evde (en azından benim gittiklerimde) ikinci bir duvarda priz yok. O yüzden bütün evlerde duvardaki tek prize takılı üçlü prizler ile çözüm üretiliyor. Böyle olunca da özellikle bilgisayar şarjı ve çalışma için odanın sabit bir noktasına çakılı kalıyorsunuz
💯 totally agree. Especially for my personal workflows, coding style, communication and writing styles, I created personal skills and regularly improving them by analyzing my session journals.
Every person has their own way of working and something reusable not always best fit. For me the AI workflows should be personal as well. The way we interact with AI agents are different, so the skills better be different/personal.
Though still we can have some reusable skills in project level that team can have it as keeping their agent on the same page.
This has become the norm nowadays. In my team, when a PR is created, cloud code review agents review the code in accordance with our structured guidelines.
On top of these, my teammates and I are using our own agents to make reviews of the same PR according to our focus preferences, flat or deep reviews depending on the size.
When the comments are added, the PR owner reviews the comments (ok, made their agent review the comments 😅) and runs their agents to handle the comments, either apply fixes or use their judgment to revise/challenge those.
Some funny things: we have a rule against restating the same thing in the PR if it's already mentioned in another comment; instead, add a thumbs up.
Another thing: don't start replies with “Good catch” or “You are right,” like a cushioning intro, and keep comments very trimmed and focused, since we are also reading the comments. I also have a token budget in my cognitive reading 😬
With this setup, I've “reviewed” over a thousand PRs in the last 10-11 months.
Fascinating to look at PRs in the Bun repo: much of it is AI reviewers talking with AI bots...!
CodeRabbit and Claude review bot talking with Robobun. Screenshot from this PR: github.com/oven-sh/bun/pull/…
