←
Bilal Çınarlı9,406 posts
Bilal Çınarlı avatar

Bilal Çınarlı

@bcinarli

Principal Frontend Engineer @atlassian • Leading Jira’s frontend • GDE for Web • CSS&JS • ☕️ , Sci-Fi, Comics • owner @htmlmag

◷ Joined May 2008▧ 722 Photos and videos
Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·15 Jun 2022

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…

Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 16

Self-improving review loop:

I have been using certain guardrails and quality gates to improve the agent's output. I'm focusing on producing high-quality work before locking in a decision or committing code. I believe that once you decide on something and lock it in, you need to maintain that decision until you supersede it or remove it. Funnily enough, I was saying the same thing in many of my conference talks about "merged code" before our agentic development era. "The time you merge your code, you start to maintain it. Think about which is more important: writing it in an hour or maintaining it for years? Pay close attention to what you deliver" In my agentic workflow, I have multiple self-improving loops for the agents. The diagram is from my plan/prepare/review loop. The good part is that I can apply it to any work, from planning to design, writing to architecture or coding. The loop starts with defining the scope of delivery and breaking it into smaller, manageable units with regular reviews. The key part of this loop is having agents self-review and improve their output before presenting it. I have two complementary skills for this that agents use after they finish their initial work; • `tighten`: focuses on simplification. The agent looks for unnecessary abstractions, possible similarities and reusability with neighbouring code or architectural elements, or unneeded details in suggested prose that don't add value. • `probe`: focuses on failures, edge cases, weak assumptions or missing validations. Explores gaps, ambiguities, claims without evidence, or other planning, analysis and prose. Both skills work with evidence and validation-based confidence. They test and validate their findings. Also, during this self-review, the model/agent tries to answer the question, "What would break in 6 months or a year when we deliver this?" The earlier I explained my "critic-orchestration" skill to get a second model opinion. After `probe` runs, if the model couldn't set a good confidence level, e.g. "high confidence" in its review, it recommends a bounded independent review to improve the judgment-heavy: • a hard-to-reverse architecture or migration decision • cross-boundary behaviour with non-obvious invariants • significant security, data, compatibility, rollout, or operational risk • an unclear root cause or fix shape after local evidence is exhausted After this self-review, the agent presents its work to the orchestrator or to me. I review the output against my thinking and how I initially planned the work, or, in short, according to my "taste". The orchestrator reviews the agent against the requirements I defined for the project, along with core engineering rules and task validation rules. Based on final approval or review, the agent continues to the next unit or revises the unit to address the feedback.

Post attachment
Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 16

Critic orchestration:

My friends and colleagues usually ask when I mention having a cross-family review of the work before finalising the code/creating a PR, or drafting an architectural or implementation plan. I sketched the workflow; Basically, how it works: either the orchestrator agent or I control the work and instruct the active agent to have a critic review. Most of the time, depending on the agent I'm in, I tell the agent, "discuss this with Claude", "discuss this with Codex", "get a review from Cursor", etc. I have a "critic-orchestration" skill that tells the agents how to start a critic review with cross-family and which model to choose, and which reasoning level to choose. Also, depending on the review type, it gives instructions on whether they need to prepare a read-only review package or just say "check this code" or "check this design" for more free-style reviews instead of bounded ones. The agent runs the cross-family agent's CLI (codex, cursor-agent or claude) with a particular setup and payload. After the critic finishes their review, they return a review document with a general verdict (_usually it always comes with "needs work" :D, feels like how we don't like other developers' work at first; they also tend to find things all the time_). Along with the verdict, they also share "must fix", "should fix" findings and follow up/optional simplification/hardening suggestions. The agent automatically verifies the findings and fixes the must-fix and should-fix ones. Depending on the criticality, the agent shares optional follow-up ideas and asks my opinion, or the orchestrator agent's, on whether to implement them or create an issue/note to handle later. The good part of the routine is that it works for architecture, design and even critical document writing like spec creation and plan preparation. It helps you spot problems before locking in decisions and reduces the overhead of reviewing tons of agent output. Believe me, I read far more documents daily than I used to write code. With this routine, I was able to deliver much higher-quality work with less inference on my side. Also, when running multiple agents/orchestrations, this review workflow catches the issues before creating a PR, deciding on scope or implementation details.

Post attachment
Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 16

I have been using certain guardrails and quality gates to improve the agent's output. I'm focusing on producing high-quality work before locking in a decision or committing code. I believe that once you decide on something and lock it in, you need to maintain that decision until you supersede it or remove it. Funnily enough, I was saying the same thing in many of my conference talks about "merged code" before our agentic development era. "The time you merge your code, you start to maintain it. Think about which is more important: writing it in an hour or maintaining it for years? Pay close attention to what you deliver" In my agentic workflow, I have multiple self-improving loops for the agents. The diagram is from my plan/prepare/review loop. The good part is that I can apply it to any work, from planning to design, writing to architecture or coding. The loop starts with defining the scope of delivery and breaking it into smaller, manageable units with regular reviews. The key part of this loop is having agents self-review and improve their output before presenting it. I have two complementary skills for this that agents use after they finish their initial work; • `tighten`: focuses on simplification. The agent looks for unnecessary abstractions, possible similarities and reusability with neighbouring code or architectural elements, or unneeded details in suggested prose that don't add value. • `probe`: focuses on failures, edge cases, weak assumptions or missing validations. Explores gaps, ambiguities, claims without evidence, or other planning, analysis and prose. Both skills work with evidence and validation-based confidence. They test and validate their findings. Also, during this self-review, the model/agent tries to answer the question, "What would break in 6 months or a year when we deliver this?" The earlier I explained my "critic-orchestration" skill to get a second model opinion. After `probe` runs, if the model couldn't set a good confidence level, e.g. "high confidence" in its review, it recommends a bounded independent review to improve the judgment-heavy: • a hard-to-reverse architecture or migration decision • cross-boundary behaviour with non-obvious invariants • significant security, data, compatibility, rollout, or operational risk • an unclear root cause or fix shape after local evidence is exhausted After this self-review, the agent presents its work to the orchestrator or to me. I review the output against my thinking and how I initially planned the work, or, in short, according to my "taste". The orchestrator reviews the agent against the requirements I defined for the project, along with core engineering rules and task validation rules. Based on final approval or review, the agent continues to the next unit or revises the unit to address the feedback.

Post attachment
Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 15

Yıllardır kontrol paneli çok kullanışsız olmasına rağmen, hem Türk firması hem de sadece alan adlarım olduğu için @ihstelekom kullanıyordum. Bugün bütün alan adlarımı @Cloudflare’a taşıdım. Hem ücret daha ucuz, hem kontrol paneli ve ek servisleri daha kullanışlı.

Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 15

Another AI/Codex 3D project that I build new lids for IKEA jugs I took the measurements with my digital caliper and shared with GPT 6.1 Sol with explanation of what I want. It handled the rest of the 3D design and slicing and printing in Bambu Studio.

Post attachmentPost attachmentPost attachmentPost attachment
Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 13

I think yesterday the @ChatGPT desktop app was updated, and since this morning I've been fighting with an agent that can’t use my terminal; it couldn’t connect to Git repos, and the “CMD + Enter” setting while prompting wasn't working. Then, while trying to debug and understand what was happening, it started saying “ChatGPT work”. I was like, “What?” Later I figured out my default Codex became ChatGPT 🫣

Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 11

I have been busy with some improvements. The app was small and fast but I wanted to have more performant and reliable app experience. The performance, CPU and RAM usage are even better now. Browsing/changing folders and editing files are up to 2 times faster now.

We have released the 1.3.0 beta 2. In the 1.3.0 we are focusing on performance improvements and search improvements. Pane switch and cycling through folders are now 2 times faster, search sort the files according to their relevance (thanks for suggestion to @omerhakanbilici)

Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 11

I probably wouldn’t imagine this a year ago. Today we had an internal survey about our IDE usage and when I was filling the survey, I figured out over the last couple of month I haven’t opened an IDE. 😱

Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 8

My friends and colleagues usually ask when I mention having a cross-family review of the work before finalising the code/creating a PR, or drafting an architectural or implementation plan. I sketched the workflow; Basically, how it works: either the orchestrator agent or I control the work and instruct the active agent to have a critic review. Most of the time, depending on the agent I'm in, I tell the agent, "discuss this with Claude", "discuss this with Codex", "get a review from Cursor", etc. I have a "critic-orchestration" skill that tells the agents how to start a critic review with cross-family and which model to choose, and which reasoning level to choose. Also, depending on the review type, it gives instructions on whether they need to prepare a read-only review package or just say "check this code" or "check this design" for more free-style reviews instead of bounded ones. The agent runs the cross-family agent's CLI (codex, cursor-agent or claude) with a particular setup and payload. After the critic finishes their review, they return a review document with a general verdict (_usually it always comes with "needs work" :D, feels like how we don't like other developers' work at first; they also tend to find things all the time_). Along with the verdict, they also share "must fix", "should fix" findings and follow up/optional simplification/hardening suggestions. The agent automatically verifies the findings and fixes the must-fix and should-fix ones. Depending on the criticality, the agent shares optional follow-up ideas and asks my opinion, or the orchestrator agent's, on whether to implement them or create an issue/note to handle later. The good part of the routine is that it works for architecture, design and even critical document writing like spec creation and plan preparation. It helps you spot problems before locking in decisions and reduces the overhead of reviewing tons of agent output. Believe me, I read far more documents daily than I used to write code. With this routine, I was able to deliver much higher-quality work with less inference on my side. Also, when running multiple agents/orchestrations, this review workflow catches the issues before creating a PR, deciding on scope or implementation details.

Post attachment
Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 7

It is my hobby project that I’m removing my need of using IDEs, text editors, project viewers etc. After working extensively with AI agents, my essential need now is just checking my what is inside a text file or render and peek quickly to markdown, code or other text file rather than reading code, editing files. Using an IDE or code editor for this need feels overkill and I wanted to have a file browser that can render or show contents of the text based files as I’m checking the directories and AI agents’ output. It is fully written with AI, mostly autonomously from my phone with an orchestrator agent coordinating feature deliveries.

My first MacOS app is available in Mac AppStore ✌️

Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 2

A small project I did for Halloween last weekend. I use AI agents a lot for daily work aside from coding. So I wanted to have a lantern and a companion for it. After explaining what I wanted to @ChatGPT in Codex, it created images of this Jack-o'-lantern and Scarecrow, which I then converted into 3D-printable models using @tripo3d_ai. Then I exported the 3ML files and adjusted the sizes and colours a bit with ChatGPT again to finalise the print setup. The lantern took 16 hours to print, and the scarecrow took 27 hours in my Bambu Lab printer. Kids enjoyed them a lot. To my bad, while cleaning the supports, I broke the ivy leaf on the scarecrow's head 🤦

Post attachmentPost attachment
Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Nov 1

Do you (we) care this any more? - either your agent decide - or your linting rule block - neither of those our preferences like spaces/tabs, brackets/single lines etc…

So const or let?

Bilal Çınarlı avatar
Bilal Çınarlı@bcinarli·Oct 31

Sometimes orchestrator works for 6-7 hours with 8-9 agents to deliver a feature. Regularly observes the agents’ work, when needed gets cross-family review, compare the output with specs/decisions/plans when needed input pings me.