Ivan Mojsilovic

Today I tried @zeddotdev for the first time and I switched from VS in less than an our. It’s incredibly fast and since I use IDE only for final code reviews their default UI setup was perfect. Did I mention how fast it is? But will we use IDE next year?

@doodlestein Codex symphony have cron jobs that will monitor whatever task list you have (although you have to setup adapter for anything other than lienar). Ive done it for my tool, monitor columns and picks anything that gets in automatically.

Since Claude Code is nearly useless to me until these new draconian rate limits go away (note: I’m not talking about usage limits; these are limits on the number of requests per minute, basically penalizing the use of concurrent agents), I thought I’d list the 3 biggest features from Claude Code that I miss in Codex. Hopefully @thsottiaux and crew can close these gaps soon, since I’ve had to move almost all of my work over to Codex instead of half of it like I normally do: 1) Ctrl-R to search across all your previously entered messages in that project. Not having this built-in is a huge bummer, and it would be so trivial to add it. Sure, I can use my cass (coding_agent_session_search) tool for this, but that requires that I ask the agent to do it or use another terminal. I use this in CC constantly and it saves so much time. 2) No built-in looping/cron functionality. This is just incredibly useful for having CC manage a swarm of agents for me using my ntm (named_tmux_manager) tool. With CC, I can give it the high-level goal and tell it to check in with the swarm every 3 minutes and give instructions to any agents that are idle or need to be restarted. Then I can come back 5 hours later, and a massive amount of useful work has been done. Codex requires babysitting or just queuing up messages, which is much worse. 3) Hooks. I know there’s finally some motion on this now, but what I specifically miss are the pre-tool use hooks so I can get my dcg (destructive_command_guard) tool working in Codex finally. I really like both models and harnesses and hope desperately that Anthropic can get its affairs in order and fix this show-stopping issue for me and others who use swarms in our workflows. Right now, it’s basically not usable.

@michael_chomsky That’s right. Every todo is should have a pr and a preview link with tests results. How we should delegate todos or whatever we gonna call them is yet to be seen.

Super interesting perspective from Theo and Linear. Why would I make an issue when I can prototype something with just a prompt? It's faster to fix it than just leave it for later. The answer is that no matter how high productivity is, there's still endless todos to be done. But now every TODO should come with a PR and a preview URL. I'd also like to decide how much compute to spend deciding and validating a solution--for hard problems it would be cool to just spin up 5 agents collaborating for 4 hours because i KNOW that's what it takes to make a decent v1

@vigneshmani1512 @rseroter E.g. you can always feed plan skill through the prompt or you can say “plan this feature “ and model will use plan skill automatically.

@Mojsilovic @rseroter So why can’t we feed those instructions in prompt ? What’s the difference ?

The difference between a "vanilla" request to the Gemini model and enabling this new skill? Pretty dramatic. More to do, but we'll all keep learning the best way to apply these to our agents and tools. developers.googleblog.com/en/closing-the…

@vigneshmani1512 @rseroter You can, instead of “installing” skills you can simply feed instructions in the prompt or path to the file. The idea of “installed” skills is that llm will figure out on its own which skill to use when you send prompt

@vigneshmani1512 @rseroter Skills are instructions. Ive witnessed this level of performance improvement with skills i wrote for my work

@rseroter But skills are also prompt written in markdown format. How does it improve the performance? Am I missing something?

@melvynx Stack-auth is fully oss, works great!

Don't use Clerk. Delegating your auth is the worst thing you can do...

@arafatkatze Absolutely!

Calling it now, This form factor of multi agent orchestration will overtake every other agentic UX in the next six months. Doesn’t matter if it’s coding agent, product management agent or something else. Every multi agent flow of frontier models currently suffers from 2 main problems 1. Inference bound: Most of the time you are just waiting on inference while the agent rips on code in the background and that can take 8-10 minutes fairly regularly so the wait time very expensive productivity wise. 2. Take Isolation: you will work on the same mutating “source code ” with multiple agents in parallel each of which are inference bound and so eventually you will run into countless merge conflicts The kanban board solves for both of them by giving you the ability to act as an engineer manager for your IC individual agents, watching them cleanly with a clear headline level out look of every parallel agent. It works seamlessly with any coding agent whether it’s cline or codex or Claude code. To me this is a mental model shift that takes 10 minutes, you use the same coding agents like you always used before this but it will fundamentally morph your coding agent experience. Try out kanban and let us know what you think. npm i -g cline Is all you need.

@oskargroth Here’s a local kanban for your agents. Just md files with folders

I really like Linear, but I decided to unsubscribe last month. I simply don't need to pay $200/yr for managing what is now mostly transient .md files for an agent. And everything in this announcement is already built-in functionality in all coding agents.

@Jeff88Ho @GergelyOrosz Right, let fuck people even more.

@GergelyOrosz Solution: Candidates need to pay to an escrow account when applying. If they don’t make it to the final round, they lose the money.

The death of inbound applications is upon us: and yes, it’s in a big part because of AI making it dead simple to apply. And so inbound applications become noisy, with increasingly more of non-qualified people. And so companies rely on referrals and recruiters to source instead.

@aakashgupta I dont see them gone. PRD is the foundation for the implementation plan. It’s a foundation for support knowledge base. It’s the idea written so ai and human can understand what are we building. Once you have prd, ai can generate accurate implementation plan which generates code.

I don't think most PMs realize the PRD is becoming obsolete. For the last decade, the PM's core artifact was a qualitative spec. Clear requirements, user stories, acceptance criteria. The engineering team interpreted it, built something close, and the PM spent two weeks reconciling what shipped with what they wrote. The best AI companies replaced that entire loop with evals. A set of inputs your product needs to handle. A task that generates outputs. A scoring function that produces a number between 0 and 1. No ambiguity. No interpretation gap. Ankur Goyal built the eval platform behind Vercel, Replit, Ramp, Notion, and Airtable. An $800M company. He walked through building an eval from zero on this episode and the score went from 0 to 0.75 in under 20 minutes. That's a PM shipping a measurable quality bar before a single line of product code exists. Here's the part that changes the PM role permanently. When the product passes the eval and users still hate it, the eval is wrong. That's on the PM. Evals make PM judgment quantifiable in a way PRDs never did. You can't hide behind "the spec was ambiguous." There's a number now. Six months ago, PM interviews asked "how do you use AI in your workflow." The next wave of interviews is going to ask you to write an eval. The PMs who can encode user intent as a scoring function are building the one skill that survives every model change, every framework swap, every agent rewrite. Write the eval.

@nickbakeddesign Open source?

went from http://localhost:3333/ to live! was looking for AI testing app and all of them cost $$$, so I built one using Claude AI logs into the app, scans for flows - come up with tasks for designers & developers -> send them to Slack (I also built a mini Slack bot) > terminal > test in local > push to github > railyway & launch

@kr0der Gemini is much better in UI than codex.

look, i love GPT 5.4 a lot, i use it for 99% of my tasks. but you’re telling me to write an entire design spec WITH screenshots, colour palettes, user stories, copy, and more when i can go to Opus 4.6 and say “use frontend design skill, make it look nice” and still get a better frontend. at some point you just gotta admit that Opus 4.6 is ahead by quite a bit in frontend. it’s kinda like when people were setting up 25 ralph loops just to get Claude Code to run as long as Codex can by default. now we’re writing up design books just to get Codex to do frontend. just use Opus for frontend until the next GPT model.

@ericlitman @odysseus0z Quick update: i made new adapter in symphony, it will monitor my column and run agent to do whatever is in the card. Ill polish the UI a bit and will make it available.

@Mojsilovic @odysseus0z Haha take your time (or don't do it at all!), I was just registering my interest

TLDR: it is a cron job dispatching tickets from Linear to workers, each of which is a Ralph loop using a Linear comment as draft pad for persisted state. Yes it is all you need. Beautifully designed and minimal. github.com/openai/symphony

@elie2222 Im 2 years cursor customer, loving it! But brothers why are you usng these shitty cover up tactics. You are just digging your credibility further. You got caught, accept it, admit it, attribute Kimi properly and move on. You are covering up like politicians do their shit man…

People are way overreacting to Cursor being Kimi 2.5. It isn't. It's based on it. That's what open source models allow for. As long as you accept the licensing terms, there's nothing wrong with it. Composer 2 outperforms Kimi K2.5 on all coding benchmarks, so no, it's not just a Kimi wrapper. It's 75% Cursor training, 25% Kimi base.

@petergyang I have this: kanban of md files. Those can be prd, plan etc. i can switch provider cli and run the same context between them and gather multiple insights before final version.

What is your preferred interface/process to switch between Claude and Codex? Composer? Cursor? Terminal?

@BuildwithOmar @BlitzitApp Last year meta took my fb and insta account. Lost it in two minutes. In that moment I got epiphany and moved all our projects out of vendors locks. It took some time but we’re free now.

@BuildwithOmar @BlitzitApp That’s why you must own your product and only pay for infrastructure. You must own it in a way so you can switch between gcp, aws, whatever in a day.

Today is a horrible day for @BlitzitApp. We're a small startup with around 10,000 total users, 3K paid. Growing every single day. If you check us on Trustpilot, Product Hunt, anywhere, you'll find a committed, raving fan base. We run one of the biggest productivity Discord servers out there, ranked number one for the keyword "productivity." Until a month ago every piece of public opinion about us is overwhelmingly positive. And today, for the second time, @Google has suspended our @Firebase backend. No warning. No lead time. No clear explanation. Here's what happened. The primary email on our Firebase account was tied to a @GoogleWorkspace that got suspended 2 weeks ago. They make it almost impossible to transfer the primary owner email... This one suspension cascaded into taking down the entire backend that our platform runs on. Every user. Every task. Every piece of data. Inaccessible. This first happened two weeks ago. We went through the appeal process, which didn't even resolve it. We ended up having to manually reinstate our Firebase project through another owner tied to the account, without access to the primary Workspace owner. And now it's happened again. We are on a pay-as-you-go plan. We are a paying customer. And this is how we get treated. Let me put this in perspective. Companies like @stripe resolves issues almost immediately. Even @Meta's ad platform, at least responds the next day. Google? 72 hours. Minimum. With zero indication that they care about the impact on your business while you wait. Last week I tweeted about how this same Workspace suspension that restricted access to my YouTube channel with thousands of followers. No warning, no communication, just a vague "policy breach" and a full suspension. That was bad enough. Now the same issue has come back and taken down our entire product. It's not just my content anymore. It's our users' data. If anyone at Google sees this, something has to change. You cannot offer services that businesses build their entire infrastructure on and then pull the rug with no warning, no escalation path, and a 72-hour response time. That is not sustainable for any founder or any business. We are going to lose trust over this. We are going to lose users. I genuinely don't know if we can recover without starting almost from scratch. Google, you are a trillion-dollar company. You should be able to move faster than 72 hours for a paying customer. Put us through a proper support process. We deserve better. Please help us.

@leerob @1slimewell “Since people really want me to say” - this is the problem. You guys should have done it on day one. You haven’t, you got caught, admit it, correct it and move on. Being defensive just fuels the fire further.

@1slimewell x.com/leerob/status/…

“our inference partner” why are they skating around kimi so hard SAY ITS NAME

@HotAisle @proxy_vector We know this for decades. It’s pay to play.

@proxy_vector I have been talking about it for a while now. Nobody cared.

It is hilarious everyone suddenly waking up. SOC2… scam. SuperMicro… felons. Cerebras… phony numbers. The list goes on and on. Reckoning is coming.