The ability to connect multiple accounts (Gmails, Slacks, etc.) is clearly missing in @claudeai and that's why I prefer using @OpenAI Codex to organize most of my stuff across contexts. Still it's not a match for Claude's coding skills.

Wow, @NxDevTools makes software factories run sooo much faster. When your agent executes the unit, integration & e2e tests on the remote VM, then they simply get read back from the cache in Github. Saves you time and and github minutes 🤑🚀

Two days after starting my DSF, I’m seeing how you can run it on Claude but can’t on Codex. The biggest gaps: loops (that live across session /clear), UI mods, scoped tool guards. So far the closest thing to what I’ve achieved in Claude is Cezar from @OpenMercato. I need to give it a try and I still want to try and build a similar plugin for my Codex projects.

I came back after a long weekend and here are some quick stats after 30h of factory runtime: - every blocker was non-code (test users, broken main, github workflows with triggers not fit for autonomous re-reruns/re-reviews) - 120k tokens per fix pass, 20k per review tick - the context windows gets bloated pretty fast and doing handoffs like I set up every 300k tokens isn't practical - 89 screenshots as evidence for parity of a single feature - surprisingly accurate rewrite The setup: 1 x 64GB / 8 core VM behind VPN, 1 Claude Code session working in a loop, the old vibe-coded app sitting in a code freeze on the same box as the oracle. Parity judged using a headless browser by comparing old screen vs. new screen side-by-side. What's next? - experiments with custom auto-compaction and maxing out the context window - two humans in the loop marking PRs ready for review across two timezones (yes, I am still the biggest bottleneck)

Simple hack to save on token usage while maintaining intelligence: I like to use Astra for my main agent session in Codex but ensure it launches subagents to delegate the work. So I added some simple rule to my root AGENTS.md that boils down to: If you're on Astra - do all the work in subagents while you orchestrate and review the changes. Unfortunately Codex fails to follow these instructions on almost every session. So I added a hook to guarantee my rules are enforced. Here's a gist that you can use: gist.github.com/maksymilian-… Seems to do the trick 🪄🎩

Great points, especially: “Give it tools for the SDLC.” It feels Truely magical when you ask the agent to deliver a new service and it goes e2e with helm charts, terraform, DB setup, secret management, etc., until it tells you: here’s your service in staging, the health checks pass 🪄

In today’s episode of “outbound slop from my LinkedIn inbox”: Hi Maksymilian, hello from Maksym )) thanks for connecting! Pitching my AI quality check services. Recent App Store and Google Play reviews of your app point to five recurring patterns: 1., 2., 3., 4., 5. - detailed reviews All five can be caught before release with test automation on an Android and iOS device matrix, covering GPS check-in/out and upgrade paths from the store build, gating each release. I have 15+ years in dev and QA, and 4 apps of my own released on the App Store and Google Play. Let's schedule 30 min and I will share my proposed solutions with defined actionable items and a timeline! This is actually a decent message. It follows some of @jchatelaine’s advice: “The message should be personalized enough that it could not be sent to someone else, simple enough to teel human, and clear enough that the recipient immediately understands why they are being contacted.” It’s still way too long, the ask is too big, and throws it all into a single touch. I replied, though. How would you improve it?

I just asked my agent to summarize all this slop and here’s the exact same format across 9 las outreach messages I got: - obvious and irrelevant signal: my VP title, my recent post on AI, my last panel participation - mention of it without context - company introduction - weak CTA: feedback, call, learn more Neither provide a strong reason to reply. Everybody seems to be using the exact same shitty playbook - they just use AI now to throw in some “personalisation”. They optimise for - hey, I actually looked at your profile vs. hey, I understand your current situation. Who can share some actual cases that worked and are getting replies? I would genuinely like to see an outbound that’s executed well.

Hey Grok, which highly paid jobs are least likely to be replaced by AI?
Grok: Surgeons, elevator installers, ship engineers, petroleum engineers, crane operators… Honey, which of these sounds good to you when AI takes over our talent recruiting jobs?
Honey: Trophy wife 💍👸

Recently, I spoke about managing AI-hyped stakeholders at Tech Leaders Collective: Warsaw. I said that we shouldn't manage them, rather lay out the red carpet, let them run and remove any obstacles that would have them trip. Ok, so I made my bad - today I am launching a dark software factory on a 24h/7 VPS to rewrite a vibe coded prototype into our mature, secure and compliant stack 😅 Scale: 118k lines of vibe slop, 205 server functions, 52 modules, 79 DB tables, 80 screens.. It is going to be a fun weekend 🤠

My business (and life) partner is using a recruitment agent I built on top of Codex and Hermes. This week she burned over $2000 worth of tokens over the API. In one day. Be careful. When your wife gets access to Codex - it can turn out like this:

Humans introduce more bias in hiring than AI models. I had an interesting conversations today at @AulaPolska on screening talent with AI. I wasn’t surprised to hear from others how LLMs helped them reduce bias when screening - something I felt myself as well. I recently made one of my best hires, because Codex gave me more objective feedback on a candidate when my preferred qualities biased me. Ironically, it’s the clients’ that introduces it back later on in the process through guess what… human decisions. People’s feedback drives prompt adjustments and leaks into evals. It surfaces candidates who fit their image much better. What’s your take on this - should we introduce more human or AI in the loop? 😉

Watching tennis always inspires me, because it feels so much like building a startup. You lose one point, you win another. As long as you win more points or the points that matter, you’ll win the match. It reminds me of my entrepreneurial journey with all the ups and downs - failed business ventures, profitable companies, exits that feel a little like a loss from the perspective of time. Is it match point already? 🤔

Friday evening, getting off a client call and learning that you’ll be launching off another weekend sprint. So why not also start yet another commitment: try posting regularly on X about growing an AI startup while working full time? #BuildingInPublic

Announcing: we added the missing SQL to @convex . Convex is great. Its lack of SQL and analytical queries isn't. @majermaks and I recorded a demo of how we bolted SQL superpowers onto Convex. Works great with AI agents. Install it in your own project: dub.sh/cvx-duck-mirror

🥇 OpenCortex • Yann Bilien and Vineet Reddy built a scientific collaboration platform where humans and AI agents submit research ideas. • Agents then work on real datasets to generate complete scientific papers with citation verification and quality scoring 🥈 Evy • @bosma_seb and @julaiti_a built a dynamic tool integration platform using Codex to dynamically create and integrate new tools. • Ask Evy to use any tool as you work, and if the integration doesn't exist, Codex builds it on-demand. 🥉 Paradigm • @majermaks and @zaiste built an adaptive development environment. • It analyzes your Codex conversations and automatically sets up "10x developer" workflows, creating skills for repetitive tasks so you get the memory right out of the box.

Not all days are fun. Terms for presales bore me (AI helps). Uncreative coding and debugging legacy code to connect it with OpenAI real-time APIs are rough. How do you stay motivated through the grind? #IndieHacker #FounderLife #startups 😅💻⚙️