Josh Owens

@bcherny @PrimeLineAI I'm having the gutted scroll back issue too, it really is messing with me. Started a few days back. No clue how to repro, just getting aggressive scroll back clearing with no understanding of how or why. Right click in a window is kind of cool tho..

Thanks for the feedback. > Scrollback gutted - session history is barely scrollable. I can't review my own conversation. can you tell me more about this? how do I repro? > Rate limits hit harder - prompt caching appears broken since March 23. Sessions that lasted hours now drain in 90 minutes. prompt caching is working correctly, and we've been shipping optimizations to make it work better. we announced reduced rate limits at peak recently due to our infra being strained because of really fast user growth. we're working around the clock to make this better, and landing improvements every day. please bear with us as we scale up -- it's really hard growing at this scale and we are working as hard as we can to keep up and serve everyone. > 1M context forced, no opt-out - extra usage required, 200k toggle removed. I didn't ask for this tradeoff.1M context not quite -- 1m context is free, and does not require extra usage. you can opt out by setting CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=20 to go back to 200k context window > Opus quality regression - execution bias overrides explicit instructions. The agent pushes into wrong directions constantly. Complex multi-step tasks that worked weeks ago now fail routinely. we have not changed opus 4.6 since rolling it out. if there's a regression, it's either in claude code, or due to a new memory or CLAUDE.md that your claude is pulling in. to debug, can you run /bug the next time you see this and post the feedback id here? > Every recent update optimizes for mass adoption. That's great - but the features power users already built themselves now come at the cost of the reliability we depend on. I'm not asking for special treatment. I'm asking: don't make your highest-paying subscribers collateral damage of growth. The tools I've built on Claude Code are my business. When the foundation shifts weekly, that's not iteration - it's instability. i totally hear you. we want to support both power users and everyone else. recent reliability issues have been a result of completely insane user growth. working as hard as we can to scale up all our services. next time you hit specific issues, feel free to tag me directly.

Today we're excited to announce NO_FLICKER mode for Claude Code in the terminal It uses an experimental new renderer that we're excited about. The renderer is early and has tradeoffs, but already we've found that most internal users prefer it over the old renderer. It also supports mouse events (yes, in a terminal). Try it: CLAUDE_CODE_NO_FLICKER=1 claude

MCPs are dead. I mass-deleted every one of them. 67,000 tokens burned before I typed a single question. One-third of my context window. Gone. Skills + CLI tools do the same thing for 1/7th the cost. And they actually compose. Full breakdown: joshowens.dev/mcps-are-dead

Your AI doesn't need better prompts. It needs a sprint planning session. joshowens.dev/ceremonies-for…

It's just... how engineering teams already work. No one person holds all the context. Specialists do focused work. A lead orchestrates. We accidentally rediscovered org design and applied it to AI. Full breakdown: joshowens.dev/ralph-wiggum-s…

But the real breakthrough wasn't restarting. It was scoping. One agent in a loop is still one agent doing everything. The unlock was fresh agents with scoped tasks; one plans, one builds, one reviews, one tests. Each gets exactly the context it needs and nothing else.

Someone put Claude in a bash loop called Ralph Wiggum. `while :; do cat PROMPT.md | claude ; done` It accidentally solved context rot, the reason Claude gets worse the longer you use it. Fresh start every cycle, progress saved to a file. 🧵

199 Claude chats. Couldn't find anything. So I dumped them into markdown files and pointed Claude Code at the folder. One month later: 500 files running my entire life. Here's the full walkthrough. youtu.be/iAHwFEbNrok

@jackrudenko @pusongqi Claude Code agents team has tmux integration, this looks like a nice UI on top of it...

@pusongqi Have you reinvented tmux:-?

You can even assign different agents under the same thread 🤯 Just like slack channels, except it's occupied with agents.

@craigzLiszt Agent teams will change it back, don't worry ;)

nearly all of the best engineers i know are switching from claude to codex

@adamwathan Both statements can be true... They aren't mutually exclusive. I was talking to someone the other day about feeling like I have an unfair advantage because I have all these patterns and systems imprinted on me. But someone new to all this may have a different way of thinking 🤷‍♂️

I don’t think AI is lowering the barrier to entry as much as it is giving the very smartest people an even more enormous advantage.

The fix: stop treating sessions like relationships. Treat them like stories in a sprint. - Start fresh often - One goal per session - Externalize decisions into files - Use subagents for isolated tasks Full breakdown: joshowens.dev/context-rot

Every developer has had this moment with AI. Hour 1: sharp, nails every suggestion. Hour 3: repeating ideas you already rejected. It's not the AI getting lazy. It's context rot. Here's what's actually happening:

At ~95% capacity, the AI starts compacting - flattening your conversation into summaries. Nuance disappears. That constraint you set in hour one? Gone. That approach you rejected? It's back, because the AI forgot you rejected it.

Every message fills a fixed buffer. Questions, answers, code, errors - all of it stays. Claude's 200K token window? Unreliable around 130K. Debugging tangents are the worst. One rabbit hole can eat 20% of your usable context in minutes.

I had 30+ Claude chats & couldn't find anything. So I dumped key conversations into markdown. Pointed Claude Code at the folder. One month later: 500 files running my life. The chat became a codebase. The AI became a collaborator with memory. joshowens.dev/life-os-markdo…

If your team's AI adoption is a ghost town, DM me

20 years shipping code. Built AI tooling at Mailchimp. Ran a consultancy that consistently cleared $1M+. Also speedran a burnout. Lessons were learned. Now I'm taking on work helping teams actually use the AI tools they're already paying for.

I spent 3 years watching engineering teams get AI tools and immediately do nothing with them. * Copilot: installed. * Usage after week one: lower than your test coverage. The tools aren't the problem. Nobody showed anyone how they fit into real work. 🧵

@kyleshevlin Dev shop life 🤷‍♂️

@stolinski @chantastic @meteorjs Crazy! What a fun trip that was 😊

@chantastic @meteorjs Shout out @joshowens

Discover Meteor