Raine Virta

@PaulRBerg Press Ctrl-G

This is the most annoying thing in Claude Code. Hiding raw text when you paste more than 4 lines. Terrible UX decision.

@rohanpaul_ai Why was it deleted?

Andrej Karpathy just put out this tool that looks at AI's impact on job. He also deleted the original Github repo very quickly. Basically, he pulled 342 job types from the Bureau of Labor Statistics and had an LLM score each one from 0 to 10 based on AI exposure. The average exposure score is 5.3. Move the score, move the probability it will get wiped out by AI. - Software developers 9/10, - medical transcriptionists are a 10/10. - Lawyers 8/10 - General Office clerks 9/10 Basically any screen-based jobs are in trouble. $3.7T annual wages in high-exposure jobs (7+) pre-computed as ∑(BLS employment count × BLS median annual wage) over exactly those occupations whose Gemini Flash score is ≥7.

@joodalooped Make a project on GitHub that people start contributing to, you will notice it's cheap indeed. People will send half-working PRs with less trouble than it would take them to write a coherent issue of the thing they are solving.

idk man, i don’t think code is as cheap as everyone is making it out to be

"it's also very messy (unnecessary utils, duplicated code, random `as` everywhere)" This is a workflow issue. Code quality is a 100% solved problem. Review your code with other agents (see e.g. consult-llm-mcp) and set up hooks to prevent tractable errors like those `as` assertions.

the bottleneck for coding agents is now testing / code quality agents are OK at writing code in the happy path, but don't consider edge cases on harder tasks it's also very messy (unnecessary utils, duplicated code, random `as` everywhere)

"The entire argument for zmx instead of something like tmux that has windows, panes, splits, etc. is that job should be handled by your os window manager." There's never going to be OS-level window manager for Mac that would allow switching between projects with same fluidity as tmux does allow me to switch between sessions.

@dmwlff github.com/neurosnap/zmx

I used to be a kitty stan who looked down my nose at tmux users, but I'm back to using it all the time. I get that it kind of sucks, but it's sooo convenient and powerful. I would love it if Ghostty added scriptability and persistence. I'd drop tmux in a heartbeat.

So what's the best way currently to have an agent interact with browser? There's so many options lately, playwright-cli, agent-browser etc. A potential benefit with CDP at least I see is that you can reuse the existing browser session and don't have to always open an adhoc Chrome session for the agent.

It took another two months but Chrome 146 is out since yesterday! And *that* means: with a single toggle, you can expose your current live browsing session via MCP and have your CLI agent do things in it. Aaand I have been waiting to deal with my LI connects until this moment.

@TheReeveOliver @SearchForRyan Anthropic has two competing interests here. They want Claude Code to be powerful, but at the same time they want you to burn tokens to compel user to upgrade. Anyone who has paid attention has noticed the trend that you're getting less mileage out of the plans. n

@SearchForRyan My question is if all of these things are needed for claude code to function well, why are the not part of the standard setup. Why do I keep discovering hacks on X instead of this shit already being setup?

okay why did nobody tell me about state.md for claude code? i've been stuck wondering why a project i build on daily was taking 8-10 mins to make changes turns out claude was reading EVERYTHING (all my code) before it made a change. i added state.md, architecture.md (alongside my claude.md) and bam, 8,000 tokens to ~1,000 tokens and from 8 mins to ~ 60 seconds. claude.md wasn't enough in this instance annnnd i'm so happy lol

@tonybuildsai @karpathy @nummanali workmux.raine.dev/guide/dashboar…

@karpathy @nummanali yeah tmux gets you like 80% there but managing 5+ agents without a proper dashboard is pain. someone is gonna build this and make a killing

Claude Code teams with tmux is really cool When you run with team mode enabled in tmux, it automatically opens the additional terminal in pane I don't really get my main agent to orchestrate, I chat to them myself CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=true claude

@hckrclws @karpathy @nummanali

@karpathy @nummanali tmux gets you there 80% of the way but the other 20% is rough. something like htop for agent state would be huge: which ones are blocked, which are waiting on I/O, which just finished and need review

@hckrclws @tom_doerr It doesn't restore, but each new worktree window opens with a configurable pane layout.

@tom_doerr workmux is exactly what my brain needed. been losing context switching between worktrees manually in tmux for months. does it restore window layouts too or just the pane/worktree pairing?

Tool for managing git worktrees and tmux windows for development environments github.com/raine/workmux

@gregpr07 Voice mode baked into a terminal app is crazy. Anthropic is scrambling to stay competitive with Claude Code and while doing so they're introducing bloaty features that are a better fit for a standalone tool. You'd want to use voice input everywhere, right ogirardot.writizzy.com/p/good-softwar…

Claude Code: *ships voice mode* Claude Code Devs: "what should we ship as a shortcut for voice mode?" Claude Code Devs: "spacebar, I'm sure nobody uses that" ...

@anttikissa Looking good - what's the font btw?

The joys of building a coding agent in a coding agent: it's excited too!

@ruuts @aarondfrancis This gave me an idea #commit-to" target="_blank" rel="nofollow noopener">github.com/raine/git-surg…

@aarondfrancis I take back my answer. I often make changes that are not specific to the working branch so I want to commit it to main / master. Not being able to checkout the main branch in the worktree dir is headache. I guess worktrees is for people that work very linear and less cluttered.

Why do people like git worktrees over discrete checkouts? (This isn't bait, it's research)

@felixge Also, you can `[core]\nexcludesfile = ~/.gitignore` to your ~/.gitconfig which allows ~/.gitignore to be your global .git/info/exclude

.git/info/exclude is one of my favorite git features, but I'm always surprised when I talk to people who never heard about it. It basically allows you to add stuff to .gitignore without modifying the file. Great for ad-hoc ignoring some files created during debugging.

@hunvreus Worktrees and multiple agents writing code is the next level from this though. Would not go back to developing without

Most of the folks in my timeline bragging about their insane multi-agent setup running over night don't have anything to show for. No SaaS, no open source projects, nothing. I call bullshit. I still don't use sub-agents, git worktrees, multi-agent orchestration, Ralph loops or Claude Code. I just use the Codex desktop app. I have usually about 2 or 3 projects I'm actively working on. On any single project, I may have multiple threads, but only one that writes code, the others may do planning or investigation. If I really need to have 2 agents working on the same codebase, I just use multiple git clones.

1. Create a new worktree: `/worktree ` where agent figures out based on the context the appropriate prompt for a new agent, and launches a new tmux window where another agent starts off with the prompt in a new worktree 2. Once the agent in the worktree is finished and I've reviewed the result, I run `/merge` which will rebase on the base branch, resolve any conflicts taking into account what has changed in the base branch in the meantime, and finally merge and clean up the worktree. There are multiple of these running concurrently. Conflicts are not a problem because agents can figure out how to reconcile conflicting changes very well. This is my workflow in personal projects for last 3 months and has worked amazing. In work context, the difference is that instead of merging to main, I do `/open-pr` and clean up the worktree manually later.

For me, the orchestrator plus atomic task pattern made the difference. I am curious how others are handling this. Are you using an orchestrator model, or have you found a better way to coordinate multiple coding agents?

Ironically, my agentic coding workflow became too slow. One synchronous coding agent. Six phases. 20 minutes each. That is two hours of waiting. So I moved to the next step: a fleet of agents running in parallel.

@bygregorr @samuelcolvin @bcherny What might those environment variables be? Deploying? Did a human write this?

I see where you're coming from, but I'm not convinced a .gitignore file is the answer here. What if your settings are environment-specific, though? A .gitignore might inadvertently ignore some essential environment variables, which could lead to issues when deploying to different environments.

Why doesn't claude code add a `.claude/.gitignore` that ignores settings.local.json? @bcherny Surely you know about .gitignore? I see no downside and significant upside to this chanage. Why hasn't it happened for a year?

Somewhat annoying part of worktrees and Claude Code has been that you cannot resume or fork a session from a another worktree. claude-history now allows forking sessions from other projects (which worktrees technically are)

@aarondfrancis Because it allows working on arbitrary number of things in parallel and I don't have to worry about keeping track of number of checkouts. Worktrees need the right abstraction to be ergonomic. For terminal users, so far it's this one: github.com/raine/workmux

@shantanugoel I tried but Zellij's shortcut system is unnecessarily complicated. Could see why the learning curve is easier on Zellij, because the shortcuts are written in the UI, but it's hard justify switching for a life-long tmux user.

If you love tmux, you will love zellij even more. It's a modern alternative with much saner keybindings.