Effort levels split the room; Karpathy ships in 2 hours
Theo moved Opus 5.5 from extra high down to high and says it is better almost all the time, while swyx says go lower still and add a critique loop. Nobody has an eval, just defaults.
The big disagreement
Theo dialed Opus 5.5 down to high, swyx says go lower still
Theo said running Opus 5.5 on high beats extra high almost all the time, reversing his own setup after Julius pushed him. swyx went further, arguing low effort plus a critique loop beats maxing the dial for anyone throughput-constrained, and noted Karpathy shipped his weekend project on medium. None of the three posted an eval behind the choice.
having strong opinions on effort levels is like the tabs vs spaces of our time
Receipts
Karpathy built a YouTube filter from 43M HN comments in two hours
Andrej Karpathy had Claude filter 43 million Hacker News comments over 20 years down to 6,454 recommended videos, in roughly two hours with Opus 5.5 on medium effort. His own manual work was creating a Google Cloud API key in the browser, booting a VPS and buying the domain; he gave Claude the IP and it deployed and hardened the box.
I only talk to Claude and have it do everything
How they actually work
dax's team sends architecture questions to a Gemini subagent
dax says that whenever his team hits a tricky architectural problem, a complex system design question or needs to simplify something, they ask for a Gemini subagent, and that it is good there. He added that Gemini's reputation is bad enough that the post drew more accusations of being paid than anything he has written.
gemini's reputation is such that i've never received more accusations of being paid to post this
Ball doubled down on pushing to main after developers balked
Thorsten Ball said the replies to his no-branches post surprised him: many developers cannot imagine working without long-lived branches, or assume pushing to main means shipping commits that break CI. He and sqs recorded answers to the two questions that kept coming back, how you avoid slop without mandatory review and how you make big architectural changes, under the heading that culture stops slop.
Seems like a lot of developers genuinely can't even imagine working without long-lived branches.
Model and agent watch
Anthropic pushed a server-side rule blocking clones of specific games
Claude Code picked up a rule around Oct 27 sent from Anthropic's servers under a flag called "tengu_heron_brook", with no setting to turn it off, telling the model not to recreate a specific existing game. Addy Osmani replied that two sentences were added to Claude Code's prompt on Opus 5.5 because using Claude to infringe IP is against usage policy, and that original and genre games should keep working.
Another reason why you wanna use your own harness and not one where companies change what you can or can’t do at will.
Open tabs
Hashimoto takes apart a rival terminal's speed claim
Mitchell Hashimoto dug into a competing terminal's benchmark claim and found the read loop overwrites a 64 KiB ring, discarding everything older unparsed per tick.
OpenCode can now use your Claude subscription
dax announced that your Claude subscription works in OpenCode. Theo said he did not see it coming and called it Anthropic loosening its grip.
Frequently asked questions
Should I max out reasoning effort on Claude Opus 5.5?
No. Developers shipping the most work are turning reasoning effort down instead of maxing it. Theo found that high effort beats extra high almost all the time, swyx argued that low effort plus a critique loop (a process where the model reviews its own work) beats maxing the dial when throughput matters, and Karpathy shipped his project on medium effort.
Can Claude deploy and secure infrastructure without human help?
Almost. When Karpathy filtered 43 million comments down to 6,454 video recommendations, the only human steps left were the ones that required a browser login, like creating an API key. Claude handled deploying the VPS and hardening the box on its own.
Can I use different AI models for different tasks in the same session?
Yes. dax's team routes tricky architectural problems to a Gemini subagent (a specialized AI component) while staying in their main session, rather than switching their entire workflow to a different model.
Does Anthropic change Claude's rules without telling me?
Yes. Around October 27, Anthropic pushed a server-side rule that prevents Claude Code from recreating specific existing games, and there is no setting to turn it off. The change was added to Claude's system prompt (the instructions the model runs on) without a release announcement.
Can I use my Claude subscription in other coding tools?
Yes. Your Claude subscription now works in OpenCode, which Theo described as Anthropic loosening its grip on where Claude can be used.