Claude Ultra drops the five-hour usage cap
Hello, prompt wranglers, if Claude cuts you off mid-build while you still have weekly usage left, this edition explains who gets relief. You'll also find a worked example of turning community recommendations into an app, plus a prompt for moving your assistant setup.
In today’s edition
- Claude removes its five-hour cap for verified Ultra accounts.
- Karpathy explains how Claude built HackerTube from recommendations.
Claude Ultra drops its five-hour usage cap
On October 30, Claude began removing the rolling five-hour usage limit for verified Ultra accounts. That lets you concentrate more of your allowance into sustained building sessions instead of stopping while weekly usage remains.
- Max 20x is now called Ultra and still costs $200/month.
- The staged rollout starts with most existing subscribers. Newer accounts keep a temporary cap until verification, with an email confirming its removal.
- Weekly limits remain unchanged, so this changes when you can use your allowance, not how much you get each week.
What it means Try Ultra if short-window limits interrupt your coding work. Skip the upgrade if you mainly need a larger weekly allowance.
Source: claudeai
Build a video browser from community recommendations
Karpathy’s October 31 HackerTube write-up shows how he turned years of Hacker News recommendations into a working YouTube-style site. The useful lesson is how to separate bulk comment analysis from building the app.
- Opus downloaded and filtered the comments, then ran Sonnet in a Claude Code Workflow to remove spam and identify 40K comments recommending YouTube videos.
- The resulting collection contains 6,454 videos recommended from 2007–2026 across 150 topics.
- His manual work covered creating a Google Cloud API key for YouTube metadata, starting a VPS server, and buying the domain. Claude handled the code and deployment.
- The accompanying source post offers the analyzed data as a ZIP file, so you can ask Claude to build your own version.
What it means Start with Karpathy’s downloadable dataset if you want your own video browser, rather than repeating the full comment analysis.
Source: karpathy.bearblog.dev
Quick hits
- Google, Microsoft, and OpenAI plan to honor a shared robots.txt AI-training preference, with support arriving on different timelines rather than taking effect now.
- Anthropic’s automated red-teaming study found harmless-looking training examples that bypassed a monitor and induced misbehavior in deliberately fine-tuned research models.
- OpenAI’s Agents API update adds automatic system-prompt caching across sessions, background context compaction, command review, and tools you can change during a session.
AI workflow of the day
How to prepare your AI assistant setup for a move
Shared by Machina (@EXM7777) · View original post ↗
If you’re switching personal AI assistants, this prompt asks your current agent to package its memories, routines, and files for handoff. The goal is a documented transfer rather than rebuilding your setup from memory.
- Send the prompt below to your current personal AI assistant.
- Answer any questions it asks, then hand the resulting ZIP file to the new assistant as the rebuilding reference.
|
package my whole setup into one .zip file so a different AI agent can rebuild it exactly
include:
- who i am, my preferences and how i like you to talk
- everything you remember about me and my work
- every skill, workflow, routine and scheduled task, with exact steps
- my custom instructions and system prompts, word for word
- every tool and integration you use and how it's set up (never passwords or API keys)
- the files and templates my workflows depend on
put a README at the top that tells the new agent how to rebuild everything, step by step
if you're unsure about anything, ask me before you zip it
|