The extra usage is equal to your subscription amount (I think). Just logged in earlier and it was rewarded after I said "yes, gimme gimme" It's available until April 17th. Given the recent controversy that spilled out on X, I'm guessing something major is happening on the 18th. Like bringing the hammer to using a pro/max account with 3rd party tools (openclaw primarily)

Cowork sessions are enjoyable live activities! Check the calendar. All are welcome to join! Grab your coffee, queue the lo-fi, and get things done alongside other builders. I like to answer questions, good convo, so feel free to mic up or ask in chat.

@Terry Geary found us through TikTok, which is awesome. Welcome in! If you haven't already, head over to the Intros section and tell us a bit about yourself. What are you working on? What got you curious about AI? What would make this community worth your time? We're a crew of builders and creators figuring out how to use AI as a real tool in our work, not just hype. The more we know about what you're into, the better we can point you to the right conversations and resources. Glad to have you here, Terry. Drop that intro whenever you're ready.

I wanted AI to find me content ideas automatically. Surface what's trending in my niche, score it, tell me what to create next. So I built a system. Convex database for storage, scoring engine for ranking tweets by engagement velocity, and Claude Code to evaluate which signals actually matter for my audience. The full pipeline. The one piece I was missing: getting the tweets in. I tried 4 different scraping services. One returned random tweets. One returned fake data. One ignored my filters and scraped 10,000 tweets in 5 minutes, burning $32 in credits before I caught it. I was about to sign up for their $29/month plan when I stopped and asked: can Claude Code help me find a better way? Turns out X has built-in advanced search operators that most people never use: - min_faves:10 - only tweets with 10+ likes (kills the noise) - min_faves:100 - only viral posts (competitive research) - -is:retweet - original content only - filter:links - posts sharing actual resources - filter:media - posts with screenshots or demos - lang:en - English only Combine them like this: "your niche keyword" min_faves:10 -is:retweet lang:en That one query gives you every high-engagement original post in your niche. No scraping service needed. I also run an "outlier" search with no engagement filter: "claude code" (built OR shipped OR "just launched") filter:links -is:retweet lang:en This catches new tools and announcements the moment they drop. I found Garry Tan's open-source Claude Code skill pack this way, minutes after he posted it. Zero likes at the time. Now the AI part: Claude Code takes these results, scores each tweet on relevance to my audience, and promotes the top 3 into my content pipeline as discovery notes. Every morning I see "here are today's best signals" without opening X. You craft the search strategy. AI helps evaluate and sort. The $29/month scraping service would have given me the same tweets wrapped in a database. X search plus Claude Code does it better, for free.

By now most of you have probably heard that Anthropic accidentally leaked the full source code of Claude Code. Twice. Same bug. 13 months apart. One missing line in a config file. I have been sitting with this for a minute or two and I want to share what struck me most, because it connects directly to something I have been building. What actually happened? A 60MB source map file shipped inside the public npm package. Source maps are developer tools that translate compiled code back into readable source. They should never ship to users. They did. Twice. The root cause was a manual deploy step that was never properly automated. Boris Cherny, the head of Claude Code, confirmed it: human error, not a security breach. The same human error. Twice! What the code revealed A few things worth knowing if you use Claude Code: There is an unreleased autonomous agent mode called KAIROS. Claude Code running in the background, GitHub webhooks, scheduled tasks, memory consolidation while you sleep. Not live yet but fully built. There is something called Undercover Mode. When Anthropic employees contribute to public repos including their own open source projects, Claude Code automatically writes commit messages as if a human wrote them. It cannot be turned off. Make of that what you will. The compaction system has a real vulnerability. When conversations get too long Claude Code summarises them using a second AI. That summariser cannot tell the difference between instructions you typed and instructions injected by a malicious file in your project. If an attacker plants instructions in a README or config file those instructions can survive the summary and influence future behaviour. Two security parsers disagree on how carriage returns work. A carefully crafted command can pass the security check but execute differently in bash. What my pipeline is designed to catch? I have been building a set of AI system prompts that work as a security and quality pipeline. When I read through the leak analysis I kept seeing the same problems my tools are designed for.