The AI Friday Brief
Good Wednesday, NOLA. Claude Code just got a major upgrade with Routines — reusable workflows that could change how you use AI for coding. Meanwhile, Google's rolling out one-click AI prompts in Chrome, and there's a wave of new tools hitting the market to turn AI into actual product builders (not just chatbots). Plus: someone built Claude Code for Wall Street, and we're seeing real conversations about what AI can and can't actually do.
Claude Code Level Up
Claude Code Routines: Save and Reuse Your Best Workflows
Claude just shipped Routines — the ability to package up your most useful Claude Code workflows and save them for reuse. This is a big deal for anyone who runs the same kinds of analysis, refactoring, or data munging repeatedly. Instead of starting from scratch each time, you build it once, save it, and call it back whenever you need it. Think of it as turning your best prompts into templates.
Anthropic
LangAlpha: What If Claude Code Was Built for Finance?
A developer just shipped LangAlpha, which is basically Claude Code for Wall Street — financial modeling, backtesting, and data analysis workflows tailored to traders and quants. It's a great example of what happens when you take a general-purpose AI coding tool and narrow it down to a specific domain. If you work in finance and have been waiting for something like this, it's worth a look.
Show HN
Tools People Are Building
Google Chrome Gets One-Click AI Skills
Google dropped a new feature in Chrome: the ability to turn your favorite AI prompts into one-click tools accessible from your browser. No API keys, no setup — save a prompt, get a button. It's low-friction enough that non-technical folks could actually use this. Early example: summarize a webpage, generate social copy, or extract structured data. The bet is that lowering the friction makes AI tools stick.
Google
Kelet: Root Cause Analysis Agent for Your LLM Apps
If you're running AI agents or LLM-powered apps in production, Kelet is worth checking out — it's a tool that helps you debug when things go wrong. It traces what your agent did, where it got confused, and what went sideways. As more people ship AI products, debugging tools like this are becoming table stakes.
Show HN
HoloTab: Your AI Browser Companion
HoloTab is a lightweight browser companion that runs AI models locally while you browse. Use it to summarize articles, extract info, or chat with content on the page. It's the kind of always-on, contextual AI assistant that could actually change how people work online — if it stays fast and doesn't drain your battery.
Hugging Face
What AI Can Actually Do (And What It Can't)
The AI Revolution in Math Has Arrived
Quanta Magazine digs into one of the clearest wins for AI right now: automated theorem proving and mathematical discovery. This isn't hype — researchers are actually using AI to find new proofs and solve problems that were stuck for years. If you want a concrete example of what frontier AI is genuinely good at (abstract, formal reasoning in a rule-based domain), read this. We covered this on HN too.
Quanta Magazine
Can Claude Fly a Plane? (Spoiler: No)
Someone actually tested whether Claude could safely operate an aircraft simulator. The answer is no — it crashed. But the experiment is less about "AI is dumb" and more about understanding where the real boundaries are. Claude can read and reason about instructions, but it struggles with real-time feedback loops and tight coupling between action and consequence. Worth reading if you care about what your AI can actually do in live systems.
Hacker News
N-Day-Bench: Can AI Find Real Vulnerabilities in Real Code?
Security researchers built a benchmark specifically for this: can AI find actual zero-day vulnerabilities in real codebases? It's a harder problem than it sounds. The results are mixed — sometimes AI spots things, sometimes it misses obvious issues. This is the kind of honest testing that cuts through the hype and tells you what AI is actually good for in security work.
Hacker News
Big Moves & Business Angles
Anthropic's Rise Is Rattling Some OpenAI Investors
OpenAI's recent valuation (north of $800B) is starting to look expensive to some of its own investors, especially as Anthropic gains ground. One backer told the FT that justifying the price requires assuming an IPO at $1.2T or higher — a hard sell when Claude is closing the capability gap. This is the kind of investor skepticism that could reshape the industry.
TechCrunch
Multi-Agent Development Is a Distributed Systems Problem
This post reframes how we should think about AI agents working together. Instead of thinking of it as a coordination problem for agents, think of it as a classic distributed systems challenge: state consistency, message ordering, failure modes. If you're building anything with multiple AI agents, this mental model will save you headaches.
Hacker News
Real-World AI Workflows
BrightBean Studio Built a Social Media Tool in 3 Weeks Using Claude
A small team shipped a full social media management tool in just three weeks, using Claude Code and LLM agents to handle the heavy lifting. No massive team, no months of engineering. It's a great case study in what's actually possible when you lean hard into AI for the boring parts and focus your human effort on product and design. The repo is public if you want to see how they built it.
GitHub
My AI-Assisted Workflow: A Real Person's Real Setup
Maio Barbero walks through their actual, day-to-day AI workflow — not theory, not best practices, but what actually works for one developer building real things. It's the kind of "I tried this and here's what stuck" post that's worth more than a dozen "AI best practices" guides. If you're trying to figure out where AI fits into your own work, this is a template.
Hacker News
Also
- Why AI Sucks at Front End (And What That Means) — A clear-eyed breakdown of where AI coding tools actually struggle. Worth reading if you're shipping UI work.
- How Top AI Agent Benchmarks Get Gamed (And What Comes Next) — Berkeley researchers explain why many AI benchmarks don't reflect real-world performance — and what honest measurement actually looks like.
- Claude Code May Be Burning Your Token Limits Invisibly — If you're using Claude Code heavily, this is a heads-up: token usage can compound faster than you'd expect.
- GAIA: Open-Source Framework for AI Agents on Local Hardware — AMD's new framework for running AI agents on consumer hardware. Useful if you're trying to keep things local.
- Building Trust in the AI Era with Privacy-Led UX — A design philosophy for AI products that treats transparency and user control as core features, not afterthoughts.
- An AI Vibe Coding Horror Story — A developer's cautionary tale about using AI for code without understanding what it's doing. Raw and honest.
- The Human Cost of 10x: How AI Is Physically Breaking Senior Engineers — An uncomfortable read about the real toll of being expected to ship 10x faster with AI tools.
- Microsoft Exec Suggests AI Agents Will Need to Buy Software Licenses — A fascinating business angle: what happens to licensing when AI agents are autonomous customers?
- To Teach in the Era of ChatGPT Is to Know Pain — Teachers are reckoning with AI in real time. Less about the AI, more about what it means for education.
- Claudraband: Claude Code for Power Users — An open-source wrapper around Claude Code that adds some quality-of-life improvements. Check it out if you're using Code daily.
- Latent Space: Top Local Models List — April 2026 — A curated list of the best open-source models you can run locally right now. Updated monthly.
Pass it on