Welcome to the Weekly Stack, serving up the hottest AI developments fresh off the griddle for PM builders. We’re adding a ludicrously tall stack of 1,987 new subscribers this week! Welcome to the stack!! 🥞

Where did all these people come from? The answer is also the reason I missed the last send. We'll get to that shortly.

But man, was it a crazy few weeks to miss! While September was a relatively slow month, October more than made up for it. Better models, new tools, and (several) completely new products.

Here's what we'll cover this week:

Let's do this.

Claude Code is one of my favorite products of all time. The immense, wonderful power of LLMs, unshackled from the browser, unleashed directly onto our machines. I literally live in the terminal almost all day now and I think many are going to follow suit.

It started when I went deep preparing for my Claude Code overview on Aakash Gupta's Product Growth Podcast. People loved the episode, and it's now sitting at number 5 of his most popular podcasts ever, edging out even legends like Marty Cagen and Claire Vo.

So I set out to build a complete guide to Claude Code for PMs to cover things we didn't have time for. As I was researching, inspiration struck – what if I created a guide to Claude Code... IN Claude Code?!

And thus, ccforpms.com was born!

It's built specifically for product managers:
📚 COMPLETE guide to mastering Claude Code
🤖 Build custom subagents for any task
🧠 Create a thinking partner to write PRDs
📊 Analyze your data and build dashboards
📑 Craft a strategy and turn it directly into slides

It was my most successful launch of anything ever, with more than 4,000 people asking for access across LinkedIn and X.

Please check it out! It’s 100% free. I'm still gathering feedback to improve it, so if you go through it, let me know what you think.

If you already have Claude Code installed (installation instructions), running this command in your terminal will launch you directly into the first module:

cd ~/Documents && \curl -L https://github.com/carlvellotti/claude-code-pm-course/releases/latest/download/complete-course.zip -o course.zip && \unzip course.zip -d claude-code-course && \cd claude-code-course && \claude "/start-1-1"

This kind of stuff – deep, tactical, fundamentally hands-on – is what the market needs and what I want to keep building. A lot more is on the way.

So, if I miss future sends, trust that I’m cooking.

My goal with this newsletter is explicitly not to cover everything in the AI building space. There are too many companies shipping too fast, and realistically a lot of it's hype.

That said, October was packed with relevant releases. So we're breaking the rules just this once to make sure you're aware of what shipped and why each matters.

Sora 2 – OpenAI's upgraded video generation model with improved physics, synchronized audio, and videos up to 25 seconds. Reality is officially broken.

Claude Sonnet 4.5 – Anthropic's new flagship model that ran autonomously for 30 hours building a production app (up from 7 hours with Opus 4). If you're using Claude Code, this is what's powering it now – same price as Sonnet 4 but significantly more capable at sustained work.

Claude Haiku 4.5 – The small, fast model that somehow scores 73% on SWE-bench Verified while costing one-third of Sonnet 4 and running twice as fast. Now free for all Claude.ai users, making near-frontier AI accessible to everyone.

Claude Code on Web – Claude Code now runs in your browser with cloud sandboxes, meaning you can assign multiple coding tasks that run in parallel without opening a terminal. Game-changer for PMs who want to experiment without local setup.

Grok 4 Fast – xAI's reasoning model that's 40% more efficient than Grok 4 with a 2M token context window, now free for all users. The web search integration is particularly useful for building apps that need real-time information.

ChatGPT Apps – OpenAI's new app ecosystem with embedded UI components from Spotify, Figma, Zillow and more, letting you interact with services visually inside the chat. Different from Custom GPTs – these are about building interfaces, not just instructions.

ChatGPT Atlas – OpenAI's AI-powered browser with ChatGPT built in, memory of your browsing context, and an agent mode for handling tasks like booking flights. Direct challenge to Chrome, launching first on Mac with agent features for Plus/Pro subscribers.

Perplexity Comet – The first truly agentic browser, previously $200/month, now completely free. It doesn't just answer questions – it actively manages your tabs, emails, and calendar, and can navigate the web autonomously on your behalf.

Gemini 2.5 Computer Use – Google's model that can click, type, and scroll through UIs like a human, outperforming alternatives on web and mobile benchmarks with lower latency. Available via API for building UI agents that automate workflows.

I will be experimenting with all this and sharing my insights in coming weeks.

DeepSeek just pulled off a counterintuitive trick: an AI model that stores text as images to compress context windows by up to 20x.

Here's the counterintuitive trick – a page of text might take 2,000-5,000 text tokens, but the same page rendered as an image only takes 200-400 vision tokens. Their model can read these image-compressed documents with 97% accuracy at 10x compression.

Context windows have been the invisible ceiling on what AI can handle – you hit limits analyzing long documents, reviewing codebases, or maintaining conversation history. DeepSeek's "vision-text compression" suggests a different future: compress older context at lower resolution (like human memory fading) so AI can handle dramatically longer contexts without exploding costs.

The researchers mention context windows spanning tens of millions of tokens becoming practical. That means analyzing entire product specs in one go, feeding AI your full codebase, or maintaining much longer agent conversations. It's open source, so expect this technique to spread fast.

Something shifted in the past few months, and I don't think enough people have noticed yet.

When "vibe coding" emerged earlier this year, it was part joke, part aspiration. Engineers could sketch ideas faster, but production-ready? Not quite. The skepticism was warranted.

But we've quietly crossed a critical threshold. The models got incrementally better with each release. The tooling matured. The techniques solidified. And now engineers are starting to say they aren't writing any code – just carefully planning what they want and reviewing what AI builds.

Three things converged:

1. Testing became automated – AI can now write comprehensive tests, not just implementation. This was the missing piece. You can verify functionality without manual QA.

2. Design systems got smarter – Tools like Storybook integration mean AI can build components that actually match your existing patterns, not just generic UIs.

3. Models hit the "good enough" bar – Sonnet 4.5 running for 30 hours autonomously isn't a gimmick. Haiku 4.5 scoring 73% on SWE-bench at one-third the cost isn't incremental. These are different capabilities.

You and I can now build apps closer and closer to production-ready, if we're willing to level up our skills. The space is infinitely deep, but many PMs are only a few techniques away from building extremely cool shit.

I've started feeling this myself. After learning how to have AI write tests and set up a Storybook design system, I can now set up AI to one-shot entire features.

Learn these things with me and we're about to accomplish more than we could have ever dreamed of before. The threshold is behind us now. Time to build.

This classic popped up again on the feed. Soon we’ll be able to do both 😈

If you only have 30 minutes this week, check out ccforpms.com and get started!

If you're on a Mac – enter this command into you terminal and it will install Node, Claude, and launch you directly into the course! (Must already have a Claude account)

curl -fsSL https://raw.githubusercontent.com/carlvellotti/claude-code-pm-course/main/scripts/quick-setup.sh | bash

Keep building,

Carl