AI Dev Essentials #29: Claude Memory & Skills, OpenAI Acquisitions, and new Vercel OSS

John Lindquist
Instructor

John Lindquist

Hey Everyone 👋,

John Lindquist here with the 29th issue of AI Dev Essentials!

Claude Memory rolled out to Pro users, providing project-scoped context that keeps work streams separate. OpenAI acquired the team behind Apple Shortcuts, which suggests they're thinking seriously about OS-level integration. Vercel shipped significant updates to the AI SDK and Agent workflows. Google AI Studio added annotation mode for visual debugging, and Claude Code launched on the web with sandbox security features. Every AI tool is getting better and better at doing the boring stuff so that we can focus on the fun stuff.

Personally, I've been diving deep into the new Claude's new "Skills" feature which it taking the AI space by storm. Skills enable you to enhance you Agents with new capabilities based on your specific needs. I'll definitely be covering Claude's Skills in my upcoming Claude Code Power User workshop:

Sign up for the Nov 7th "Claude Code Power User" workshop here: https://egghead.io/workshop/claude-code

🎓 New egghead.io Lessons This Week

I launched "The Essential Guide to Claude Code Skills" which is a free course that covers the basics of Skills and how to use them to enhance your Agents. Lessons 1-3 are live and I'll be adding more lessons ~daily:

  1. Create Your First Claude Code Skill
  2. Control Claude Skills Output with References and Examples
  3. Stacking Claude Skills to Create Complex Workflows

Check out the course here: https://egghead.io/lessons/create-your-first-claude-code-skill~sds93

🚀 Major Announcements

Claude Memory Rolls Out to Pro and Max Users

Anthropic announced on October 23, 2025, that Claude Memory is now available to Pro and Max plan users, bringing project-scoped context management to individual subscribers after initially launching for Team and Enterprise users in September.

Memory capabilities:

  • Project-Scoped Memory: Each project maintains its own focused context, preventing cross-contamination between different work streams
  • Immediate Access: Max users get immediate access; Pro users receiving rollout "over the coming days"
  • User Control: View, edit, and manually update memory summaries at any time
  • Import/Export: Transfer memories to/from other AI tools for backup or migration
  • Privacy Options: Opt-in system with incognito chat mode for conversations that don't save to memory
  • Safety Testing: Extensive safety testing conducted to prevent harmful pattern reinforcement

(Anthropic Official, MacRumors, Axios)

Memories leaking across projects is the reason I never turn memories on. I've had too many experiences where it would suggest TypeScript in a Swift project or even mention my deadlifts when planning a vacation. It just felt creepy and invasive. Scoping memories sounds like it will finally give me a level of control to actually make it useful.

OpenAI Acquires Software Applications Inc and Apple Shortcuts Team

OpenAI announced on October 23, 2025, the acquisition of Software Applications Incorporated, maker of Sky—a natural language interface for Mac—along with the founding team behind Apple's Shortcuts automation framework.

Acquisition details:

  • Sky Product: Natural language interface for Mac that understands screen context and executes actions
  • Key Team Members: Ari Weinstein and Conrad Kramer, former Apple employees who created Workflow (which became Apple Shortcuts)
  • Design Talent: Kim Beverett, 10-year Apple designer who worked on Safari, Messages, FaceTime, WebKit, and user privacy
  • Valuation: Deal valued in multibillion-dollar range; specific terms not disclosed
  • Previous Funding: Software Applications had raised $6.5 million; Sam Altman's investment fund held passive stake
  • Timing: Acquisition announced 2 days after OpenAI launched ChatGPT Atlas (AI-powered browser for Mac)
  • Strategic Goal: Integrate Sky's deep macOS integration and product craftsmanship into ChatGPT ecosystem

(OpenAI Official, TechCrunch, CNBC, MacRumors)

Automation specialists + OpenAI Agents working on Mac... I've always loved the general idea of the Shortcuts app on both iOS and Mac, but the execution and UI has always felt clunky. I'm curious to see what Automation/Agent design/UI will look like when they get around to releasing the tools. I'm already deeply aware of the capabilties of Mac automation (thanks to my work on Script Kit), but I hope this isn't just another connect-the-nodes UI.

Anthropic and Google Announce Multi-Billion Dollar Compute Deal

Google and Anthropic announced on October 23, 2025, a major cloud infrastructure deal that will provide Anthropic with up to 1 million custom-designed Tensor Processing Unit (TPU) chips to support training and deployment of next-generation Claude models.

Partnership details:

  • Scale: Up to 1 million TPU chips worth tens of billions of dollars (industry estimates: ~$50B for 1-gigawatt data center)
  • Compute Capacity: Over 1 gigawatt of AI compute capacity expected online in 2026
  • Purpose: Support training and deployment of next-generation Claude models
  • Rationale: TPUs selected for "price-performance and efficiency"
  • Business Context: Anthropic's annual revenue run rate approaching $7 billion; Claude powers 300,000+ businesses (300x increase over past 2 years)
  • Infrastructure Strategy: Builds on initial 2023 partnership; ensures capacity to meet exponentially growing demand
  • Multi-Cloud Approach: Amazon remains primary cloud provider and training partner with $8 billion investment (more than double Google's $3 billion)

(Google Cloud Press, Bloomberg, CNBC, SiliconANGLE)

It's interesting how OpenAI and Grok always talk about needed more compute, but then Google (with Gemini) can lease compute to direct competitors. I don't really have a good read on the situation and requirements for what these teams are doing, but it makes me think that Google is in a much better position than the other providers.

Vercel v0 Launches iOS App in Public Beta

Vercel announced in October 2025 the public beta launch of the v0 iOS app, extending AI-powered development capabilities to mobile devices with voice and camera inputs.

App features:

  • Mobile Development: Build apps using voice and camera inputs on iOS
  • Waitlist Access: Users can join waitlist for mobile beta access
  • October Updates: Team Templates, Stripe payment integration (public beta), Claude Sonnet 4.5 rollout
  • Cross-Platform Strategy: Expansion extends v0 beyond web development to mobile platforms
  • Usage: Enables on-the-go prototyping and development workflow

(Vercel Community, Vercel Blog)

I'll be honest, I switched to Android for all the deep Gemini integrations. It's fine and I've got high hopes for it, but seeing OpenAI only releasing Codex on iOS, Anthropic only releasing Claude Code on iOS, and now Vercel only releasing v0 on iOS is making me sad. I'm super curious to see if my non-AI friends will be able to use v0 to build an app for themselves. Being able to launch it on a phone removes a ton of friction from the casual user to get building. This is a super interesting space to watch over the next year to see how far this can be pushed to letting anyone build whatever they need.

🛠️ Developer Tooling Updates

Vercel Open Sources Workflow Dev Kit for Durable Workflows

Vercel released the Workflow Dev Kit on October 23, 2025, as an open-source TypeScript framework for building durable, long-running workflows with built-in observability and no vendor lock-in.

Framework features:

  • Simple API: Uses use workflow and use step directives with no manual queue/scheduler setup
  • Durable Execution: Functions can pause for minutes or months, survive crashes, and resume exactly where they stopped
  • Built-in Observability: CLI and web UI for monitoring workflow execution
  • Webhook Support: External data triggers for workflow initiation
  • Framework Agnostic: Runs on any framework, platform, and runtime
  • Target Use Cases: AI agents and data pipelines requiring reliable long-running operations
  • Open Source: MIT licensed and available on GitHub

(Vercel Blog, Vercel Changelog, GitHub)

I don't know how many times I can say "Agents!" in a single newsletter, apparently they're kind of a big deal. Everyone is figuring out how to host them and Vercel is throwing its hat in the ring. I'm definitely all for anything that makes these easier to understand and manage. I'm almost suprised it's not just a text prompt describing the workflow so that it's a "true" Agent, but I understand that comes with massive limitations and unexpected results.

Google AI Studio Adds Annotation Mode for Visual Debugging

Google announced on October 18, 2025, the addition of annotation mode to Google AI Studio, enabling developers to highlight and comment on UI components directly for precise visual debugging.

Annotation capabilities:

  • Visual Highlighting: Click and mark specific components, buttons, or text on the canvas
  • Collaborative Comments: Add visible comments, error pointers, and highlights for team collaboration
  • Screenshot Sharing: Annotated screenshots shareable in chat for reference context
  • Precise Targeting: Enable exact UI region specifications for Gemini model adjustments
  • Vibe Coding Integration: Part of broader "vibe coding" mode revamp with context-aware suggestions
  • One-Click Deployment: Deploy directly to Google Cloud Run from AI Studio

(Google Developers Blog, Testing Catalog)

Awesome feature! Absolutely love this direction of just busting out a red pen and writing all over what you want changed rather than trying to describe it in words. I need to show this to my young kids and see what they make.

Claude Code Launches on Web with Parallel Task Execution

Anthropic announced on October 20, 2025, that Claude Code is now available on the web as a beta research preview for Claude Pro and Max users, enabling parallel coding task delegation on cloud infrastructure.

Web platform features:

  • Cloud Infrastructure: Delegates multiple coding tasks simultaneously on Anthropic-managed infrastructure
  • GitHub Integration: Connects to GitHub repositories for seamless code management
  • Real-Time Progress: Track progress with ability to steer Claude during execution
  • Parallel Execution: Support for running tasks across different repositories simultaneously
  • Automatic PRs: Creates pull requests with comprehensive change summaries
  • Growth Metrics: 10x user growth since May 2025 launch; generates $500 million annualized revenue
  • Mobile Support: iOS app available in early preview

(Anthropic Official, TechCrunch, VentureBeat)

Claude Code actions have been available in GitHub actions for a long time, so I was curious when they were going to release their own hosted version. I haven't had a chance to poke at it yet, but it's definitely next on my list.

Claude Code Introduces Sandbox Security Model

Anthropic launched sandboxing for Claude Code on October 20, 2025, implementing a two-boundary isolation system for filesystem and network access with OS-level enforcement.

Security features:

  • Filesystem Isolation: Claude can only access and modify specified directories
  • Network Isolation: Restricted to approved servers via proxy service
  • OS-Level Enforcement: Uses Linux bubblewrap and macOS seatbelt for security
  • Reduced Friction: 84% reduction in permission prompts in internal testing
  • Open Source: Sandboxing approach open-sourced for other teams building AI agents
  • Easy Activation: Run /sandbox command in Claude Code to enable
  • Prompt Injection Protection: Protects against malicious prompt injection attempts

(Anthropic Engineering, WinBuzzer, Claude Docs)

Permissions are one of those areas of AI Agent development that are a pain to manage. You have a lot of files on your system you don't want AI's to read, but you also want your Agent to run autonomously and not have to constantly approve permissions. This is a big step in the right direction.

Gemini CLI Enhances Interactive Shell Experience

Google released Gemini CLI versions 0.9.0 (October 15) and 0.10.0 (October 22), 2025, introducing pseudo-terminal support and interactive shell tool calling for complex command execution.

CLI improvements:

  • PTY Support: Version 0.9.0 introduced pseudo-terminal for interactive command execution
  • Interactive Tool Calling: Version 0.10.0 added intelligent shell tool usage without explicit "!" prefix
  • Complex Commands: Run vim, nano, git rebase -i, top, htop, npm init, and interactive REPLs
  • Real-Time Visualization: Colors and cursor position maintained within Gemini CLI context
  • Bidirectional Communication: Keystrokes and window resizing transmitted to processes
  • Installation: npm install -g @google/gemini-cli@latest

(Google Developers Blog, Testing Catalog, GitHub)

I'm curious when Gemini/Claude/Codex will become a default shell. I know the Warp terminal already takes this kind of approach, but my guess is that in the near future, every command you run will be part of an AI session that can help fix things along the way.

🤖 AI Ecosystem Updates

Google Veo 3.1 Tops Video Arena Leaderboard

Google announced on October 15, 2025, the release of Veo 3.1 with enhanced capabilities, which subsequently topped both Text-to-Video and Image-to-Video leaderboards on Video Arena on October 20.

Model improvements:

  • Video Generation: Up to 60 seconds of 1080p video generation
  • Native Audio: Enhanced audio generation including dialogue, sound effects, and ambient audio
  • Improved Fidelity: Better prompt adherence and texture realism
  • Availability: Gemini App, Flow, Gemini API, Vertex AI
  • Pricing: $0.40/second standard tier, $0.15/second fast tier
  • Leaderboard Performance: First model to break 1400 in Video Arena history

(Google Blog, TechCrunch, LMArena)

Alibaba Qwen Releases Qwen3-VL Vision Models

Alibaba's Qwen team released Qwen3-VL-2B and Qwen3-VL-32B on October 21, 2025, bringing state-of-the-art vision-language capabilities to both edge devices and cloud deployment.

Model variants:

  • Qwen3-VL-2B: Optimized for edge devices (smartphones, laptops)
  • Qwen3-VL-32B: Cloud deployment with performance matching models up to 235B parameters
  • Context Window: Up to 1 million tokens
  • OCR Support: 32 languages with strong multilingual capabilities
  • Model Modes: Both Instruct and Thinking variants available
  • Deployment Options: FP8 quantized versions for efficient deployment
  • Benchmark Performance: Qwen3-VL-32B reportedly outperforms GPT-5 Mini and Claude 4 Sonnet on STEM benchmarks (MathVista, MMMU)

(GitHub, Alibaba Cloud, TechNode)

Microsoft Launches Mico Avatar for Copilot Mobile

Microsoft introduced Mico, an expressive blob-shaped avatar, at the Microsoft Fall 2025 Copilot Sessions on October 23, 2025, for Copilot on iOS and Android.

Avatar features:

  • Expressive Design: Blob-shaped avatar with shape-shifting animations and emotive reactions
  • Optional Feature: Can be disabled; includes Easter egg transformation to Clippy when tapped repeatedly
  • Voice Integration: Displays during voice conversations with dynamic reactions
  • Availability: US launch with UK and Canada rollout in coming weeks
  • Additional Features: Part of broader fall release including group chat and debate modes

(TechCrunch, Windows Central)

Anthropic Launches Claude for Life Sciences

Anthropic announced Claude for Life Sciences on October 20, 2025, built on Claude Sonnet 4.5 with specialized integrations for research and pharmaceutical development.

Life sciences features:

  • Model Performance: Scores 0.83 vs 0.79 human baseline on Protocol QA benchmark
  • Platform Integrations: Benchling, PubMed, 10x Genomics, Synapse.org
  • Use Cases: Literature reviews, hypothesis development, data analysis, regulatory submission drafting
  • Real-World Impact: Novo Nordisk reduced clinical study documentation from 10+ weeks to 10 minutes
  • Availability: Claude.com and AWS Marketplace; Google Cloud Marketplace coming soon
  • Strategic Focus: Formal entry into life sciences sector with domain-specific optimizations

(Anthropic Official, CNBC)

⚡ Quick Updates

OpenRouter Launches Exacto Endpoints for Precision Tool Calling

  • Announced: October 21, 2025
  • Precision Routing: Route requests to providers with measurably superior tool-calling accuracy
  • Access Syntax: model_slug:exacto (e.g., glm-4.6:exacto)
  • Launch Models: Kimi K2, DeepSeek v3.1 Terminus, GLM 4.6, GPT-OSS 120B, Qwen3 Coder
  • Selection Criteria: Top accuracy, normal propensity, low user blacklisting
  • Benchmarks: Shows materially fewer tool-calling failures on tau2-Bench and LiveMCPBench

(OpenRouter, Hacker News)

Google Releases ADK v1.17.0 for Agent Development

  • Released: October 22, 2025 via PyPI
  • Framework: Open-source Python toolkit for building, evaluating, and deploying AI agents
  • Agent Types: LLM agents, workflow agents (Sequential, Parallel, Loop), custom agents
  • Design Philosophy: Make agent development feel like traditional software development
  • Model Agnostic: Works with any LLM provider
  • Installation: pip install google-adk

(Google ADK Docs, Google Cloud)

Google Research Publishes FLAME for Rapid Model Specialization

  • Published: October 23, 2025
  • Approach: One-step active learning for model specialization
  • Speed: Complete specialization in ~1 minute on standard CPU (no GPU required)
  • Performance: 53.96% AP on DOTA benchmark, 53.21% on DIOR benchmark with 30-shot adaptation
  • Key Innovation: Lightweight classifier refinement layer on top of frozen base model
  • Use Cases: Rapid adaptation for specialized object detection tasks

(MarkTechPost, arXiv)

Google Launches Skills Learning Platform

  • Launched: October 21, 2025
  • Content: Nearly 3,000 courses, labs, and credentials consolidated into single platform
  • Sources: Google Cloud, Google DeepMind, Grow with Google, Google for Education
  • Developer Features: Gemini Code Assist integration in 20+ hands-on labs
  • Gamification: Achievement streaks, leagues, shareable badges
  • Pricing: 35 monthly lab credits for developers; many courses free; paid certifications available
  • Career Integration: 150+ employer partnerships including direct interview pathways

(Google Blog, Google Cloud Blog)

Framecap CLI Screen Recorder Launches for macOS

  • Launched: October 13, 2025 (v1.0.0); October 22, 2025 (v1.1.1)
  • Size: Lightweight 1.2 MB native binary with no external dependencies
  • Features: Multi-track recording (screen, camera, audio to separate files)
  • Codecs: Lossless H.264 and ProRes formats
  • Special Modes: Window-only recording with transparent backgrounds
  • Performance: Handles multiple 5K sources simultaneously
  • Pricing: $49 USD for lifetime updates

(Framecap Official, Hacker News)

Research: Self-Improving Language Models Survey Published

  • Published: October 3, 2025 (submitted to EMNLP 2025)
  • Focus: Self-improvement in Multimodal Large Language Models
  • Structure: Data collection, data organization, model optimization
  • Key Finding: Recent advances enhance capabilities without significant cost increases
  • Related Work: Self-Evolving LLMs paper shows 53.8% relative improvement with less than 1.2% additional data
  • Paradigm: Models generate and solve their own tasks for data-efficient learning

(arXiv Survey, arXiv Self-Evolving)

✨ Workshop Spotlight

Transform into a Claude Code Power User

Date: November 07, 2025 Time: 9:00 AM - 2:00 PM (PDT) Platform: Zoom

Pricing:

What You'll Learn:

Master the essential skills to ship reliable AI-generated code with confidence. This hands-on workshop covers everything from foundational prompting to advanced automation using the Claude Code SDK and custom integrations.

Core Skills:

  • Prompt for Developers: Learn to communicate effectively with Claude for coding tasks
  • Context Engineering: Control what context Claude sees for reliable, consistent results
  • TypeScript SDK: Script Claude programmatically to build custom workflows
  • Custom Hooks: Automate repetitive tasks with Claude Code hooks
  • Model Context Protocol: Integrate APIs securely to extend Claude's capabilities
  • Live Q&A: Get your specific questions answered by John Lindquist

Register: https://egghead.io/workshop/claude-code

Read this far? Share "AI Dev Essentials" with a friend! - https://egghead.io/newsletters/ai-dev-essentials

  • John Lindquist

https://egghead.io

Share with a coworker