Welcome back. Anthropic is winning the AI talent war so decisively that its CEO now has an unusual problem. In a recent report, he voiced concerns that new hires are prioritizing paychecks over the company mission. In response, the dev community sparked a meme fest, but one senior engineer's ironic take cut deeper than the rest (read below). 

Also: Inside Uber’s AI Native Engineering team, a vibe-coded solution for data transfer using QR codes, and a Claude Skill that watches youtube videos and podcasts for you.

MiniMax's new video model runs on one consumer GPU: The Chinese AI lab just dropped H3, a 33B model capable of generating 2K video clips up to 15 seconds long, complete with synced stereo audio. While 2K regeneration still happens server-side, early tests indicate it can run locally on a single RTX 5090 using ComfyUI. This allows US-based teams to prototype without racking up massive cloud bills. The open weights are currently licensed for use in the US, EU, UK, and South Korea.

Anthropic's CEO fears engineers chase hefty salaries over mission: The AI lab reportedly outpays every rival. When an Axios report on the talent wars went public, the developer community was quick to jump in. Veteran engineer Gergely Orosz remarked that the mission-driven lab should try underpaying than its rivals and then see who shows up for the work. Meanwhile, others flooded the internet with memes mocking the company's concern. See how much Anthropic pays its engineers. 

Cursor plugs coding agents into Google Workspace: The AI code editor just shipped a suite of plugins that give its agents direct access to Gmail, Google Drive, Calendar, Docs, and Sheets. This means agents can now pull context straight from your inbox, draft or update files, and manage schedules without you ever having to leave the editor. On top of that, cloud agents now use tokens 20-30% more efficiently, helping engineering teams offload bigger tasks while staying under budget.

Not because you don't care. Because after 6 hours of building, typing paragraphs about what you built feels
like punishment.

So the PR description stays vague. The README gets "TODO." The architecture decision lives in your head until you leave the company.

Wispr Flow: syntax-aware voice dictation that works in every app. Speak your PR descriptions, commit messages, and docs. camelCase, snake_case, and acronyms stay intact.

Engineers at OpenAI and Vercel use it daily. Available on Mac, Windows, iPhone, and Android. The best docs are the ones that actually get written.

Try free

Product managers at Uber operate within systems too complex for any one person to fully grasp. Their blueprint, the PRD (Product Requirements Document), cycles through design, engineering, and several other departments before a feature ships. Uber recently shared how its team uses AI to review PRDs. It’s basically a version of an RFC review, where engineers vet a design doc to catch flaws before anyone writes code.

The usual suspects. PRDs often arrive with the same issues: shaky assumptions, blind spots around adjacent systems, or ideas another team already tested and dropped. The context to catch them sits scattered across docs, dashboards, and institutional memory. Reviews then become scavenger hunts. Teams rebuild lost context and surface problems that were cheaper to fix early, leaving slower decisions, uneven feedback, and rework that takes far too much time.

A better first pass. To combat this, Uber first ran a structured checkpoint process, but checkpoints only work as well as what goes into them. The team then asked, what if every PM had a fast first-pass reviewer before a PRD hit broader approval?

This reviewer can:

Why it matters. The reviewer widens a PM's field of view by offering a specific diagnosis. Helping later reviews run sharper and faster. Reviewing the doc is half the job. Drafting it and spinning up tickets is the next part of the grind.

P.S. This tutorial is a good place to start for engineering teams who want to learn to automate PRDs.

Senior engineer’s AI coding workflow that works: You’ll learn a structured AI-assisted coding workflow using spec files, roadmaps, automated testing, and mandatory manual review gates. This tutorial shows how setting strict repo guidelines and using automated tools for code reviews prevents token waste while producing clean, reliable software.

Kimi-k3-in-c (1.3K ⭐): You can now run Moonshot AI’s Kimi K3, a 2.78 trillion parameter model, on a single CPU with just 8.24 GB of RAM.

How to design effective prompts (by Qwen): Vague or unorganized prompts often lead to inconsistent and inaccurate AI responses. To fix this, use this prompting guide that gives you clear constraints, step-by-step reasoning, and a structured framework.

When a long Codex run drifts, it's tempting to fire off constant corrections mid-run. But each minor adjustment clutter the context and often makes things worse.

An engineer at OpenAI suggests a better approach: interrupt the run once and provide a single, detailed course correction. The Codex CLI even includes specific commands to handle this.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

A Vercel engineer just used 60 billion tokens to perfect an AGENTS(.)md file that manages long-running agent tasks without any issues. Go ahead and grab it.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team