Another Wednesday, another selection of AI news and resources to help you become more AI native. This week:

  • Anthropic locks up SpaceX's Memphis supercomputer and doubles Claude Code limits

  • OpenAI's Codex moves into your Chrome tabs

  • Voice agents finally got a brain with GPT-Realtime-2

  • The one-line prompt that forces any AI to audit its own work

Next Gen

Your moat is now a markdown file.

What's Happening: Three of the biggest AI newsletters this week landed on the same thesis from different angles. The companies pulling ahead are not picking better models or buying more seats. They are writing down what their best operators know in plain English so their agents can read it. Anthropic's keynote pushed Skills, Outcomes, and overnight Dreaming so agents can build their own memory. OpenAI's Codex Chrome extension lets agents inhabit your actual logged-in apps. The connective tissue is documented context, not model choice.

Why It Matters: Most leadership teams are still budgeting AI like a vendor RFP, picking platforms and counting seats. The teams winning figured out that the platform is mostly interchangeable and the real asset is the library of skill files describing how their best people actually work. Without it, every agent you deploy shows up like a brilliant new hire with no onboarding doc and burns weeks of trust. With it, you can swap models, vendors, or file formats five times in two years and the institutional knowledge ports forward intact.

In the Wild: Browserbase's Autobrowse graduates each successful browser-agent task into a reusable skill the next agent loads instead of re-deriving the work. Cost and time on a Craigslist task dropped roughly in half by iteration four. Matt Van Horn shipped 30+ atomic agent-native CLIs covering Linear, ESPN, and Flight GOAT, plus a factory that mints new ones on demand.

Looking Ahead: Pick one workflow your team runs constantly, the one where someone keeps saying "I wish AI could just do this." Have the person who runs it best spend a Friday afternoon writing down in plain English what they do and the edge cases that matter. That is your first skill. Do that five times across five workflows in a month and you have a starter library any major AI platform can load tomorrow.

AI First

The 100% Confident Prompt

A one-line prompt loop that forces any AI to audit its own work instead of nodding along.

LLMs are trained to please you, so most plans and code reviews come back with the suspicious vibe of "yes, this is great." This prompt flips the model into self-audit mode.

  1. Use it on any plan, strategy, code review, or decision you just got from an AI.

  2. Paste it as a follow-up turn after the model's first answer.

  3. Run it 2 to 3 times. The model finds tighter holes each cycle.

  4. Stop when the model says it's actually confident or the fixes get nitpicky.

  5. The prompt: "Are you 100% confident in this strategy? If not, find all possible loopholes, suggest proper fixes, and run this loop until you are factually 100% confident."

Pro tip: The word "factually" is doing the work. Without it the model vibe-checks itself. With it the model grounds the answer the way it would for a fact-check.

AI News

Anthropic signed SpaceX's Memphis supercomputer and doubled Claude Code's limits.

At Code with Claude SF, Anthropic announced full access to Colossus 1, the 300+ MW Memphis data center formerly run by xAI, locking in 220,000+ Nvidia GPUs coming online within the month. Five-hour Claude Code limits doubled across Pro, Max, Team, and Enterprise, peak-hour throttling is gone, and Opus API limits went up considerably. Anthropic also floated multi-gigawatt orbital data centers with SpaceX. The catch: Claude itself briefly went down mid-keynote.

OpenAI's Codex moved into your Chrome tabs.

OpenAI shipped a Codex Chrome extension that runs inside your live, authenticated browser sessions instead of analyzing screenshots from the outside. It can drive Salesforce, Gmail, and internal dashboards using your existing cookies, parallel-task across tabs in its own isolated tab group, and respond to prompts like "@Chrome open Salesforce and update the account from these call notes." Mac and Windows only at launch, not yet available in the EU or UK.

Voice agents finally got a brain.

OpenAI shipped three new voice models led by GPT-Realtime-2, which jumped from 81.4% to 96.6% on Big Bench Audio and from 34.7% to 48.5% on Audio MultiChallenge versus the previous reference model. Context grew from 32K to 128K tokens. The trick: short conversational preambles like "let me check that for you" play while reasoning runs in the background, hiding the silence that used to expose voice AI as AI. Zillow launched voice home search and Deutsche Telekom deployed translated voice support across 14 markets the same day.

That's it for this week. See you next Friday.

- Cam

PS: Get a personalized one-page report that tells you which 5 AI tools you should implement right now.

Use code "AI40" for 40% OFF

Got 3 seconds to spare?

We want to make the newsletter as valuable as possible for you... Click one of the links below to let us know how we did.

Login or Subscribe to participate

Keep Reading