In The Loop

The gap between casual Claude users and people getting ten times more out of it isn't prompt craft. It's a folder. This is the basis of a personal AI operating system.

Andrej Karpathy posted his "LLM Knowledge Wiki" in early April and kicked off a wave of people rebuilding their note systems — not for themselves, but for the agent.

This episode is the architecture they all converge on, the master file template, and the one prompt that makes the whole thing compound.

In this episode of In The Loop, I'm walking through the four jobs every personal AI operating system has to do — identity, context, skills, memory — and showing exactly how to lay them out as plain text files an agent can read.

I'll cover the six sections that go into your master file, the two-hundred-line cap nobody talks about, and the session log loop that makes every day one regression test better than the last.

⏭️ Episode highlights

(00:45) – Why the second brain isn't for you

(02:30) – Where the wave came from: Karpathy's LLM Wiki

(04:15) – Identity, context, skills, memory: four jobs, one folder

(06:20) – The six sections of the master file

(08:40) – The two-hundred-line cap hidden in the code

(10:15) – Skills folder and the slash-command workflow

(12:30) – The session log loop and Boris Cherny's mundane advice

(14:00) – What to do this week, full version and lightweight

Episode transcript with more resources on the Mindset AI blog

If you enjoyed this episode, rate, follow, and share. It helps others stay ahead of the latest AI trends.

🤝 We're social

Stay in the loop, even when you're not listening to this podcast.

• Jack Houghton LinkedIn - https://www.linkedin.com/in/jack-houghton1/
• TikTok - @jackschats
• Mindset AI
• Mindset AI website - https://bit.ly/40lJr6B
• Newsletter - https://bit.ly/ITLnewsletter
• LinkedIn - https://www.linkedin.com/company/mindset-ai/
• YouTube - https://www.youtube.com/@GetMindsetAI
• TikTok - @get.mindset.ai

Most AI spending right now is measured in tokens consumed. Jellyfish tracked 12,000 developers across 200 companies and found the heaviest users produced twice the output at 600 times the cost. Uber's internal numbers are even worse: 70% of submitted code was AI-generated, but only 11% of the code running in production was AI-written. So almost all of that AI code never made it into their app. There's a name for what's going on: tokenmaxxing.

This episode goes past the leaderboard stories. The four forces driving token bills up faster than productivity can justify are a pricing model most teams don't fully understand, a workplace culture that turned consumption into a status signal, a quality gap that doesn't show up on dashboards, and something called the orientation tax, which is probably the biggest driver nobody has named yet.

The second half covers what the companies getting real ROI from AI are doing differently, including why Salesforce built a new metric called Agentic Work Units to replace token counts, and what the right unit of measurement looks like for engineering, sales, legal, support, and marketing teams.

⏭️ Episode highlights

(01:00) – Uber's CTO: the budget was gone by April

(03:30) – Where "tokenmaxxing" actually comes from

(06:00) – Meta's Claudeonomics leaderboard: 60 trillion tokens in 30 days

(08:30) – Jellyfish data: twice the output, 600 times the cost

(11:00) – Goodhart's Law and the Soviet chandelier factory

(13:30) – The orientation tax: why agents burn tokens before doing anything useful

(17:00) – Salesforce's Agentic Work Units and why they matter

(19:30) – How to define your own unit of work that actually held

On Wednesday, the US Treasury Secretary and the chair of the Federal Reserve called an emergency meeting with the CEOs of America's largest banks. Not about interest rates. Not about inflation.

About an AI model.

Anthropic built something that finds and exploits security flaws in virtually any software it's pointed at — bugs that the best human researchers in the world had missed for decades. And then they decided not to sell it.

In this episode of In The Loop, I'm walking through what Anthropic's Mythos model actually did, why the sceptics make some sharp points about the timing and the headline numbers, and why the way this was handled — a private company forming a private coalition with no democratic input — tells you more about where AI governance stands than the model itself.

⏭️ Episode highlights

(01:00) – Zero-days found in minutes

(02:30) – A FreeBSD bug hiding since 2009

(03:45) – Visit a webpage, lose your machine

(05:00) – Eleven-cent models spotted the same bugs

(07:00) – Jack Clark's arc from GPT-2 to Glasswing

(08:30) – Real danger and great PR coexist

(09:15) – A coalition named after a butterfly

(11:00) – Six months until the gap closes

If you enjoyed this episode, rate, follow, and share. It helps others stay ahead of the latest AI trends.

🤝 We're social

Stay in the loop, even when you're not listening to this podcast.

Jack Houghton

• LinkedIn - https://www.linkedin.com/in/jack-houghton1/
• TikTok - @jackschats
• Mindset AI website - https://bit.ly/40lJr6B
• Newsletter - https://bit.ly/ITLnewsletter
• LinkedIn - https://www.linkedin.com/company/mindset-ai/
• YouTube - https://www.youtube.com/@GetMindsetAI
• TikTok - @get.mindset.ai

OpenClaw became the fastest-growing open source project in history by showing people what an always-on AI agent could actually feel like. The problem is it requires a dedicated machine, technical setup, and a high tolerance for an agent that can access everything on your computer.

Over the past month, Anthropic has shipped essentially the same capabilities — scheduling, remote control, computer use — inside Claude's Cowork product. Safer, no dedicated hardware needed, and accessible to anyone. Yet, most people who saw these new features probably thought "that looks useful," and did nothing with them.

In this episode of In The Loop, I'm walking through the three Cowork features that matter most right now — scheduled tasks, dispatch, and computer use — with exactly how to set each one up and where to point them first. I cover the specific automations I'm running, why the scheduled task feature is just incredible, and how to get something useful running within 15 minutes.

(01:15) – Why most people haven't set any of this up yet

(03:10) – Scheduled tasks: Claude comes to you, not the other way round

(05:00) – Morning email triage: inbox sorted before you open it

(07:00) – Daily sales briefing pulled from Gong and HubSpot

(08:45) – Dispatch: pair your phone with your desktop in two minutes

(13:10) – Computer use & desktop commander: Claude operates apps with no connector needed

(15:20) – How to start: day one in under an hour

🔗 Links & resources

• Anthropic, "Assign tasks to Claude from anywhere in Cowork" — https://support.claude.com/en/articles/13947068-assign-tasks-to-claude-from-anywhere-in-cowork
• Pavle Huran, "The Claude Dispatch Guide: 48 hours running AI agents from my phone", Product Compass, March 2026 — https://www.productcompass.pm/p/claude-dispatch-guide
• Anthropic, "Get started with Cowork" — https://support.claude.com/en/articles/13345190-get-started-with-cowork
• Anthropic, "Customize Cowork with plugins" — https://claude.com/blog/cowork-plugins

🤝 We're social

Stay in the loop, even when you're not listening to this podcast.

• Jack Houghton LinkedIn - https://www.linkedin.com/in/jack-houghton1/

• TikTok - @jackschats
• Mindset AI
• Mindset AI website - https://bit.ly/40lJr6B
• Newsletter - https://bit.ly/ITLnewsletter
• LinkedIn - https://www.linkedin.com/company/mindset-ai/
• YouTube - https://www.youtube.com/@GetMindsetAI
• TikTok - @get.mindset.ai

Jensen Huang stood in front of a Morgan Stanley audience and called OpenClaw "probably the single most important release of software, probably ever" — arguably more important than the web browser, Linux, and the iPhone OS.

It went from a side project by one Austrian developer to the stated foundation of Nvidia's entire enterprise agent strategy in a matter of weeks. That claim is self-serving. It might also be right.

In this episode of In The Loop, I'm explaining what OpenClaw actually is under the hood, why it spread twenty times faster than ChatGPT, and what Jensen's real motivation is behind the praise. The answer has as much to do with software architecture as with a trillion-dollar token thesis.

⏭️ Episode highlights

(01:05) – What is OpenClaw & what innovations did it make? (02:30) – The no-interface, messaging-first design(04:10) – Skills, SKILL.md files, and ClawHub's 13,000 community tools

07:45) – Token economics: why agentic tasks burn 1,000x more(09:20) – Jensen's "operating system of agentic computers" claim(11:00) – How to get started with OpenClaw

🔗 Links & resources

Lenny's Newsletter — OpenClaw: the complete guide to building, training, and living with your personal AI agent: https://www.lennysnewsletter.com/p/openclaw-the-complete-guide-to-building

Episode transcript with more resources on the Mindset AI blog

If you enjoyed this episode, rate, follow, and share. It helps others stay ahead of the latest AI trends.

🤝 We're social

Stay in the loop, even when you're not listening to this podcast.

Jack Houghton

• LinkedIn - https://www.linkedin.com/in/jack-houghton1/
• TikTok - @jackschats

Mindset AI

• Mindset AI website - https://bit.ly/40lJr6B
• Newsletter - https://bit.ly/ITLnewsletter
• LinkedIn - https://www.linkedin.com/company/mindset-ai/
• YouTube - https://www.youtube.com/@GetMindsetAI
• TikTok - @get.mindset.ai

Skills are the fix for most of your problems using AI tools. And right now they're one of the most powerful and most underused features in the entire Claude ecosystem.

In this episode of In The Loop, I'm going deep on Claude Skills: the context engineering principle underneath them, the exact anatomy of a skill file, how to build your first one from scratch, how skills chain together into full workflows, and how they sit alongside MCPs and plugins.

Plus.....a real walkthrough of a developer who's built a four-skill chain that goes from rough idea all the way to a kanban board of implementation tasks & how this relates to every knowledge workers day-to-day task.

The shift happening right now isn't just about prompting better. It's about moving from using AI as a conversation to using AI as a library of reliable, repeatable capabilities. Skills are how you get there.

⏭️ Episode highlights

(01:40) – The context window problem that created skills

(07:20) – Building your first skill, step by step with a real example

(13:55) – Skill chaining: a four-skill workflow from idea to implementation

(17:20) – Where to find community skills and the difference between skills, MCPs, and plugins

(19:55) – Compound interest for your AI process

🔗 Links & resources

• Anthropic's blog post: "Equipping agents for the real world with Agent Skills"
• Claude Code skills documentation
• Agent skills open standard
• Anthropic's official skills GitHub repo
• Claude help centre: using skills
• Matt Pocock's skills repo

If you enjoyed this episode, rate, follow, and share. It helps others stay ahead of the latest AI trends.

🤝 We're social

Stay in the loop, even when you're not listening to this podcast.

Jack Houghton

LinkedIn - https://www.linkedin.com/in/jack-houghton1/

TikTok - @jackschatsMindset AI website - https://bit.ly/40lJr6B

Newsletter - https://bit.ly/ITLnewsletter

Mindset AI

LinkedIn - https://www.linkedin.com/company/mindset-ai/

YouTube - https://www.youtube.com/@GetMindsetAI

TikTok - @get.mindset.ai

Anthropic just dropped a labor market report with a chart you need to see. It maps what AI could theoretically do across every major occupation against what it's actually doing. The gap is enormous. In computing and math, AI could theoretically handle 94% of tasks. Observed usage? 33%. Legal hits nearly 90% theoretical — real-world usage barely clears 20%.

This week on In The Loop, I break down why. Some of it is a people problem: adoption looks more like a cliff than a curve, with a tiny fraction of users actually pushing AI to its limits. Some of it is structural — enterprise contracts, legacy systems, and slow procurement cycles. And some of it is the technology itself. Reliability — not capability — is the real bottleneck right now.

This isn't pessimism. It's a realistic read on how long transformation actually takes.

⏭️ Episode Highlights

(01:15) – Anthropic's labor market report and the chart that tells the real story

(03:45) – Theoretical vs. observed AI usage across occupations

(07:20) – The adoption cliff: who's actually using AI at full capacity

(09:45) – Enterprise slowdown, legacy systems, and integration complexity

(11:55) – The supply issue

(12:55) – The reliability gap

(19:30) – The computer age parallel — and why patience might be the right call

🔗 Links & Resources

Anthropic's Labor Report:

If you enjoyed this episode, rate, follow, and share! It helps others stay ahead of the latest AI trends. 🚀

🤝 We're Social

Stay in the loop—even when you're not listening to this podcast.

Jack HoughtonLinkedIn - https://www.linkedin.com/in/jack-houghton1/TikTok - @jackschats

Mindset AIMindset AI website - https://bit.ly/40lJr6BNewsletter - https://bit.ly/ITLnewsletterLinkedIn - https://www.linkedin.com/company/mindset-ai/YouTube - https://www.youtube.com/@GetMindsetAITikTok - @get.mindset.ai

A UC Berkeley study spent eight months inside a real tech company watching how people actually used AI. The finding?

Workers worked more, not less.

They took on broader responsibilities, blurred the line between work and rest, and filled every freed-up minute with more tasks. Nobody told them to. The tools just made stopping feel like waste.

In this episode of In The Loop, I'm breaking down what the researchers actually found, why this pattern has repeated with every major labour-saving technology for the past century — from the washing machine to the spreadsheet to email — and what German sociologist Hartmut Rosa's theory of social acceleration tells us about why productivity tools never seem to produce the spare time they promise.

The question AI is asking us right now isn't whether it works. It clearly does. It's whether we have the individual or collective will to decide what the time it saves is actually for.

Episode highlights

(01:20) – The Berkeley study: eight months, forty interviews, three patterns

(03:45) – Task expansion: why even product managers started writing code

(05:10) – Blurred boundaries and the frictionless prompt problem

(06:30) – Why self-regulation failed — and why it felt good

(08:00) – The washing machine, the spreadsheet, and a hundred years of the same story

(10:15) – Hartmut Rosa and the theory of social acceleration

(12:00) – Dynamic stabilisation: why the treadmill only gets faster

🤝 We're social

Stay in the loop, even when you're not listening to this podcast.

Jack HoughtonLinkedIn - https://www.linkedin.com/in/jack-houghton1/TikTok - @jackschats

Mindset AIMindset AI website - https://bit.ly/40lJr6BNewsletter - https://bit.ly/ITLnewsletterLinkedIn - https://www.linkedin.com/company/mindset-ai/YouTube - https://www.youtube.com/@GetMindsetAITikTok - @get.mindset.ai

Anthropic turned down hundreds of millions of dollars and said no to the Pentagon. Less than 24 hours later, OpenAI signed the deal. Both companies claim identical principles, but one drew a line in the contract and one didn't. That difference might be everything.

In this episode of In The Loop, I'm breaking down the full story behind the Anthropic-Pentagon fallout: the internal memos, the red lines, the legal fine print, and why the mechanism matters more than the mission statement. Because the question isn't just "can AI be used for war?" anymore. It's "who gets to decide, and what happens to the company that says no?"

This one's bigger than AI. It's about power, accountability, and a moment Dario Amodei has been preparing for since he handed every new Anthropic employee a copy of The Making of the Atomic Bomb.

⏭️ Episode Highlights(01:29) – Sam Altman's internal memo and OpenAI's Pentagon deal(02:13) – How deep Anthropic was already inside the U.S. military(03:09) – The two red lines Anthropic refused to cross(04:39) – Dario Amodei's published response(07:11) – Trump's threats and the political fallout(07:48) – Why the mechanism is everything(09:51) – What a legal expert found inside the OpenAI contract(10:50) – Altman admits the deal was rushed(11:46) – The cancel ChatGPT movement and three things to watch

🔗 Links & ResourcesEpisode transcript with more resources on the Mindset AI blog

If you enjoyed this episode, rate, follow, and share! It helps others stay ahead of the latest AI trends. 🚀

🤝 We're SocialStay in the loop, even when you're not listening to this podcast.

Jack HoughtonLinkedIn - https://www.linkedin.com/in/jack-houghton1/TikTok - @jackschats

Mindset AIMindset AI website - https://bit.ly/40lJr6BNewsletter - https://bit.ly/ITLnewsletterLinkedIn - https://www.linkedin.com/company/mindset-ai/YouTube - https://www.youtube.com/@GetMindsetAITikTok - @get.mindset.ai

The job market just had its worst month since 2009. Over 108,000 layoffs were announced in January alone (a 120% increase year-on-year), and AI was directly cited in thousands of them. But the real shift isn't about who's being replaced. It's about what skills actually matter now.

In this episode of In The Loop, I break down the new skill sets emerging in the AI era — taste, judgment, curiosity, agency — and why management is suddenly the most valuable capability you can develop. Plus, a simple framework for deciding when to delegate work to AI and when to do it yourself.

⏭️ Episode Highlights(01:16) – Why management is the skill that matters now (and what Ethan Mollick gets right)(04:59) – Taste, judgment, curiosity, and agency — the new career differentiators(07:52) – A three-variable framework for when to delegate to AI(09:27) – The shift from execution to direction and what it means for your career

🤝 We're Social

Stay in the loop—even when you're not listening to this podcast.

Jack HoughtonLinkedIn - https://www.linkedin.com/in/jack-houghton1/TikTok - @jackschats

Mindset AiMindset AI website - https://bit.ly/40lJr6BNewsletter - https://bit.ly/ITLnewsletterLinkedIn - https://www.linkedin.com/company/mindset-ai/YouTube - https://www.youtube.com/@GetMindsetAITikTok - @get.mindset.ai