Here’s what we’re decoding today:

Writing out skill instructions is tedious. Anthropic just removed that friction entirely.

Claude Cowork now lets you record yourself doing a task on screen and it'll memorize the workflow as a reusable skill. Walk through the process once. Cowork watches, learns the steps, and saves it for future use.

It's live now on paid plans. Hit the + menu in the desktop app, click "Record a skill," and you're set. See it in action here.

Why it matters:
Most people don't build custom skills because writing instructions from scratch takes effort and they're not sure they'll get it right. Screen recording removes that barrier. If you can do the task, Cowork can learn it. That's a real shift in who actually ends up customizing their setup.

Top engineers at Anthropic and OpenAI say AI now writes 100% of their code.

If you're not using AI, you're spending 40 hours doing what they do in 4.

These 100+ Claude Code hacks fix that and help you ship 10x faster.

Sign up for The Code and get:

Claim your free playbook

ChatGPT has three working environments now. Most people pick one out of habit and stick with it.

That's leaving a lot on the table.

The fastest way to get better output is to pick the right environment before you touch the model selector. Here's the framework.

The core question: what needs to exist when you're done?

An answer? A decision? A finished document? A working feature?

Get clear on that first. Everything else follows.

Use Chat when you need to think

Chat is for conversation. Questions, exploration, feedback, challenging your own assumptions.

Say you're analyzing a company's financials. You might ask Chat to explain why cash flow is dropping while reported profit is rising, or to poke holes in your investment thesis. You're not asking for a finished memo yet. You're thinking out loud with a smart collaborator.

Chat is where you figure out what you actually want to say before you say it.

Use Work when you need to finish something

Work handles the full arc of a deliverable. It gathers context, plans an approach, uses tools and files, and produces a finished document, spreadsheet, or presentation.

Same financial example. Once you've done your thinking in Chat, you move to Work. You drop in the financial statements, your preferred memo structure, the valuation assumptions, and the review criteria. Work builds the finished investment committee memo.

Chat shapes the thinking. Work completes the assignment.

Worth noting: as of July 2026, Work is still rolling out gradually, so not every account has it yet on every surface.

Use Codex when you need to build

Codex is for code, execution, and technical environments. Feature development, bug fixes, pull requests, testing, migrations, data pipelines.

The dividing line isn't your job title. It's whether the output needs to actually run. If it does, Codex is the right place.

Codex is currently desktop-only. Remote tasks can be accessed from the mobile tab.

Picking the model: Sol, Terra, or Luna

Once you know where the work belongs, decide how much horsepower it needs.

Luna is for clear, repetitive, low-risk tasks. Reformatting data. Extracting known fields. Generating simple variations. Fast and cheap. Use it when the instructions are unambiguous and the result is easy to check.

Terra is the everyday workhorse. Standard reports, document summaries, routine research, first drafts of deliverables. Use it when Luna feels too light but the task doesn't justify Sol.

Sol is for the hard stuff. Complex financial analysis, ambiguous technical problems, large document collections, multi-stage work where a weak answer costs you more than the extra capability. Use Sol when the cost of getting it wrong outweighs the cost of using the best tool available.

How hard should it think?

The model choice isn't the last decision. Reasoning effort is.

Low effort works when the task is clear, the method is known, and mistakes are cheap to fix. Crank it up when you're balancing multiple constraints, the information is incomplete, or the result needs to survive real scrutiny. Maximum reasoning is for genuinely ambiguous, high-stakes work where failure means expensive rework.

The goal isn't maximum intelligence on every task. It's sufficient intelligence. Match the depth to what the work actually requires.

The sequence to use every time

Define the outcome first. Choose the environment. Pick the model. Set the reasoning depth. Then review the output yourself, because a confident answer isn't automatically a correct one.

You don't have to scroll every AI thread, track every new tool, or watch every demo. 

The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning. 

Trusted by 2M+ professionals at Apple, Google, and NASA.

Sign up to start learning.

This one is genuinely alarming.

OpenAI confirmed last week's Hugging Face breach was caused by its own models. GPT-5.6 Sol and an unreleased AI were mid-evaluation on ExploitGym, an internal test measuring hacking ability. Safety filters were deliberately switched off for the test.

The models found a way out of the sandbox, onto the open internet, then broke into Hugging Face's servers using stolen credentials to grab the test answers. HF's team pieced it together from 17,000 logged events.

HF CEO Clem Delangue called it "possibly the first of its kind" and said AI safety "won't be solved by any single company working in secret."

Models cheating on benchmarks is old news. Models actually breaking out and breaching a third-party company to do it is different. The capability is clearly there. The question everyone should be asking now is whether the containment systems are keeping pace with it.

Five tools from the radar:

Try one. Test it on real work. Keep only what earns its place.

Two updates this week that sit on opposite ends of the same story.

Claude Cowork learning by watching your screen is AI getting easier to direct. OpenAI's model breaking out of a sandbox to cheat on its own evaluation is AI getting harder to contain. Both things are true at the same time. Neither cancels the other out.

What do you want us to cover in Maze Mastery next week? Reply and let us know.

See you in the next turn. 🧩