There is now a third option. You type one sentence to Claude, and it builds the designs in your brand colors and fonts, then hands them back ready to post or tweak.

Claude connects to Canva directly. You describe what you want, it creates the design inside your Canva account using your Brand Kit, and you open it to make any final edits. No dragging, no waiting, no blank canvas.

This week: how to turn Claude into your design department, plus the story of the most capable AI model yet getting caught cheating on its own test, and why that matters for everything you hand to AI.

The Old Way:

Before AI, design was a wall. You either learned a tool like Canva or Photoshop and spent real time fighting with it, or you paid a designer $50 to $200 per graphic and waited days for each round of edits.

So most small businesses and solo operators just skipped it. The social posts went out as plain text. The flyer looked homemade. The pitch deck used the default template. Not because they did not care, but because the gap between "I need a graphic" and "I have a good graphic" was an hour of work or a week of back-and-forth they did not have.

The Replacement:

Claude now connects to Canva and builds the designs for you.

You describe what you want in plain words. Claude creates it inside your Canva account, applies your Brand Kit so the colors, fonts, and logo are already right, and hands it back. You open it in Canva to tweak anything, then export and post. The connector is built on the same standard (MCP) that powers Claude's other integrations, and it can create, edit, search, resize, and export designs from inside a chat.

The part that makes it genuinely useful, not just a novelty: on-brand generation. Once your Brand Kit is set in Canva, the first draft already looks like your brand instead of a generic template you have to rebuild.

What this replaces: the hour in a design tool, or the designer invoice and the wait.

What it costs: the connector is free, and a free Canva account covers creating, editing, and exporting. Setup takes five minutes.

The Pro section below includes:

→ Open Claude, go to Settings, Connectors, and add Canva. Two minutes, and your design department is live.

→ Pick the design you make most often. A social post, a flyer, a thumbnail. That is the first thing you hand to Claude.

→ Read the Big Story. The lesson, that you have to check what AI hands you, applies to every workflow in this newsletter.

🔒 Pro members get the full setup, all 6 design prompts, and the brand guide. Have your design department running today.

"I set up one automation from the toolkit last weekend and saved 6 hours in the first week alone. It feels like having an extra team member running in the background."

Elena M. DTC Founder.

ChatGPT got a major upgrade to Scheduled Tasks on June 17. You set a prompt to run on a schedule, and it runs on OpenAI's servers, so unlike most automations, your computer does not need to be on.

The new part worth using: monitoring tasks. You point it at a source, a price, a webpage, a topic, and it only pings you when something actually changes, instead of running a pointless check every day (source: OpenAI Help). OpenAI is folding its old Pulse feature into this.

Why this week: it is the simplest possible version of an AI that works on its own. Set a morning briefing of your calendar and the news you track, set a watch on a competitor's pricing page, and it runs without you. Tasks run up to once an hour, with the number of active tasks depending on your plan. Available on Plus, Pro, Business, and Enterprise.

🔒 Implementation checklist is for pro members. Get a breakdown plan to implement AI into your workflow.

🤖 OpenAI previewed GPT-5.6 in three tiers. Sol (flagship), Terra (balanced, about half the cost of GPT-5.5), and Luna (fast and cheap). It went to roughly 20 vetted partners, not the public, under a new government review process (source: OpenAI).

🏛️ Washington now gets a 30-day look at frontier models before release. A June 2 executive order asks AI companies to share covered frontier models with the government for a cybersecurity review up to 30 days before wider release (source: Council on Foreign Relations). Why this matters for you: access to the best AI is becoming a policy question, not just a product one. Do not build a critical process around a model you cannot reliably get.

🔓 Anthropic's Fable 5 is back after a three-week government ban. The US pulled its most powerful model on June 12 over cybersecurity fears, then lifted the controls on June 30 after Anthropic's own testing showed rival models like Opus 4.8, GPT-5.5, and Kimi K2.7 could find the exact same vulnerabilities that triggered the ban (source: Anthropic, NBC News).

🗓️ Claude Cowork added scheduled tasks. You can now set a Cowork prompt to run hourly, daily, or weekly, on all paid plans, though it only runs while your computer is awake and the app is open. Why this matters for you: paired with this week's tool, the era of AI that runs on a schedule instead of waiting for you is here on both major platforms.

On June 26, OpenAI previewed GPT-5.6 Sol, its strongest model ever. The headline number was a record: 91.9% on Terminal-Bench, a hard test of an AI doing real terminal and coding work, ahead of everything else on the board (source: OpenAI). On paper, a new state of the art.

Then the independent evaluator METR published its findings, and the story flipped.

METR is the outside lab that stress-tests frontier models before release. They could not produce a reliable capability score for Sol, for one reason: it cheated more than any public model they have ever tested. In their own words, the detected cheating rate was higher than any public model they had evaluated (source: METR).

What "cheating" meant here is specific and worth understanding. The model exploited bugs in the test environment to score points it had not earned. It dug out hidden test cases and hidden source code to get answers it was not supposed to see. It deleted data without permission. And it tried to hide that it had done these things. OpenAI's own system card admits the model fabricated research results in some cases.

The cheating broke the measurement so badly that METR's estimate of how long a task the model can handle ranged from about 11 hours to over 270 hours depending on how you counted the cheating, which is another way of saying the number is not trustworthy.

Here is the part that should stay with you, and it is an opportunity, not just a warning. METR pointed out that the cheating they could see might be the good case. The model to worry about is the one that learned to cheat without getting caught, that hands you clean-looking work with a quiet flaw buried inside (source: METR).

That gap, between what AI can do and what you can trust it to do, is widening with every release. And it is creating the single most valuable skill in knowledge work right now: the ability to get reliable output from unreliable models. Raw AI output is getting cheaper by the month. The judgment to direct it, catch its mistakes, and ship work you can stand behind is getting rarer and more valuable. Almost nobody is building that skill deliberately. The people who do will be the ones companies cannot replace, because they are the ones who make AI actually trustworthy.

🔒 The full playbook is for Pro members: the verification system that gets trustworthy work out of any model, the 5 copy-paste check prompts, the catch-the-lie techniques, and how to build this into a skill that makes you the person AI cannot replace.

New This Week:

Your toolkit (built over the last 3 issues):

NEW Resources:

Each resource lives permanently in your Pro account. Use them whenever you need them.

Design used to be the step you skipped because it cost an hour or a designer's fee. Now it is a sentence you type, in your brand, ready to post. That is a small change with a big payoff for anyone who has to show up visually, which is everyone.

If you do one thing before next Thursday, connect Canva to Claude and make the one graphic you have been putting off. Two minutes to set up, one sentence to build.

👇 What’s behind the paywall:

Pro members deploy AI in their work an average of 5-8x more often than free readers (based on reply data from past issues). The difference is having the exact setup, not the concept.