Good Morning! Here is today’s breakdown:
Microsoft ships computers-using agents
OpenAI Codex can now run your Mac while it is locked
AI now beats the average person on creativity tests
How to make AI-generated text sound actually human
4 new AI tools worth trying
AI AGENTS
Microsoft has shipped its biggest Copilot Studio update of the year, taking computer-using agents to general availability across every commercial region. These agents drive software the same way a person does, by reading the screen, clicking buttons, and typing into fields, which means they can finally automate the legacy internal tools that never had an API.
Computer-using agents are now generally available in all commercial Power Platform geographies, so any maker can build one without waiting for a vendor to expose an API.
Uses vision and reasoning instead of brittle, selector-based scripts, so the agent adapts when a layout shifts, a field moves, or a workflow branches.
Ships a redesigned workflows canvas with conditional branching, parallel execution paths, and a step-by-step debugging console that traces every agent action.
Adds real-time voice with sub-500ms latency, opening up customer service and IT helpdesk use cases that were previously too slow for voice.
Logs every session to Microsoft Purview and Dataverse, with configurable human-in-the-loop checkpoints for steps that need approval.
Most companies run dozens of critical internal tools that were never built for automation. By making vision-based agents generally available and wrapping them in audit logs and approval gates, Microsoft turns the messy long tail of enterprise software into something an agent can actually operate. This is the moment "computer use" stops being a research demo and becomes a production feature any team can deploy.
AI RESEARCH
OpenAI has expanded Codex with a set of upgrades that push it further toward a true background worker, including the ability to control a Mac even when the machine is locked. The days of leaving a laptop open and awake just to keep an agent running are ending.
Remote, locked computer use lets Codex keep operating on a Mac after you have stepped away and locked it.
Goal mode is now generally available across the Codex app, IDE extension, and CLI, so you can define an outcome and let Codex keep working toward it.
Appshots on macOS let you attach any app window to a Codex thread with a hotkey, sending a screenshot and its text for instant context.
Plugin sharing lets teams reuse locally built plugins across a workspace, turning one person's tooling into shared infrastructure.
The friction in agentic work has never really been intelligence, it has been babysitting. Every capability here removes a reason a human has to stay in the loop: locked-machine execution removes the open-laptop tax, goal mode removes constant re-prompting, and plugin sharing removes redundant setup. Agents are quietly shifting from tools you operate to coworkers you delegate to and check on later.
AI RESEARCH
A new study compared more than 100,000 people against today's top generative AI systems and found that AI now outperforms the average human on standardized creativity tests. The models were measured on classic tasks like generating unusual uses for everyday objects and finding hidden links between unrelated words.
The 100,000-person sample makes the result statistically robust, not a quirk of a small benchmark.
AI was tested on the Alternative Uses Task, the Remote Associates Test, and divergent thinking measures, all long used in organizational psychology.
These exact tests have been used for decades to screen for creative potential in hiring, which makes the result directly relevant to recruitment.
Researchers stress the caveat: the tests measure structured creative output, not the open-ended, culturally embedded creativity behind great art, music, or literature.
The takeaway for builders is about the shape of the distribution, not a doomsday headline. If AI is better than average at structured creative tasks, the work most exposed is the kind that needs only average creativity, while genuinely exceptional or culturally rooted work stays protected longer. The practical move is to use AI to lift the floor on routine creative output and spend your own time where real originality is the differentiator.
HOW TO AI
🗂️ How to make AI generated text sound actually human
There's a specific feeling when you read AI-generated text.
You can't always name it. But you feel it immediately. Something is slightly off. The sentences are too balanced. Every paragraph is the same length. Words like "delve," "tapestry," "multifaceted," "it's worth noting" show up constantly. The rhythm never varies. Nothing sounds like a person said it.
People scroll past it. They trust it less. And if you're using AI to write anything that needs to sound like you, newsletters, LinkedIn posts, cold emails, proposals, this is costing you.
The fix isn't prompting harder. It's building a skill that teaches Claude exactly how you write, and making it run automatically every time.
This guide covers three things: why AI text sounds the way it does, how to build a humanizer markdown file that Claude uses as a permanent reference, and how to install a GitHub skill that enforces anti-AI writing rules on every output.
Why AI text sounds like AI.
Claude, ChatGPT, Gemini and every other model was trained to be helpful, clear, and balanced. That training produces a very specific style.
Predictable structure. Every response has an intro, a middle, a conclusion. Every paragraph is 2-3 sentences. Every list has 3-5 items. Real human writing is messier. Some paragraphs are one sentence. Some are seven. Structure follows thought, not a template.
Safe vocabulary. Models default to words that are broadly understood and unlikely to offend. This produces corporate-speak: "leverage," "utilize," "it's important to note," "in conclusion." Real writers have verbal fingerprints, words they use, words they never use.
No voice. AI writing is neutral by design. It doesn't have opinions, pet phrases, strong preferences, or anything that sounds like a particular person said it. That neutrality is exactly what makes it feel hollow.
The word problem. There are about 20 words that models lean on constantly: delve, tapestry, elevate, comprehensive, robust, crucial, pivotal, multifaceted, nuanced, seamlessly, harness, foster, realm, underscores, in summary. If your output contains three of these in one paragraph, it reads as AI to anyone paying attention.
The solution isn't to avoid AI writing. It's to teach Claude your voice and make it enforce your rules before it shows you anything.
Part 1: Build your humanizer markdown file.
This is one of the most important things in this guide.
A humanizer markdown file is a document that describes exactly how you write. Claude reads it before producing any content. It knows your sentence style, your vocabulary rules, your structural preferences, what you say, and what you never say.
Here's how to build one.
Step 1: Interview yourself (or have Claude do it).
Open a new Claude chat. Paste this:
I want to create a writing style guide that captures exactly how I write,
so I can give it to AI and have it produce content that sounds like me.
Interview me with specific questions about:
1. My sentence length preferences and rhythm
2. Words and phrases I commonly use
3. Words and phrases I never use or hate
4. How I open pieces, do I use questions, bold statements, stories?
5. How I close pieces
6. What tone I aim for (direct, warm, funny, formal, conversational?)
7. What I think makes writing good vs. bad
8. Examples of writing I admire
Ask me one section at a time. After I answer everything, create a
clean markdown file I can download and use as a permanent reference.
Answer the questions. Be specific. "I don't like formal language" is useless. "I never use the word 'leverage' and I hate business-speak like 'circle back' and 'touch base'" is useful.
Step 2: Add the anti-AI word list.
After the interview, add a section to your markdown file called Words to never use. This is where you list every AI cliché you want eliminated from your outputs forever.
Start with these, they're the most common AI tells:
Want to read the full guide?
Become a paying subscriber to get access to all premium content.
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial
Google has confirmed that Gemini 3.5 Pro is arriving in June, with Sundar Pichai telling the I/O 2026 stage to "give us until next month."
Anthropic has closed a funding round of more than $30 billion at a pre-money valuation above $900 billion, making it the world's most valuable private AI startup and pushing it past OpenAI's $852 billion mark for the first time.
Reporting this week revealed that China is widening overseas travel restrictions on people working on advanced AI, extending beyond private firms like DeepSeek and Alibaba to researchers employed or funded by government agencies.
🎨 MAI-Image-2: Microsoft’s improved image generator
🚀 Google AI Studio: Google’s upgraded all-in-one AI builder
🌊 Mistral Workflows: Mistral's new orchestration layer for enterprise AI. Python-defined, durable multi-step processes with human-in-the-loop approvals and full audit trails.
⚡ GPT-5.5 Instant: The new free-tier default in ChatGPT with 52.5% fewer hallucinated claims than GPT-5.3 Instant.
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas