The "delve," the "leverage," every sentence the same length, the confident vagueness. People learned the tells this year, and the second they spot them, they trust you a little less.

I spent last week fixing this properly, not with a prompt I paste in every time, but with a Skill: instructions you save once that Claude follows automatically, forever. Built five of them. The first one strips the AI voice out of anything I write, and I have not edited for it since.

That one is below, in full.

A Skill is instructions Claude follows automatically whenever the task matches. Save it once, no code, and it fires every time without you asking.

This week's Skill does one job: it catches the patterns that make writing sound AI-generated and rebuilds the text in a real voice. Turn Skills on, paste it in, and everything you write from then on comes out clean.

Time: two minutes to switch on, five to build it. You need: Claude on a paid plan, one setting flipped on.

The old way: you write with AI, then either send it sounding robotic or spend twenty minutes stripping the tells out by hand. Every time.

The replacement: the Skill does the stripping automatically. Ask Claude to write or fix anything, and it applies your de-slopping rules without being told.

A toggle, not a coding thing.

That is the step everyone misses. Without it, the Skills menu does not appear. Claude also ships with built-in Skills for Excel, Word, PowerPoint, and PDF, so the moment you flip that switch, those start working too.

In the Skills menu, create a new Skill. Give it this description, which is how Claude knows when to use it:

Then paste these instructions:

What you get: every piece of writing Claude makes for you comes out sounding like a person, first try. No paste-in rules, no editing pass.

Test it: ask Claude to write a short promotional email for your business. Watch it come back clean and specific, because the Skill fired on its own.

The Anti-AI-Voice Skill saves you an editing pass on everything you write. Now imagine four more, each doing a whole job the moment you ask: growing on Twitter, reading your business numbers, writing cold emails that get replies, and applying real persuasion psychology to your copy.

That is behind the wall this week, plus The PromptWire Skills Bank: 100+ ready-to-use Skills you can paste straight in.

The Skills Starter Pack. The free Skill saves you an editing pass. These four save you whole jobs. Each is the description plus the paste-in instructions, built and tested, no code:

The PromptWire Skills Bank. 100+ ready-to-use Skills across writing, marketing, sales, operations, finance, and admin, each written as plain-English instructions you paste straight into your Claude app. No code, no install. Plus the one security check to run before you trust any Skill from elsewhere.

The build-your-own recipe and the cheat sheet, so you can turn any task you keep repeating into a Skill in five minutes.

If your business runs on Notion, this update is worth two minutes. Notion AI 2.5 shipped in July, and the change is that it now works across your whole workspace instead of one page at a time. It pulls context from linked databases, attached documents, and calendar entries to answer you.

Here is why that matters. The old version could only see the page you were on, so anything spanning projects was useless. Now you can ask "what did we decide about pricing across all our client docs" or "pull everything on the Henderson project" and it reads across the workspace to answer.

🧮 OpenAI's next model solved ten decade-old math problems. On August 1, OpenAI said an unreleased version of its next model, Astra, produced genuine solutions to ten open problems in maths and computer science, each with a machine-checkable proof anyone can verify, for roughly $2,000 of compute (source: TechTimes).

📈 A key AI stock just posted proof the money is real. Palantir reported August 3 that revenue jumped 92.8% year over year, with its US commercial business up 149% (source: The Signal). Businesses are now spending serious money to actually deploy AI, not just experiment. The companies moving from testing to using it are pulling ahead, which is the whole point of building your Skills now.

🛡️ Microsoft launched AI agents that defend your business from AI attacks. Project Perception went to public preview on August 3: teams of AI agents that probe for security weaknesses, investigate threats, and deploy fixes on their own, with humans signing off on the big calls (source: Redmond). Attackers now use AI to move faster than human security teams can, and this is the first mainstream answer aimed at closing that gap.

📜 Anthropic launched official Claude certifications. Four tracks went live July 8, delivered through Pearson VUE with digital credentials (source: ClaudeLog). AI skill is becoming something you can formally prove, which will start showing up on job listings and freelance profiles sooner than you would expect.

Quick version: OpenAI's unreleased next model, Astra, solved ten math and computer-science problems that had been open for at least a decade, and shipped machine-checkable proofs anyone can verify for about $2,000 of compute. The headline is the maths. The lesson that matters for your work is the verification: the proofs count because a machine can check them, not because you trust OpenAI. That is the exact discipline you need for AI in your own business.

On August 1, OpenAI announced that an internal version of its next model, called Astra, had produced new results on ten problems in mathematics and theoretical computer science, each open for at least ten years (source: TheNextWeb). One had stood since 1999. The whole run cost roughly $2,000 in compute (source: implicator.ai).

Impressive. But the number ten is not the important part.

Every proof shipped with a Lean certificate. Lean is a proof assistant that forces every step of an argument into machine-readable detail, so a program checks it line by line, without anyone having to trust the lab that produced it. The files are public on GitHub. Anyone with a laptop can verify them.

That is the shift. When OpenAI's model disproved a famous conjecture back in May, validation meant nine mathematicians reading it by hand and vouching for it (source: TechTimes). This time, trust moved from "believe the humans who checked it" to "run the checker yourself." The claim carries its own proof.

Now bring that back to your desk. You use AI that is confident and often wrong. The Astra story is the whole problem in miniature: a brilliant output is worthless if you cannot verify it, and worth a lot if you can. The labs solved this for maths by making the output checkable. You need the same habit for the AI you use every day, because nobody is shipping you a Lean certificate with your marketing copy or your financial summary.

Being clear-eyed: this was maths, where a proof can be mechanically checked. Most of your work is not like that. None of the field's hardest problems fell either. But the principle holds, and it is the one skill that separates people who get burned by AI from people who get value from it.

🔒 The Pro breakdown: how to get trustworthy, checkable output from AI in your own work, the four ways to make any model show its proof, and the specific tells that a confident answer is actually wrong.

Workflow Library Updates:

NEW Resources:

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

You have been briefing Claude from scratch every morning like it is its first day. It does not have to be. Set it up once, the way you would train a good assistant, and it remembers how you like things done from then on.

If you do one thing before next Thursday, turn the switch on and build the Anti-AI-Voice Skill. Two minutes plus five, and the next thing you write stops sounding like a robot.

👇 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.