💌 Claude caught a mistake nobody asked it to look for / £20 off the course until Monday / Video 5 is live
22 October 2026
buildwitholi
Hello people,
Like clockwork, it's Thursday, and here's your update on the world of AI.
And make sure you read to the end, there's a discount waiting for you...
[IN SHORT]
This week
- Part five is live: how to write the perfect claude.md (the brief it reads every time you send it a message).
- Case study: an accountant's first week, and the mistake Claude caught without being asked.
- Worth your time: what an Anthropic engineer asks Claude instead of telling it it's an expert.
- £20 off the course: the code NEWSLETTER20 takes it from £69 to £49 until midnight on Monday.
[NEW YOUTUBE VIDEO]
Part five, write the brief for your business
The brief is a short note to Claude about you and your business, a file called CLAUDE.md, and Claude reads it every time you send it a message.
So if the brief is wrong, you spend your days fixing what comes back.
But if it's right...
It's like having that star employee you never have to remind of anything.
In this one I go through mine and show you how to write yours, including:
- The six sections: from who Claude is and who it works for to your standing rules and the things it must never do.
- The interview: a prompt that asks you questions and writes the brief for you.
- The right length: 60 to 120 lines is the sweet spot.
- The canary: a one-line check that tells you a chat has gone stale.
- Andrej Karpathy's rules: where he says AI goes wrong when it works for you, and the standing rules that stop it agreeing with you and charging ahead.
Everything else you build, the connectors, the skills and the jobs on a timer, works from this one file, so it's worth getting right first.
[CASE STUDY]
An accountant's first week
On Monday I started an eight-week build with an accountant who runs her own small practice.
Her biggest pain is email, the slow follow-ups to new clients and the documents stuck in her inbox.
So in the first session we let Claude read her shared drive, without being able to change anything, and connected her inbox so Claude can write drafts, but nothing leaves until she presses send.
Then we gave it a real job, finding a long thread with a prospective client and drafting the reply with her quote.
The draft was close to what she'd send, but the best bit was a note Claude left at the bottom.
Earlier in the thread she'd told the client they'd move to Dext, the app for receipts, and the quote she drafted never charged for it.
In her words:
"I didn't say to it, read through and see if I've missed anything in that quote, but it's gone, 'she's missed that, so I'll point it out.'"
That's the lesson I keep coming back to, tell Claude what you want rather than exactly what to do.
The models are clever enough now that a long list of must dos gets in the way of them spotting things like this.
Next week she turns her edits into two skills that draft emails the way she writes them, and we connect her accounts software.
[WORTH YOUR TIME]
Stop telling Claude it's an expert
Felix Rieseberg, one of the people who builds Claude at Anthropic, wrote an essay last week on getting more out of it, and his post about it passed four million views in three days.
Most of us were taught to start with "you are an expert", but it changes how the answer sounds, not how good it is.
A study at the University of Southern California found smaller AI models, not Claude itself, got about five more answers in every hundred wrong when they were told they were experts.
Felix asks Claude four questions instead, to get it to rewrite his request:
- The top professional: what would they do here that you're not asking for?
- The raw material: what would they start from, rather than what's already been summed up for you?
- The methods: what tools and methods would they use?
- The proof: what, other than Claude or you, would show the result is right?
Here's his prompt, word for word, to paste in before any request where the thinking matters:
I want a deeper version of this task, not a bigger one. Before rewriting it: (1) say what a top professional in this field would do here that I'm not asking for; (2) name the rawest material they would start from instead of what has already been digested for me; (3) name the specific methods and tools they would use; (4) say what, other than you or me, would show the result is right. Then rewrite my prompt to ask for exactly that.
It was my most watched video this week, so I've also put it into a free part of the course as a skill you can call on any time, with a before and after.
[THE COURSE]
£20 off until Monday
The course takes you from the free first stages to a Claude that knows your business and takes work off you:
- Mapping your business.
- Writing the brief.
- Connecting your inbox and calendar.
- Building skills that run on their own.
Every step comes with the prompts to paste in, so you build it as you read.
A CEO running two companies told me on a call last week:
"You read theory and then you give me something that I can actually go and copy and then test. And that's really, really powerful."
And someone on stage six sent me this:
"The course is really well structured, super clear, and I love the copy/paste prompts, they're like customisable cheat sheets that tell Claude how to build everything I need."
And if you don't believe the text, here are some of the people who've built theirs with me:
Every stage is written up with its prompts, but while not all the videos are recorded, it's £69 for now.
It will go up to (at least!) £149 when they're all filmed.
With the code NEWSLETTER20 it's £49, until midnight on Monday.
The button below takes you straight to checkout with the £20 already off, and if you're new it signs you up free on the way.
If it isn't for you, there's a full refund within 14 days, and once you're through stage seven you get a 30-minute call with me.
Any questions, just shout, I read every response.
And if you'd rather have someone help you set up your ecosystem, reply with a few times that suit and we can talk it through.
Thanks for reading.
Chat soon,
Oli