The newsletter

Claude found something an accountant missed / Socrates' way of questioning as a skill / Video nine

19 November 2026

An accountant had Claude check four sets of accounts after she had, and on one it found something she'd missed, plus Socrates' way of questioning as a skill, video nine, and last week's Fable and ChatGPT followed up.͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ ͏‌ 

buildwitholi

Hello all,

Pretty awesome, we've just crossed 8,000 on Instagram.

The @buildwitholi Instagram profile at 8,044 followers

Thank you for all of your support.

(Will we reach the goal of 10k by the end of the year?)

There's a small line at the bottom about this too, but please let me know about anything you want from me to support you on your AI adoption journey.

[IN SHORT]

This week

  • A second pair of eyes: the accountant who had Claude check her accounts after she had.
  • Great minds: Socrates' way of questioning, so Claude asks how you know before it helps.
  • Video nine: growing Claude from one folder to every part of your business.
  • Worth your time:
    • Why losing Fable isn't a big deal.
    • OpenAI's new GPT-6.2 Astra.
    • Claiming your business on ChatGPT, step by step.

[CASE STUDY]

A second pair of eyes on the accounts

A couple of weeks ago I told you about the accountant whose inbox Claude now keeps tidy.

Since then she's had it check her accounts too.

One of her team prepares each client's accounts, and she reviews them before they go out.

So she did her own review first, then asked Claude to review the same accounts, without showing it any of her notes.

On three of the four sets, it found the things she'd already spotted.

On the fourth, it found something she hadn't, in her words:

"I did my review, and then I got Claude to do a review with no comments from me on it. And it actually spotted something that I didn't."

Four sets of accounts, three ticked as she found them and the fourth flagged with something she missed

Then it wrote the summary email for the client, and an email to her accounts assistant with the changes to make in the client's folder.

She still does her review, and Claude is the second pair of eyes after it.

I'll keep you updated on what else we get Claude doing for her.

[GREAT MINDS]

Socrates' way of questioning

Claude is incredibly agreeable, but I'm on a mission to build skills that stop it behaving like that.

Today's is based on the way Socrates questioned people, nearly two and a half thousand years ago.

He was once told that no one was wiser than him, so he went to ask the people everyone called wise.

He kept finding the same thing, that everyone thought they knew things they didn't.

A message from Socrates, "He knows nothing, and thinks that he knows", from Plato's Apology

Which is exactly how you or I behave without knowing it.

And Claude just nods along, and we only find out when it's too late.

So to stop this, I've turned it into a skill you can use any time you're working through something big:

  1. First, it says back to you what you're sure of, in one line.
  2. Then it asks how you know, one question at a time, each one building on your last answer.
  3. When your answers don't add up, it puts your own words side by side and shows you where.
  4. After five questions at most, it tells you whether it holds up, and what you'd need to find out before you act.
Claude putting the coach's own words side by side, none will leave, and two clients asked about cheaper options

All it's allowed to do is ask, so it can't just tell you what you want to hear.

Get the skill →

[NEW YOUTUBE VIDEO]

Video nine, the whole business

Once Claude knows one part of your business, the next step is every part of it, with a folder for each area of your work and its own brief, so you never have to explain the context again.

Part nine of the agent ecosystem series, Claude across your whole business, and a play button

In this video I cover:

  • A brief in every folder: so Claude reads an overview instead of opening every file.
  • Keeping the top one up to date: so it sends Claude to the right folder.
  • Smaller models for smaller jobs: so your weekly allowance goes further.
  • Closing out a chat: so what it learned is written down before you leave.

I also use my own setup as an example, from a different email address for my life and my work to separate writing skills for Instagram and LinkedIn.

Watch it →

[WORTH YOUR TIME]

Fable has left Pro, and that's fine

Last week I said to make the most of Fable 5.2 before Monday, and since Monday night it costs extra on Pro again.

But Anthropic's own advice is that Fable is for big technical work in coding, scientific research and finance, and that for most everyday work you start with Opus 5.5, which you've still got.

Anthropic's Fable page, recommending Opus 5.5 for most everyday work

So if what you get back isn't quite right, chances are it's not the model, it's that you're telling it how to do the job.

Most people ask for specifics and over-bound the model, so it does exactly what it's told and never stops to think what else might help.

Instead, tell it what you want, not how to do it, and if it's connected to your email, try this after your next call:

"Knowing what you know about me and the information you have access to through the connectors, draft a response."

Then only add a rule where it keeps getting something wrong.

Read it →

GPT-6.2 Astra, from OpenAI

On Tuesday OpenAI released two new models, GPT-6.2 Astra, which they call their new frontier model for your hardest projects, and GPT-6.2 Sol, their go-to model for everyday work.

They say Astra's slides, spreadsheets and documents read less like AI, and follow the conventions of fields like law, finance and accounting more closely.

OpenAI's announcement of GPT-6.2 Astra and Sol

Both are in ChatGPT Work and Codex on Plus, Pro, Enterprise and Edu, rather than the regular chat, so if you can't find them, that's why.

And it's the same story as Fable, with the biggest model for the hardest work, and many of OpenAI's own people using Sol as their default.

Read it →

Claim your business on ChatGPT, step by step

Last week I told you OpenAI had launched Businesses in ChatGPT, and I've since claimed mine and written up every step.

The fiddly part is proving the website is yours, which means adding one line of text where you bought your domain.

The five steps to claim your business on ChatGPT, from finding it to writing your profile

Claiming it won't make ChatGPT recommend you more often, and OpenAI's own help page says so, but it decides what people read once your name has come up.

Read it →


Thanks for reading.

If you run a small business and you feel you're not getting as much out of AI as you could be, I've got some slots for a free 30-minute chat.

We'll talk through your setup, how you're using it now and how you could get more out of it.

If you're keen, I'd love to chat.

Book a chat →

And if a chat isn't for you, hit reply and let me know what you'd like more of in these emails.

Chat soon,

Oli

@buildwitholi on Instagram

Get the next one

Your address goes on the list and nowhere else. Privacy policy.