01 / 18

Before anything else

Give me something to build.

Shout one out. Anything you'd want on a screen. I'll start it now and it runs while we talk.

02 / 18

Running

It's going. We'll come back to it at the end.

If it falls over, you'll see that too. That's the point of doing it live.

03 / 18

Quickly, so you know why to listen

Everything I do goes through Claude.

£0

on AI tooling since May. Claude, OpenAI, Anthropic.

£0

on a PC built to run it faster.

0M

tokens through the API in five weeks.

0

automated calls I never watched.

0h

a day, every day, since March.

04 / 18

Some of what came out of it

Nobody wrote any code. Me included.

StudioCanal UK licensing census — **StudioCanal licensing census**266,522 stores read, 24,071 checked by hand. Ask it a question and it answers.

Superco AI Visibility Check — **AI Visibility Check**Superco's own lead magnet. 503 brands scored on whether AI recommends them.

Helix Health pharmacy group calculator — **Helix Health calculator**A pharmacy group works out its own numbers before anyone gets on a call.

Last-mile delivery cost calculator — **Delivery cost calculator**For a last-mile carrier. The prospect sees what the number is made of.

Kennis Europe calculator — **Freight cost calculator**Same idea for a European forwarder, in their brand, on their domain.

AGNODOS proposal, first page — **A proposal, from a call recording**Discovery call ends, the proposal writes itself and lands in the inbox.

05 / 18

Your ideas

Here's what you told me you wanted.

06 / 18

Tom

Pulling receipts and paid invoices for bookkeeping. About an hour a month. Adobe hides them somewhere in the back end.

Already running

It watches Gmail for invoices, converts them to GBP and adds them to the tracker, every evening. Once a month a browser script collects the ones that never email a receipt. I haven't typed a transaction by hand since July.

Happy to tailor this to your workflow. About thirty minutes on a call, sometime this week.

07 / 18

Tom

The sales spreadsheet, pulling from Attio. I spent a lot of time and credits and it didn’t deliver anything useful.

Already running

This is the Superco tracker. A live sheet, fed on a schedule from the campaign system, that people actually open. The same shape pulls from Attio.

Happy to tailor this to your workflow. About thirty minutes on a call, sometime this week.

08 / 18

Lisa

Some of the solutioning is quite poor. Sometimes absolute gibberish. I think its sources are out of date.

You're right

It's answering from memory. Every tool I use has a folder, and the learnings file sits on top. It reads the docs, adds what it discovers, and updates the file itself.

Start here

Take your Shopify call transcripts, hand them over, and let it write them into a learnings file. It stops guessing the moment it has something to read.

integrations/ plusvibe/ LEARNINGS.md 48 dated gotchas, indexed plusvibe.py openai/ LEARNINGS.md email-verify/ LEARNINGS.md ...33 of them # PlusVibe - tool learnings Terse, one-line insights. Insight, why, how to apply. - Lead label vs status are two different API fields - Writing a lead label uses POST /lead/data/update - ...

Happy to tailor this to your workflow. About thirty minutes on a call, sometime this week.

09 / 18

Sophie

The first draft usually sounds very robotic. I want it to ask me the questions rather than the other way round.

Proof it's buildable

That's the output of a voice skill built from 128 real posts, measured rather than described. The tone is deliberately a joke. The method is the point: Superco has years of published work, and a skill that interviews you before it writes is a day's build.

A prospect's AI assistant has been negotiating with my AI reply agent for 11 days. Neither of them told us. The thread is 34 emails long. They agreed a meeting time, moved it four times, and complimented each other's response speed. At email 29, my agent offered a case study. Her agent requested a security questionnaire. My agent completed it. We only found out because the calendar invite landed: "Intro call - AI x AI". We are letting them take the meeting.

Happy to tailor this to your workflow. About thirty minutes on a call, sometime this week.

10 / 18

Rachel

Find the out-of-office replies from people who’ve left and see if they name someone else. And sense-check anything going out, without writing it for me.

Out of office

That's a sequencer feature and the one you're moving to has it built in. I'm setting it up anyway, so it'll just be on.

Sense-check

The copy gate reads a draft against a written standard and tells you what's wrong. The writing stays yours, and your boundary stays where you put it.

Happy to tailor this to your workflow. About thirty minutes on a call, sometime this week.

11 / 18

Farah

A pitch deck that doesn’t eat credits. A proposal from a template with the exact services.

Already running

After a discovery call, the proposal writes itself from the recording and emails itself. One spec file makes the review copy and the branded document, so they can't drift. The deck is the same idea in a different shape.

Happy to tailor this to your workflow. About thirty minutes on a call, sometime this week.

12 / 18

How I've set up Claude

Five files. All of it runs in Claude Code.

CLAUDE.mdRead at the start of every session. It says:

how to reply to me, and in which English
what to do every time, without being asked: give an ETA, sample the output, render before handing over
the hard rules it may never break
where everything lives

SkillsForty-eight. Each one is a job I used to do by hand. The ones I use most: research a prospect before a call, write a client's cold email in one of three styles, build an interactive lead magnet, write the monthly client report.

opinions.mdWhat I believe about my craft. It came out of a 44-question interview it ran on me.

writing-style.md and design-system.mdWhat good looks like, as a check that can fail. The writing one strips the AI slop out of everything before I see it. This page went through both.

MemoryFour hundred and thirty-five things it got wrong once and won't get wrong again.

13 / 18

How I work with it

Three habits that make the files work.

Build loopFor anything big. It writes the scoring rubric before the build exists, builds, renders, then something that didn't do the building checks it. The check can't be bent to fit what got made.

Knowledge baseOne folder of business truth it reads before it writes: who we are, what we believe, what the client said. Nothing lives in a chat.

Prompting processWhen I want it to keep something: tell it once, then remind it the next two or three times. After that it does it on its own, and I write it into a file.

14 / 18

The parts I'd show a friend

Things that run without me.

FinanceInboxDaily at 19:00. Gmail to the ledger, in GBP. No typing since July.

LucaA WhatsApp number in my client chats. Message it an instruction and it runs my whole setup.

LocalWatchChecks every automation hourly and repairs the ones that fell over.

MonthlyLoopEvery client's 30-day review, written and waiting for me at 06:30.

The census266,522 stores read, 24,071 checked by hand. One page.

This deckWritten, checked and hosted the same way. It reads on its own afterwards.

15 / 18

Make whatever you want

Four steps, and none of them needs a terminal.

01Write down what you believeOne page on how your job should be done. Let it interview you if that's easier. That's your opinions file.

02Make a folder and tell it who it isA short CLAUDE.md in the project folder. It gets read first, every time.

03Do the task once with a recorder onHand it the recording and ask for a skill. Customize, then Skills, then upload. That's the whole format.

04Run it through a check before you trust itA plan written before the work, a build, a render, and a check by something that didn't do the building.

16 / 18

Or send me the thing you want built, and I'll run mine on it.

Step four is the part people skip. My build loop does it: the rubric gets written before the build exists, so the check can't be bent to fit what got made. You get it back with the score, and the transcript of this session plus a starter repo if you want to run it yourself.

17 / 18

Back to the start

So. Did it work?

18 / 18

That's it

Ask me anything.

This page was made the same way. It's yours to keep, and it reads on its own.