1st March 2026

How This Email Gets Made

A human–AI collaboration, every Saturday

A later note

I wrote this in week two of the weekly email, on 2nd March 2026. The process has evolved significantly since.

Two things stand out that this version doesn’t properly capture:

  1. Everything in the email comes from me. Something I’ve read, something I’ve done. There’s no web trawl filling in what I might have missed. It’s a week of my working life. (A separate research artefact, How the AI Story Was Told, reads the press; it is not where this email comes from.)
  2. I’m much clearer now on how big my role is at every step. I know exactly what I’m doing versus what the language models are doing, and I’ve done some clever things to keep them in their box along the way.

How it worked in week two. The diagram below is the process as it was then. It has changed since: David chooses the material and the argument; AI assists with processing, drafting and production within defined instructions; automated checks support human review; David approves each edition and scheduled systems publish it.

David
Claude
Both
How it started

Edition 1 was written largely by hand — from a voice memo, through multiple drafts, with Claude helping on wording and research. Claude Code then studied that first issue and built a reusable framework: tone of voice, structure, template, style guide. That framework now shapes every future edition (from a task to recurring delegated work).

Otter Claude.ai Claude Code
Gather
Monday – Friday
Throughout the week, organically, whenever something strikes:
  • Voice memos recorded between meetings
  • Screenshots of the relevant bit
  • Links sent via Telegram
  • Discussions sparked by reader replies
Otter Telegram → Claude Code Gmail
Process
Automated
Screenshots and links are analysed and summarised. Voice memos are transcribed. Everything lands in a running notebook of observations, quotes and half-formed thoughts.
Claude API Otter Claude Code
Draft
Saturday morning
The week’s material is combined with the framework — tone of voice, past issues, style guide — and Claude produces a few suggested drafts.
Claude Code
Rework
The hard part
Three layers of human judgment:
What’s worth including
David 50%
Claude 50%
The “so what” — why each point matters
David 33%
Claude 67%
Style and voice
20%
Claude 80%
Verify
Automated
A separate agent finds supporting links and fact-checks claims before publication. Checks support human review; they do not transfer ownership (what checking and ownership mean here).
Claude Code
Personalise
Saturday
Claude drafts a personalised opening line for each engaged reader, based on their replies and the conversations we’ve had. David reviews hundreds of these, makes tweaks, then approves.
Claude Code Gmail
Publish
Saturday
Final review, then sent. In week two David pressed send himself; today he approves the edition and scheduled systems publish it.
Claude Code Gmail
Saturday AI Thoughts — David Boyle

This is an artefact created for David's weekly email. See the others here.