HOW THIS IS MADE
The method behind everything published here, and where AI sits in it.
Everything published under Orion Playbook runs the same way: research, argument, drafting, editing, and a line-by-line pass by me before anything ships. Newsletter issues, social posts, images, audio, and what's coming next. AI is in that process at every stage. So is my judgment, and my name is the one on the result.
Here is the method, in the order it runs. I am writing it down as I am using AI in public, and my learnings are worth sharing.
Research.
Every piece starts from a question I cannot already answer. The first pass, which I run with Claude Opus, gathers primary sources, the numbers, and the strongest arguments against the position I walked in with. Every fact that reaches a published piece traces to a source I can point to. Claims that can't be verified get killed here. A lot of them do.
Argument.
No prose gets written until the piece commits to three things: a thesis that could be proven wrong, a mechanism explaining why it holds, and one move a reader could make on Monday morning. Ideas that fail that test get reworked as ideas. Writing them more attractively is no fix at all.
Writing.
The draft runs against a voice specification I started building in May 2026 and have rewritten continuously since: how I sound and the constructions that give machine writing away. English is my third language. Working this way gets me closer to what I actually mean than writing unaided does. That is why I do it, and it took me a while to say so out loud.
Editing.
Every piece gets a structural edit, then comes to me line by line. I cut, rewrite, argue with it, and send it back. Most pieces take several rounds. Grammarly Pro takes a final pass for spelling and grammar, which is the smallest job on this page. Everything here involves judgment (including accepting Grammarly changes).
Time.
A 700-to-900-word article takes about four hours of my time. I give you the number because it is what makes this different from a short prompt and a publish button. A monthly batch of LinkedIn posts takes an afternoon of my time to go from idea to being scheduled in Buffer.
Images.
I generate every image with ChatGPT Image or Gemini Nano Banana. Each one comes from a prompt written against a defined visual system, usually over several attempts. The hero image is a combination of multiple outputs. I would rather make something specific to the argument than reach for another overused photograph from Unsplash or Pexels.
Audio.
Each article ships with a narrated version, generated with ElevenLabs from a customized voice that has nothing to do with mine. The player says "AI-generated," and the recording says so in its first sentence, before anything else. If you ever hear my actual voice here, it will be because I recorded it.
The stack, as of October 2026.
Research, drafting, and editing
Anthropic Claude Opus
Final spelling and grammar
Grammarly Pro
Images
OpenAI ChatGPT Image
Google Gemini Nano Banana
Audio
ElevenLabs Eleven
Speech-To-Text
Superwhisper
Claude Voice/Speach-to-Text
Background operations
Anthropic Claude Cowork, on the Sonnet model.
This runs the scaffolding: gathering sources, moving files, queuing posts I have already approved. It writes nothing you read, and it publishes nothing on its own
I use whichever version is current when I sit down. This list will evolve over time, which is covered below.
Where this applies.
The method holds everywhere. What changes is the depth.
Newsletter issues and their audio. The full process above, about four hours each.
LinkedIn and other social media channels. Same voice specification, compressed to the format. Posts are drafted in batches, and each one gets edited by me before it queues (using Buffer).
This website. Written like everything else here.
Consulting and speaking engagements. Delivered by me, in the room. AI helps me prepare. It does not show up in my place.
What stays off the table.
I do not publish anything I have not read closely, do not agree with, and cannot defend. I do not use AI to invent experiences or fabricate quotations. Scheduling is automated; approval never is. Replies to comments are mine, written in the moment by a person who read what you wrote. Every piece carries my editorial responsibility, in the ordinary sense of that phrase and in the specific sense the EU AI Act gives it.
The method changes, and that is the point.
Until a few weeks ago, drafting was one step. I split it into separate research, argument, writing, and editing stages because doing them at once produced pieces that read well and proved less than they claimed. That was a defect in how I worked, and the split fixed it. Expect this page to go out of date, and expect me to update it when it does.
One more thing: The major AI providers now embed machine-readable marks in generated text and images, and detectors are being built into the platforms you read on. Anthropic, whose model I use most, is direct about the limits of its own mark: "Detecting a Claude mark tells you that the content may have been processed by Claude. It does not, on its own, confirm the full provenance of the content."
A mark tells you a model touched the words or image, and I intentionally leave them in. It cannot tell you who did the thinking, who checked the sources, or who answers for a claim that turns out to be wrong. Only a person can tell you that, which is what this page is for.
If any of that is a dealbreaker and you don't derive value from the Orion Playbook, unsubscribing takes one click, and there are no hard feelings.
Gérard Métrailler
Last updated: October 23, 2026
