This newsletter is about giving you insights into the world of exec comms at the biggest AI company on the planet and actionable ways to write and create better in the age of AI. 

Last week I spilled the tea on what’s the tea and optimising for time with comms.

This week, I wanted to dive into how I use Claude Code for writing.

The interface looks a bit like a movie…. Black terminal, blinking cursor, the whole screen stripped down to a single line of plain text waiting for input. The first time I opened Claude Code I was like - I’m in the matrix. I’m a hacker now. 

For the past few months I've been experiment with some comms inside Claude Code rather than the regular chat window: newsletters, LinkedIn drafts, town hall scripts, briefing memos — and the output has gotten sharper while the process has gotten faster. 

The surprise, after being intimidated by the interface, is how adaptive the system is and how it learns as you go. 

Here's what I've learned.

The chat window, Claude.ai, ChatGPT, whichever one you happen to be using — was built for one-off questions, the kind where you type something in, read the answer, and paste the answer into a doc somewhere else. 

The annoying thing about that is that every conversation starts from scratch. The model meets you fresh every time, lives outside your files entirely, and forgets the careful decision you made yesterday about how to handle a tricky paragraph the moment you close the tab.

Claude Code is the same underlying model with the same intelligence, with one important addition: it can read your files, write to them, run commands on your behalf, and hold on to what you taught it last week. 

The terminal is the way in. 

This is where most people give up because it’s annoying tbh. They open the tool, type a vague request, get a generic answer back, conclude the experiment is over, and quit.

The fix lives upstream of all of that. 

The model writes like you when it has direct access to you. You need to give it stuff: your past writing, your voice notes, your favorite turns of phrase. Mine lives in a folder of reference documents that includes a voice guide built from a recorded interview with myself, a style sheet of phrases to not use (its not just x, its y. That’ the whole game, etc) and a running log of feedback I've given the model across months of conversations. Every session reads through all of it before doing anything else.

Then I have loops. 

Loops are sick. Basically just ask it to do a post mortem everytime a task is complete, and remember and update an ongoing .md file so it doesnt make same mistake twice. 

The system gets smarter the more I use it, in a way the chat window has no architecture to match.

Once the system kind of works. The temptation is to scale. 

“Let’s do 10 thousands posts. I don’t have to write ever again” you bellow. 

Slow down soldier… it ain’t that good yet.

Here is what AI can do for you, easily: 

Here is what AI struggles with: 

That part is still your job, and the job has only been compressed in time. You spend less time generating and more time choosing, and the choosing has always been the art of communication.

The model treats a sentence that merely sounds smart and a sentence that actually is smart as identical inputs, but you can tell the difference between them, and discerning between those two things is the thing that matters.

Outsource the discernment and you end up with what everyone else has – SLOP.  The same flattened executive voice, the same hedged paragraphs, the same vague optimism, the voice that belongs to no one.

Taste is what’s important.

If this is valuable and you liked it - please consider sharing 🙂

Mischievously, 

Connor

Each week I break down how I'm pushing the envelope in exec comms — the tools, the frameworks, the bets.