Jim VandeHei: Writing with AI

Few AI use cases elicit more outrage than writing: Using AI makes writing duller ... dumber ... robotic. It kills thinking ... creativity ... originality. It produces sameness and slop!

Why it matters: All that's true — if you, as the teacher or writer, allow it to be true. Lazy AI outsourcing means lazy thinking and writing. Used with persistence and skill, AI can enhance both your thoughts and expression. I'm thinking and writing more and better than ever. But I want to be frank and useful about what I've learned about writing and the limits I've hit since turning myself into an AI lab rat: Caution 1: I'm 55, a writer by training, went in aware of AI dangers, and am a prolific writer with or without it. Anyone lazily letting AI do their thinking or writing is playing with fire: You're destined to create a mushy, "blah brain." Hope the tips that follow help you build a bionic brain instead. Caution 2, from Autumn, my wife, who's more of an AI skeptic: "Don't conflate the way you write and the utility of AI to convey information with soul writing that many of us need to live, breathe and understand the world around us. Living without that is like trying to live without air for many of us — it needs to be said somehow."

Between the lines: As AI tools get more powerful, Axios will stay transparent with you about how we use them. (Our website includes a detailed explanation of how we use AI in our journalism.) This column captures my latest thinking on best practices. So here's the nitty-gritty of how I use AI for writing: Set rules based on your standards, not AI's. I write into the model's memory: My writing flows from my thinking. So challenge me, never flatter me; press me with wise skepticism. Then write like me — always in Smart Brevity — to match my style. Be very precise about the writing and editing style. I have mine wired into the AI memory (just tell it to commit style to memory) and in a skills document for my AI agents. My rules: short, sharp sentences ... clinical, fact-based emphasis ... context flagged as "Why it matters" ... and supporting points stacked in bullet form in order of importance. Pour in examples. You need to feed in the original work you're most proud of. It can be as simple as diary entries, memos or a school project. Make sure your writing and editing priorities are reflected in it. I've dumped in every Axios column, memo and strategic document, as well as four books (two unreleased). Let's pause for a second. Those are very clear parameters for writing and editing with AI. That will suffice for most people in most instances. I've taken my own AI journey much further. I'm doing this to test limits and capabilities. Advanced writing with AI: Create specific advanced skills. I use both projects (mainly in Claude, but increasingly in ChatGPT) and agents (in ChatGPT's Codex). Inside projects and agents, my thinking, writing and editing files are much more specific and richer. They're called skills or instructions, depending on the platform or use case. I detail a fact-checking skill to double-check data points, then move outside the agent for another fact-check with a safety-net skill. Interact with projects and agents a lot. You need to create a conversation loop, where you get better at fine-tuning the writing output, and the AI gets better at understanding your style. It definitely gets better with time. You want to find a mind meld that serves and betters you. Put it to the test. My laboratory is my Axios C-Suite newsletter, which I publish Saturdays for CEOs and other executives. I'm testing and using two main things: I have a project inside of Claude that knows everything I've written for this audience, plus my thinking about business and leadership, and the high standards I have for ideas and data. I have a similar agent that operates autonomously, powered by Codex, that constantly scans high-quality publications, data projects and research for relevant information. It writes up ideas in my style and delivers them via email before I wake up. The items are good, smart starters or enhancers — but they're never camera-ready, at least by my standards. Push and pull — a lot. Funny thing: I write naturally in an AI kind of way — direct, sparse, lots of dashes. And AI naturally writes a lot like me (and naturally loves Smart Brevity). So I find the output often excellent, well-documented and edited better than my favorite human editor, Mike Allen. It still does weird stuff, but rarely. It's wordier and less direct than I am. Increasingly, I find myself using its edits or proposed Jim-like phrases as naturally as the notes from Mike or any other human editor. Fun new test: I'm preparing to hike Kilimanjaro, so I'm doing a lot of rucking. I've been using Claude in voice mode to have long conversations about things I want to write. I have Claude keep detailed notes, pull out my best phrases and ideas, challenge my assumptions, help organize my thinking. Then we go back and forth by voice before producing a draft column. That's how I got both the idea and the output for my "Confessions of an AI lab rat" column. The end product was a hybrid of my thinking and my best phrases alongside Claude's real-time editing and editor-like pushback.