Axios: Jim VandeHei on writing with AI — rules, memory, and autonomous agents

Jim VandeHei, co-founder and CEO of Axios, published a direct account of how he uses AI in his own writing practice, specifically for his C-Suite newsletter aimed at senior business leaders. The piece is notable for its specificity: rather than discussing AI in general terms, VandeHei describes the exact setup he uses, the rules he writes into Claude’s memory, and the separate autonomous agent he built for idea sourcing.

The core workflow

VandeHei uses a dedicated Claude project for the C-Suite newsletter. The project holds everything he has written for that audience, which means the model has context on his past angles, his typical examples, and the specific reader profile he is addressing. Into this project he has written a standing instruction to challenge him with “wise skepticism” — meaning the AI should push back on weak arguments rather than simply follow directions.

The style rules are equally specific. He requires short, sharp sentences with a clinical emphasis on facts. Every piece should flag context in a section called “Why it matters.” Supporting evidence should be stacked in bullet form ordered by importance. These are the defining features of Axios’s Smart Brevity format, applied as explicit constraints rather than leaving the AI to infer them.

An autonomous agent for idea discovery

Separately, VandeHei runs a Codex-powered agent that scans high-quality publications for material relevant to his C-Suite audience. The agent writes potential ideas in his style and delivers them to him by email. This keeps the sourcing step mostly automated while leaving editorial selection — which ideas are worth developing, which angles are true to his perspective — entirely in his hands.

What this demonstrates

The workflow separates two tasks that are often conflated: finding what to write about and deciding how to write it. The autonomous agent handles the first; VandeHei retains control of the second. The style rules in Claude’s memory mean the drafting step already operates within the constraints he cares about, rather than producing output that then needs to be reshaped to fit his voice.

The underlying logic is to set rules based on your own standards rather than the AI’s defaults. VandeHei trains the model on his past writing and writes the instructions precisely enough that the output needs editing, not rewriting.

Who this is useful for

Journalists, newsletter writers, and editors who publish on a regular cadence and have an established voice they want to protect. The approach is most applicable when the writer has enough of a body of work to train a project on and a clear enough style that it can be written down as rules. It is less applicable to writers who are still finding their voice or who work across topics where consistent tone matters less than contextual adaptation.