Write Down Your Editorial Standards

I’ve spent too much time asking AI to fix writing habits I had already corrected.

But asking for better prose left the model to decide what I meant by better. “Be more concise” helped for a few paragraphs, then the unnecessary introductions returned. So did the conclusions that repeated what the reader had already understood. Sometimes I was making the same correction within a single conversation.

The experience convinced me that I needed to explain the editorial decision behind the correction. If an introduction takes too long, the useful instruction is about where the piece should begin. If a conclusion repeats the argument, the agent needs to know where to stop. Those decisions can be written down and reused, instead of being worked out again with every draft.

Give your agent the standards you use when you edit. You still have to decide whether the finished piece is worth publishing. An opinion column needs to develop a position and answer objections; a technical essay needs to explain how something works before drawing conclusions. A field guide needs enough practical detail for a reader to apply it, which makes cutting every explanation a poor rule for that form.

I’ve put those distinctions into Writing Skills, an open-source repository of twelve skills for Claude Code, Codex, ChatGPT and other compatible agents. Its humanizer identifies recurring patterns of machine-generated prose and asks the agent to preserve the author’s voice, facts and citations. Eleven further skills describe particular writing forms. They draw on close readings of published work to explain how each form develops an argument and uses evidence.

Two Python checkers let agents review drafts against those instructions. One flags common AI-writing patterns; the other checks the structure of the selected form. A good sentence can trigger a warning, and a generic draft can pass every check. The warnings need judgment, and responsibility for the published words remains with the author.

The repository includes the supporting research and checking scripts under the MIT licence. You can inspect the standards, disagree with them and revise them. When an instruction keeps producing an edit you would reject, you can change the instruction before using it again.

Teach your agent the decisions you keep making yourself.