Skip to content
Article Every Sep 2026

Every: good writing with AI starts before you open the chat

Published in Every on September 23, 2026, this piece by Mike Taylor examines the stage of AI-assisted writing that most workflow guides skip: what happens before the first prompt. Taylor argues that the quality of the final piece depends almost entirely on decisions made before any generation happens.

What the article covers

Taylor works as head of evals at Every and has written extensively using AI tools. The article is structured as a workflow, but its argument is a design principle: AI can handle drafting, but it cannot substitute for the judgment about what is worth writing, what material is relevant, and what structure will work. Writers who use AI for drafting but invest no time in what precedes it tend to produce pieces that are technically competent but don’t say anything specific.

Key steps in the workflow

Start with the title. Taylor argues the title drives most of the decision about whether anyone reads a piece at all, and a title the writer finds genuinely interesting sustains motivation through the drafting process. He drafts the title before writing anything else.

Gather material deliberately. Before drafting, he collects sources, including research that is hard to find and pieces he values for their style. He also talks to people whose views might belong in the piece. This material goes into a Sources tab that the AI can draw on later.

Speak before writing. He records unstructured spoken monologues to capture his distinctive takes and memorable phrasing before they are processed through the more formal register of writing. An LLM can later organize these transcripts. The logic is that the idiomatic, specific language in spoken thought is often better than what emerges when writing directly.

Test the structure early. He drafts an introduction and a rough outline, turns his transcript into story beats, and runs them past an editor or trusted readers before writing the full draft. The test is whether the argument holds before investment in polishing.

Generate from multiple models, then combine. He produces drafts from several models, identifies the strongest passages in each, and assembles them. He expects to rewrite 20–30% of the text himself during self-editing.

Write the conclusion yourself. Endings generated by AI tend to fall back on summarization and generic uplift. Taylor writes his own, aiming for an effect that lasts after the reader stops.

Defend your voice when editing. He accepts most grammar and style corrections but pushes back when an edit weakens the idea or eliminates a distinctive phrase. Feeling resistant to a suggested change is sometimes a signal that the original was carrying a point of view worth keeping.

Who it is useful for

Writers who are already using AI for drafting and find the output technically adequate but somehow flat. The article’s intervention is upstream of the tools: if the title is uninteresting, the material is shallow, and the structure is untested, a better prompt changes little. Taylor’s workflow is a case for treating pre-prompt decisions as the actual work, with AI handling the mechanical parts of drafting.