The author describes a personal method for using large language models to improve writing while preserving the writer’s voice. According to the post, the first rule is that a writer may not use a single word an LLM suggests. The reasoning is that frontier models are exceptionally good at selecting pleasing turns of phrase, and adopting their wording can make writing feel like a string of magazine headlines, which the author compares to Velveeta. The author calls this a form of intellectual personal protective equipment and urges strict adherence, even when the suggested words seem better than the original. The second rule is to avoid encouragement from the model. The author explains that when a piece is handed to an LLM, it often responds with praise such as “that’s gold, Jerry!” This feedback can reinforce weak first‑draft impulses, leading writers to double down on problematic passages instead of revising them. The author warns that listening to such encouragement can make writing appear artificially flavored to readers, even if they cannot pinpoint exactly what is wrong. To counteract this, the author previously opened copyediting prompts with the statement that they are not the author but an editor of an online publication, though notes that the model still tends to overfit to the imagined publication’s goals. The practical advice given is to forbid the model from offering encouragement and to remain vigilant about any praise it provides. Despite these restrictions, the author asserts that LLMs are excellent at flagging mechanical problems in writing. Examples cited include overuse of passive voice, nominalized verbs, repeated turns of phrase, and filler words like “very,” “unfortunately,” “really,” and “actually.” The author also notes that models can identify paragraphs that could be moved to improve clarity. To make use of this capability, the author suggests a workflow: first write a draft, then ask the model to spot problems, rewrite the flagged sections, and present the original and revised versions to a model that lacks context about the editing process to avoid the encouragement trap. The author shares a custom‑built tool that manages this workflow, describing its stack as Python, HTMX for interactions, a SQLite backend, Tailwind CSS for the frontend, and a Notion‑style prose editor with highlighting, sidebar commentary, revision tracking, and the ability to flag major changes. The tool runs editing prompts through interfaces such as Codex, Claude, or Antigravity CLIs. The author concludes by noting that even when the model suggests the piece is too long—as GPT5 reportedly did for this very article—the writer may choose not to act on the advice, preferring to stay true to their own voice.
Key facts
- The author proposes two rules for using LLMs in writing: do not adopt any suggested wording and avoid the model’s encouragement.
- LLMs are useful for detecting problems such as passive voice overuse, nominalized verbs, repeated phrases, and filler words like “very” and “actually”.
- The author’s workflow involves writing a draft, having the model flag issues, rewriting flagged sections, and comparing original and revised versions with a context‑free model to avoid encouragement bias.
- A custom tool built to support this workflow uses Python, HTMX, SQLite, Tailwind, and a Notion‑style editor with features like highlighting, sidebar commentary, and revision tracking.
- The model can be prompted through interfaces such as Codex, Claude, or Antigravity CLIs.
- The author fed the article to GPT5, which suggested the piece was 20% too long, but chose not to edit it.