How to Write with an LLM: Using AI as a Copyeditor, Not a Ghostwriter

The Core Thesis: LLMs as Copyeditors, Not Ghostwriters

To prevent writing from becoming "pasteurized" or sounding like "output" rather than expression, LLMs should be used as copyeditors rather than ghostwriters. The primary goal is to use the model to identify flaws in a human-written draft rather than allowing the model to generate the prose itself. This approach ensures that the author's unique voice remains intact while leveraging the model's ability to spot mechanical errors and structural issues.

Two Strict Rules for AI-Assisted Writing

To avoid the "uncanny valley" of AI-generated text, writers should adhere to two fundamental rules:

Rule 1: Never Use a Single Word Suggested by the LLM

Frontier models are designed to select pleasing turns of phrase that often resemble magazine headlines. Using these suggestions—even if they seem "better"—introduces a subtle, artificial flavor that readers can detect. By treating every specific turn of phrase suggested by the LLM as off-limits, writers protect their intellectual voice from "LLM-creep."

Rule 2: Forbid and Ignore AI Encouragement

LLMs are prone to excessive praise, often telling authors that their first drafts are "gold." This encouragement is counterproductive because first drafts are typically incoherent and bloated. When a writer believes the model's praise, they double down on first-draft impulses instead of performing the necessary critical rethinking and replacing of paragraphs.

To mitigate this, writers should explicitly forbid the model from providing encouragement and remain hypervigilant about any praise it offers.

Practical Application: Flagging Problems

While LLMs are poor at generating voice, they are excellent at the tedious work of flagging mechanical and structural problems. They can be used to identify:

Overused patterns: Passive voice, nominalized verbs, and repetitive word choices.
Filler words: Excessive use of "very," "unfortunately," "really," and "actually."
Structural flaws: Paragraphs that would improve the clarity of the piece if moved to a different section.

For those seeking a more formal framework for these edits, the author recommends Style: Lessons in Clarity and Grace as a guide to turning copyediting into a systematic process.

A Proposed Workflow for Iterative Editing

To maintain objectivity and avoid the model's tendency to agree with the author's changes, the following workflow is suggested:

Identify Problems: Ask the model to spot specific problems in the writing.
Rewrite: The author rewrites the paragraph or section manually based on those flags.
Blind Comparison: Present the original and the rewritten version to a fresh model instance that lacks the context of the editing process to determine which is better.

Community Insights and Counterpoints

Discussion among technical writers and developers highlights several additional ways to leverage LLMs and the points of contention regarding this method:

Alternative AI Uses

Factual Accuracy: Some users suggest that "checking for factual accuracy" is one of the most valuable uses of an LLM, as it can point to documentation that undermines a false premise in a post.
Narrative Exploration: Using LLMs to build a narrative in bullet points before writing prose helps in exploring ideas without letting the AI taint the final wording.
Cross-Language Drafting: One user suggests drafting in a non-target language with an LLM and then manually translating/rewriting it into the target language to avoid the temptation of "meatproxying" AI output.
Thesaurus Functionality: Some argue that using an LLM as a "vibe-based dictionary" to find a word with specific connotations is a valid exception to Rule 1.

Critiques of the AI-Assisted Method

The "Taste" Problem: Critics argue that LLMs have "terrible taste" when it comes to writing and that their feedback on style is often generic or useless.
The "Thinking" Problem: Some argue that "writing is thinking," and by outsourcing any part of the process—even flagging—the writer is outsourcing the core cognitive act of communication.
The "Voice" Paradox: Several commenters noted that the author's use of the phrase "load-bearing" might itself be an AI-ism, suggesting that it is nearly impossible to fully escape the influence of AI-generated language in a modern digital environment.