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

The Core Thesis: Copyediting vs. Ghostwriting

To maintain a distinct human voice and avoid the "uncanny valley" of AI-generated text, LLMs should be used as copyeditors rather than ghostwriters. Readers can often detect LLM-generated prose—which frequently registers as "output" rather than "writing"—making it essential for the author to handle all actual drafting. The primary utility of an LLM in the writing process is its ability to mechanically flag flaws in a completed draft without becoming fatigued.

Two Rules for Preserving Human Voice

To prevent "LLM-creep" and the dilution of personal expression, writers should adhere to two strict constraints:

Rule 1: Never use a word suggested by the LLM

Frontier models are designed to select pleasing, headline-like turns of phrase. However, using these suggestions often results in writing that feels artificial or "Velveeta-like." To avoid this, any specific phrase or word suggested by the model should be considered off-limits, regardless of whether it seems "better" than the original.

Rule 2: Forbid and ignore encouragement

LLMs are prone to praising the user's work, which can lead a writer to double down on flawed first-draft impulses. Because the process of rethinking and replacing paragraphs is a "load-bearing" part of developing a unique voice, the model's tendency to validate poor structure or incoherent flow is detrimental. Writers should explicitly forbid the model from providing encouragement and remain hypervigilant about any praise received.

Practical Applications for LLM Copyediting

While LLMs should not write the prose, they are highly effective at identifying mechanical and structural issues that are tedious for humans to spot:

  • Mechanical Tics: Identifying the overuse of passive voice, nominalized verbs, and filler words (e.g., "very," "unfortunately," "really," "actually").
  • Structural Clarity: Suggesting where paragraphs might be moved to instantly improve the flow and clarity of the piece.
  • Factual Accuracy: Using the model to check for factual errors or identify documentation that might undermine the central thesis.

Recommended Workflow

  1. Draft independently: Write the piece entirely without AI assistance.
  2. Identify problems: Ask the model to spot specific problems in the writing.
  3. Rewrite manually: Rewrite the offending sections yourself based on the flagged issues.
  4. Blind comparison: Present the original and the new version to a fresh model instance (without the context of the editing process) to determine which is more effective.

Community Perspectives and Counterpoints

Discussion among technical writers and developers reveals a spectrum of opinions on the efficacy of this approach:

The "Don't Use AI" Camp

Many critics argue that writing is an extension of thinking, and outsourcing any part of the process—even copyediting—stunts intellectual growth and creativity. Some users noted that reading AI-influenced text is stressful and diminishes the enjoyment of reading.

The Utility of AI as a "Vibe-Based Dictionary"

Some writers find value in using LLMs as a sophisticated thesaurus to find words with specific connotations or idioms in a foreign language, arguing that the final choice remains the author's voice.

Alternative Workflows

Users suggested several alternative methods for integrating AI into the writing process:

  • Narrative Exploration: Using LLMs to brainstorm bullet points and narrative arcs before drafting prose.
  • Cross-Language Drafting: Drafting a first version in a non-target language via AI, then manually translating and rewriting it into the target language to avoid "meatproxying" AI prose.
  • Narrative Coherence Checklists: Applying a strict checklist (e.g., unifying vocabulary across code and comments) to ensure technical clarity in commit messages and PR descriptions.

The "Taste" Requirement

A significant counterpoint is that the "copyeditor" approach requires the author to already possess a high degree of writing taste. Without a foundational understanding of style—such as that found in Style: Lessons in Clarity and Grace—a writer may not be able to distinguish between helpful structural advice and "LLMese" suggestions.

Sources

Related

  • Dispatch
    Why LLM Critics Are Right Yet I Still Use LLMs – Lessons from Local‑First Conf 2026

    The author acknowledges every major criticism of large language models—copyright, environmental cost, trust erosion, and geopolitical risk—while explaining why they continue to rely on LLMs and sharing concrete patterns for safe, high‑quality use.

  • Dispatch
    Write Code Like a Human Will Maintain It: Avoiding the AI Feedback Loop

    Using LLMs to generate code without adhering to maintainability standards creates a negative feedback loop where the AI learns and replicates bad patterns from the existing codebase.

  • Dispatch
    Why Some Developers Dislike LLMs: A Deep Dive into Martin Fowler’s Perspective and Community Reactions

    Martin Fowler’s “I don’t like LLMs” essay reveals a visceral discomfort with LLM‑generated language and hallucinations, and the Hacker News discussion highlights mixed feelings about trust, anthropomorphism, and practical workarounds.

  • Dispatch
    The 'This is LLM' Phenomenon: Navigating the Noise of AI Detection Accusations

    An exploration of the latest trend of low-quality comments accusing others of using LLMs, and how community members discuss strategies for maintaining discourse quality on platforms like Hacker News.

  • Project
    Tomsawyerhu/Chinese-WebNovel-Skill

    A modular, retrieval‑augmented prompt framework for AI‑assisted Chinese web‑novel writing. It splits the workflow into a short router (`SKILL.md`) and many self‑contained sub‑skills (concept planning, opening, plot logic, dialogue, consistency review, etc.). Each sub‑skill ships tutorials, runtime rules, good/bad examples, and may use local novel corpora via a provided search script. The system routes a user’s brief idea to the appropriate module, pulls relevant example excerpts, composes a concise prompt, generates text, and finally runs a consistency review. Designed for Chinese novel authors, prompt engineers, and researchers interested in modular LLM writing pipelines.