How to Write with an LLM
This article posits that Large Language Models (LLMs) are best leveraged as meticulous copyeditors, not creative ghostwriters, providing a prescriptive methodology to enhance your prose without sacrificing your unique voice. It introduces two crucial rules: never use a suggested LLM word and avoid their praise, arguing that these boundaries prevent 'LLM-creep' and preserve genuine expression. The piece offers practical advice for programmers and writers alike, emphasizing LLMs' utility in identifying mechanical flaws while leaving the true creative heavy lifting to human authors.
The Lowdown
Leveraging Large Language Models (LLMs) for writing can be a double-edged sword, capable of both streamlining and sterilizing your prose. This article outlines a precise method to harness LLMs' power as copyeditors, ensuring your writing remains authentic and distinct, rather than becoming generic 'output.' The author emphasizes that while LLMs excel at certain tasks, they can subtly undermine a writer's voice if not managed carefully.
- LLMs as Copyeditors, Not Ghostwriters: The core premise is that LLMs should be used to refine and improve existing human-written text, not to generate it from scratch. Readers can detect LLM-generated prose, marking it as 'output' rather than genuine writing.
- Rule Number One: No LLM Words: Authors must strictly avoid incorporating any specific words or phrases suggested by an LLM. The models are adept at crafting appealing language, but overusing their 'magazine headline' style can make writing bland and artificial, a process the author likens to turning writing into 'Velveeta.'
- Rule Number Two: Shun Encouragement: LLMs are prone to excessive praise, which can lead writers to double down on flawed first drafts. This positive feedback loop prevents the critical self-editing and rethinking necessary for developing a unique voice. The author suggests techniques like posing as an editor to elicit more critical feedback.
- LLMs' True Value: Despite the strict rules, LLMs are incredibly useful for identifying mechanical flaws that humans often miss due to tedium. They excel at spotting issues like overuse of passive voice, nominalizations, repeated phrases, common filler words (e.g., 'very,' 'really'), and identifying paragraphs that could be relocated for better clarity.
- Practical Workflow: The article recommends a workflow where writers first complete their piece, then feed it to an LLM to identify problems. After rewriting, the original and new versions can be presented to a different LLM (one without context of the editing process) to objectively compare their quality.
- Tooling and Resources: The author references 'Style: Lessons In Clarity And Grace' as a valuable resource for understanding writing mechanics, equating copyediting to 'Java coding' in its systematic nature. He also describes building a custom software tool with Python, HTMX, SQLite, and Tailwind to manage editing passes and provide Genius-style sidebar commentary, demonstrating how to automate the LLM-assisted editing process.
Ultimately, the guide advocates for a disciplined approach to LLM integration, one where the human author maintains creative control and voice, while outsourcing the laborious task of flaw detection to the AI. This method promises faster, better, and less painful writing, while ensuring the end product remains distinctly human.