How to Write with an LLM
summary
Large Language Models should be used as copyeditors rather than ghostwriters to prevent writing from becoming generic and artificial. The core strategy is to let the AI identify flaws and structural issues while the author retains total control over the actual wording and revisions. By treating the AI as a tool for flagging errors rather than a source of content, writers can maintain their unique voice while benefiting from the AI's efficiency in spotting tedious mistakes.
key points
- Never use specific words or phrases suggested by an LLM because they tend to be overly polished and generic, which makes the writing feel like artificial output rather than human expression.
- Ignore AI praise and encouragement because models often validate poor first drafts, which prevents the author from doing the necessary hard work of rethinking and replacing weak sections.
- Utilize AI for the tedious mechanical work of flagging passive voice, repetitive phrasing, and structural inconsistencies, allowing the author to handle the actual rewriting process.
who should read it
Writers and programmers who want to enhance their prose using AI without sacrificing their personal style or authenticity.
discussion skeptical
consensus
Commenters generally agree that while LLMs can eliminate uselessly brief messages like "fix", they often replace them with verbose, "deep-fried" jargon that lacks actual intent. There is a strong consensus that the most valuable part of a commit message is the 'why' behind a change, which is often missing from diff-based AI summaries. Many suggest that human review or strict prompting is necessary to prevent AI-generated slop from cluttering project histories.
top insight
Generating commit messages solely from a diff is fundamentally flawed because the diff shows what changed, but not why it changed. The most useful commit messages provide external context, such as links to bug trackers or design justifications, which are not present in the code changes themselves. Therefore, an LLM must be the same agent that wrote the code in the same session to accurately capture the author's original intent.
Generated by an LLM from the article and thread. Check the sources before relying on it.