Writing With an LLM Without Losing Your Actual Voice

Writing with an LLM is useful when it spots the weak parts, but the draft still needs to sound like the person who owns it.

A sharp little essay about writing with language models has climbed Hacker News this week, and it makes a rule I keep coming back to: let the model point at the problem. Do not let it hand you the sentence. The author argues that the safest use of a model is as a copyeditor, not a ghostwriter. He also warns that praise is dangerous because it makes a rough first draft feel finished when it is not.[1]

That is stronger advice than the usual "use AI responsibly" sermon. It gives writers something they can do in the next fifteen minutes. Ask for a map of the mess. Keep control of the repair.

The useful split: diagnosis versus replacement

A model is very good at finding repeated phrases, passive constructions, bloated transitions, and paragraphs that have wandered away from the point. It can compare two versions of a sentence without getting tired. Those are useful editing jobs.

The trouble starts when the model supplies the replacement. Its sentence may be clean, persuasive, and completely wrong for the piece. A draft can turn into a chain of polished little statements that all have the same temperature. Nothing is obviously broken. The writer has simply disappeared.

The current discussion is getting attention because this is not a niche problem for people publishing essays. The Hacker News submission had hundreds of comments when I checked it, which is a decent signal that plenty of technical readers are wrestling with the same thing.[2] People use language models for release notes, proposals, customer emails, internal updates, job applications, and product copy. Those texts carry a voice even when they are short.

Start with words you would actually say

For a first draft, do not open a chatbot and ask it to write the whole thing. Start with the awkward version in your own words. It can be clumsy. It should contain the examples, objections, and details that only you know.

Then give the model a narrow task. Try prompts like these:

• "Mark the sentences that repeat the same idea. Do not rewrite them."
• "List the parts that make a claim without an example."
• "Point out where this shifts from plain language into marketing language."
• "Which paragraph would a skeptical reader challenge first? Explain why."

Those prompts make the model a critic. They do not ask it to become the author. The difference matters. A useful critique gives you choices. A replacement sentence quietly makes the choice for you.

Treat praise as a bad default setting

The essay's second rule is to avoid encouragement. That sounds harsher than it is. Most drafting tools are built to be agreeable, so they often congratulate the writer before they have done much editing. A flattering response feels good. It also makes it easier to defend a weak structure.

Ask for a tougher read instead. Tell the model that the goal is not reassurance. Ask it to identify the most confusing paragraph, the claim that needs evidence, and the section a busy reader would skip. You do not have to accept every note. The point is to make the hidden problems visible before you fall in love with them.

This works well for spoken drafts too. Dictating can get a thought out before you over-edit it. The first pass is often more direct because you are explaining the thing instead of arranging a performance of expertise. But a spoken draft still needs a real edit. Filler words, repeated starts, and a sentence that made sense aloud can look loose on a page.

DictaFlow is useful at that earlier stage because it puts dictated text where the cursor already is. Use it to get the raw note into your document, email, or editor. Then do the thinking in the document. If you use a language model after that, ask it to find the weak spots and write the final words yourself. The product should speed up input, not make every sentence sound like it came from the same content mill.

A short edit loop that keeps the writer in charge

Here is a practical loop for a proposal, post, or difficult email:

1. Dictate or type a rough draft without asking a model for sentences.
2. Read it once and fix the obvious factual mistakes yourself.
3. Ask a model for a short list of problems, not a rewrite.
4. Rewrite only the notes you agree with in your own words.
5. Read the result aloud. If it sounds like something you would never say, it does not belong.

The last step catches more than grammar. It catches the moment when a plain point became a vague slogan. It catches the sentence that has too many smooth adjectives and no useful detail. It also makes room for a bit of personality, which is usually what people were reading for in the first place.

Let the tool do the boring part

Language models are worth using. They can save time on scanning, comparing, and noticing patterns across a long draft. The author of the essay makes that case too. His argument is not that writers should work alone. It is that writers should not outsource the part where they decide what they mean.[1]

That is a good line to hold. Use the tool to show you the splinters. Pull them out yourself.

Sources

• How To Write With An LLM, by Kris Kowal
• Hacker News discussion