# How I write with LLMs (it's not faster, it's just different)

By Nikolai Skavinskii (https://nikska.com) — 2025-08-15
Canonical: https://nikska.com/blog/writing-with-llms

> *Originally posted on [LinkedIn](https://www.linkedin.com/in/nikska/) in August 2025, back when GPT-5 and Claude 4.1 were the news of the day. Every word still stands.*

![Using LLMs to write text — tips to keep your human voice](/blog/writing-with-llms.webp)

GPT-5 is out, Claude 4.1 is here. Writing texts with them still sucks (unless you're okay with "YOU'RE ABSOLUTELY RIGHT!")

Here are my quick tips on how to write with LLMs.

**TL;DR — it's not faster, it's just different.**

Writing is a human task. Nature's ["Writing is Thinking"](https://www.nature.com/articles/s44222-025-00323-4) kinda says it all.

With this off the table, how do I write with LLMs to avoid cringe and slop.

## 🧠 Split the work by strengths

Leverage the best in humans — expertise, wit, passion, and LLMs — speed, summarization, reasoning.

## 🎙️ Talk, don't type

Use voice mode and throw all your ideas on the topic into it. Do not care too much for the structure just yet. The most important part — talk to the machine (yes, use your vocal chords).

## 🏗️ Ask for structure only

Ask the LLM to produce only the structure, a Table of Contents.

## ✍️ Read it... then delete it

As counterintuitive as it sounds, you now have way more than 15 minutes ago when you just started. Let me explain:

- You spoke through the main points, which actually reinforced the best of your ideas and highlighted the weak parts way better than if you just typed them
- You used the best feature of LLM — restructuring data. You're lucky if it made mistakes or allowed itself some inaccuracies — now you know what to avoid and how to write it better
- And now you have a better understanding of the subject. Go on and write it down, make your own mistakes!

## 💫 Bonus: turn the LLM on your draft

Send the final text to the same LLM and ask it two things:

- List some questions as if they were a member of your audience. What would they want to expand on?
- List some harsh critiques from the standpoint of a nitpicky ICP. While this part is usually already pretty sloppy, some of the points GPT or Claude produces are quite useful.

---

Okay, this process might seem more lengthy. Because it is. But for me, it's way, way more engaging, and it allows me to produce more meaningful texts, not less. I still don't use LLMs to produce text anyone would read (if you catch me doing that, you're allowed to hit me in the knee.)

This post was written using this approach, and I even decided to pretend to be an LLM and place emojis here and there. Did I do a good job?