AI-assisted writing (without the Slop)
Why LLM generated text is so weird, and how to make it work
I suck at writing. I open too many parentheses, and my thoughts scatter (everywhere).
So when ChatGPT launched, I thought it would finally replace Grammarly.
But LLMs have their own problems:
“It’s not just x—it’s y,”
Rhetorical questions? Affirmative answers!
“Here’s the kicker”: That preface was entirely unnecessary,
And in the end, it ends with recaps — that repeat everything already said, now with bullet points.
The problem with AI text is that when you read it, your first thought is: “Did this person actually invest time in this, or did they write a two-line prompt and expect me to read something they never even thought about?”
And as some people put it: “I’d rather just read the prompt.”
LLMs can’t be genuine because they don’t know how to be a person. They read text from multiple public sources and average it out. They weren’t trained by eavesdropping on authentic conversations or messages. (At least I hope not)
The more the AI creates for you, the worse the output becomes.
That’s why when you ask it to keep it casual, it turns into “How do you do, fellow kids?” and when you ask for a professional tone, it becomes “Alas, who’d’ve done this?”.
What You Need to Provide
If you want LLMs to cook, you need to provide ingredients.
As a general writing (and cooking) tip, start it raw. Don’t use autocorrect. In fact, don’t even look at what you’re typing. Close your eyes and let raw ideas flow, along with grammatical mistakes and misconstrued sentences. Just make it coherent enough. Make bullet points to answer: “What’s the point of me writing this?”
Connect those bullet points with your personality, which dictates how you link sentences. A serious person uses serious connectors; a casual person throws in verbal expressions (and memes).
How LLMs can help
When you have the first draft, the key is using the right edits. The biggest mistake people make is while prompting. If you prompt like a casual writer, it treats you like one.
Saying, "Improve the text below for my email,” makes the AI slopify everything: it accesses the neural latent space of “This person needs my help immensely.”.
You need to signal, “Hey, I know what I’m writing. I just need help improving the flow while keeping my own words.”
You can do this by using the verbiage editors and publishers use during the different editing phases, from solidifying the overall scope to minor edits like correcting grammar.
While the LLM won't write for you, it can help you immensely, because writing words is not the hard part once you get the hang of it. For me, the editing takes 80% of the overall time. Most people start as slow writers because they try to write and edit simultaneously.
Types of editing (and how they help)
With chain-of-thoughts in newer models, you don't need much prompt engineering anymore. You just need to know the right words so the LLM's thinking can go into the right embeddings.
When publishing a book, the editing steps go from the most destructive (major rewrites) to the least destructive (removing typos).
Also, as a metaexample, I’ve added the editing phases of this article here. You can see how the content changed from draft to final edit.
Let's go in that order as well:
Content Editing
Content editing improves flow, structure, and clarity. It is useful when you know what you want to say but are unsure how to connect thoughts.
It’s the most destructive, so it's better only to use it once.
Example Prompt: “You are a content editor. Improve the flow of the sentences and make the text stronger and more structured. <content>”
The AI will make many edits to make your text “make sense”, and the places where the AI misunderstood your intentions will stick out like sore thumbs. You will need to adjust them and add points that solidify your premise.
As you add (and cut) content for a second draft, it's time to move to line editing.
Line Editing
Line editing is where AI shines, especially for short texts like announcements. Use this when you know what and how you want to say something, but specific words escape you, or phrasing could be simpler.
I spend most of my time here, line editing multiple times and rewriting content until no paragraph is a waste.
Example Prompt: “Line edit this (Slack message / blog post): <content>”
Proofreading
Proofreading happens when you’ve “mastered” the copy. It’s always safe to run multiple times without fearing the AI will destroy your voice, because you will tempt yourself to write small additional bits here and there.
Example Prompt: “Fix any grammatical mistakes in the text:”
This is basically a cheap Grammarly.
As you can see, “writing” is actually two things: drafting and editing. AI struggles with the first but can greatly help with the latter. Even with better models, I don’t think we’ll completely remove the AI scent from text generated during the drafting phase.
So we as humans will need to write until we’re exhausted and don’t even want to finish what we wro—