Lessons from using LLMs as writing assistants, from someone who's trying to write more

After years of avoiding these tools for writing, I’ve finally given up and I am now trying to see if they can help me write more, better, or both. Just for clarity: there's a lot more to say about what these tools do and don't do, but for now, I only want to share how they affect my very own experience of writing with their help.

Feed it 95%, and it'll do the remaining 3%

When given an almost-ready-to-publish piece, assistants can take care of all the small cosmetic issues I can't even see: grammar, syntax, correct tense suggestions. It fixes my non-native English without killing the meaning, knows when “I was” should really be “I’ve been”, which I still can't do intuitively. It’s like having a native speaker clean up my final draft.

The last 2%? I found myself editing every single piece that was presented to me as "ready for publication", because I felt like it was erasing too much of my personality, tone or accent. It uses "better" words, but these words aren't part of my vocabulary. It's like over-ironing all the rough edges that make something uniquely mine.

Feed it 50%, and it'll take all the shortcuts to 95%

Think about writing as a puzzle, or a maze you're trying to get out of. An idea takes me somewhere, then takes a left turn, straight into a dead end. I walk back to the intersection where I came from, take a deep breath, then take another path, and repeat the process until I find the exit. A writing assistant will take the same left turn, but just destroy the wall in front of it, and do the same at every intersection or dead end. It will repeat the process at every obstacle, and technically exit the maze. But did it actually solve the maze?

The main flaw I can think of when using an assistant at this stage is that it only thinks forwards. Never sideways, nor backwards. It doesn’t care about idea progression or evolution. If I’m writing a three-part idea and ask it to expand part one, it won’t realize it’s breaking the flow to part two. It won't realize it drifted too far away from the original point. It just keeps going forward, all the time.

It never pushes back. It never asks “are you sure this belongs here?”, and will seamlessly connect two ideas based on the words they contain, and not because of what these words mean when put together. You can get from a rough draft to a bloated piece in just a few prompts. It doesn’t know when to kill a paragraph, rethink a sentence, or leave something unsaid. You have to explicitely ask it to stop, go back, or undo. Otherwise, it just keeps pushing forward.

Feed it 10%, and you still need to feed it 40% more

Writing assistants don’t have good or bad ideas. They only return a well-formatted version of your own idea, whether it's good or bad.

I started writing a fictional script for a TV commercial, and fed it with only a few sentences as the setup. My following prompts were about other ideas and tangeants connecting this initial setup. I ended up with a perfectly structured campaign pitch that read well, and that could easily be used as the foundation of a pitch in a real agency.

The problem is that this was a bad idea to start with. The core concept of this pitch was just weak, and it wasn't challenged at any point in the conversation. When I reread the thread from the start, I know I should have given up at the third or fourth prompt. I was just too distracted by the shiny appearance of my own bad idea, nicely organized in bullet points.

Feed it 5% and it'll make you believe you wrote 100%

I started writing an idea that clearly borrowed from authors who’d explored similar ground before. I asked it for references, and it didn’t recognize or find the authors I was clearly referencing. I then fed it with the exact names and titles I was thinking about, and asked how their works related to what I had in mind. This is what it returned to me:

Your angle is fresh: taking [...] and using it to expose a deeper inconsistency in [...]. If you're looking to cite existing voices, people like [author 1] and [author 2] come close — they [...] — but they don’t explicitly argue that [...]. That distinctive claim appears to be yours to own.

At this moment, I realized it was just unable to read between the lines. And also, how dangerous it can be to have an echo chamber giving you that level of confidence, telling you exactly what you want to hear.

When you start with just a seed of an idea, hoping the assistant will “take it somewhere,” be assured that it absolutely can do it, effortlessly. But once you're there, you still need to wonder if this is even where you wanted it to go in the first place. Assistants will never tell you “this is a bad idea”: they’ll just polish it, structure it neatly, tell you how good you look today, and hand your repackaged idea back with a smile.

So?

Writing assistants are brilliant proofreaders and reliable autocomplete machines, all in one. But they only shine after you've done the hard part: figuring out what you actually want to say. Or critical thinking, if you prefer calling it that way.