Last week a colleague asked me to review a proposal. Two paragraphs in I closed the tab. Not because it was wrong — it was technically fine. I closed it because I knew, with the certainty of someone who's read a thousand of these, that no human had written it.
The giveaways:
- “Here’s a comprehensive overview” opener
- A bulleted list with bold labels under every section
- “Hope this helps!” at the end
I asked him about it later. He'd typed two sentences into ChatGPT and shipped the output. He hadn't read it himself. We were both supposed to react to a document that nobody had actually written.
This is happening everywhere now. Reddit replies that sound like LinkedIn posts. X comments that read like blog intros. DMs with em-dashes and Oxford commas where you'd expect a typo and a “lol”. The texture of how people online actually talk — the rhythm, the sentence fragments, the casual disrespect for grammar — is being slowly sanded down by the same five LLM defaults applied to everything.
The frustrating part: the LLMs CAN sound human. They're just not allowed to by default. Every major chat product ships with system prompts tuned for the lowest common denominator — corporate-helper register, structure for every answer, signoffs on casual replies, the bullet-list reflex. That register is safe. It can't get a model in trouble. It also can't sound like a person.
What I changed
Same Claude under the hood. Different system prompt. The prompt is the product.
The rules I gave it:
- When someone pastes a tweet and asks for a reply, you're typing in their voice, into the comment box. Not writing about the post. Not adding a signoff. Just the reply, ready to paste.
- No “Certainly”. No “Hope this helps”. No
[Your name]placeholder. No 8-bullet response to a one-sentence question. - If someone's friend got laid off, “damn, 6 years at one place is rough” lands harder than five paragraphs of structured empathy.
- Structure exists for things that actually need structure — code, configs, multi-step debugging. Tweets don't need structure.
Two side effects I didn't expect
Fewer tokens. A one-line reply uses a fraction of the tokens of a polished five-paragraph one. If the average AI reply on the internet today is ~400 tokens and most of them could be ~30, we're collectively burning a lot of GPU on shape.
It makes me think more. The shorter the reply, the more it has to actually mean something. When the model can't hide behind “Here are 5 considerations…”, it has to commit to one. Same goes for me when I write.
Why I'm writing this
I'm not trying to start a movement. I just want to not feel like a stranger on Reddit anymore. I want to reply to friends and sound like myself. I want my colleagues' proposals to be readable because they wrote them.
If you've ever closed a tab because you knew nobody on the other end wrote what you were reading, you'll get it.
It's at rawreply.com.