Show HN: I built a tool to fix the problem in LLM replies

1 points by rocksalad 4 hours ago

Hi HN,

I've been using LLMs to help draft replies and content, but I hit a wall: the "default" voice of standard models is painfully obvious. It’s always too cheerful, uses emojis incorrectly, and loves words like "delve" and "testament."

I realized I was spending more time editing the AI's output to sound like me than it would take to just write the thing myself.

So I built PostOwl to solve the "Style Alignment" problem.

How it works: Instead of a generic wrapper, it uses your previous posting history to build a style profile. When you generate a reply, it dynamically injects your specific vocabulary, sentence length patterns, and tonal constraints into the context window (using a mix of RAG and few-shot prompting) to force the model to mimic your syntax.

The Challenge: It’s effectively a "Doppelgänger-as-a-Service." The hardest part has been balancing the speed of generation with the accuracy of the style mimicry.

I’d love feedback on the output quality. Does it actually sound human, or just like a "pirate mode" version of ChatGPT?

(It has a free tier, no credit card required. I'm just looking for load testing and feedback on the prompts right now.)