WRITING STYLE LLM
I fine-tuned an open-weight language model on a corpus of my own writing so it could reproduce my voice — cadence, word choice, the way I structure an argument. This is the method, the tradeoffs, and a before/after.
// THE GOAL
Generic models write competent, forgettable prose. I wanted a model that sounded like me — useful for drafting at speed without losing voice. The task: teach an open-weight base model my style from my own writing.
// METHOD
Assembled a personal corpus — essays, notes, and long-form posts — cleaned and formatted into instruction/response pairs that isolate voice from topic.
Chose LoRA over full fine-tuning: adapters train a small set of low-rank weights, so it's cheap, fast, and easy to iterate without touching the base model.
Blind side-by-side comparisons against the base model on held-out prompts, scoring for voice match and coherence — not just fluency.
// BEFORE / AFTER
Same prompt: "Explain why our onboarding flow needs work."
Our onboarding flow presents several opportunities for improvement. By streamlining the user journey and reducing friction points, we can enhance the overall customer experience and drive higher activation rates across the funnel.
Onboarding is where we lose people, and we lose them for a boring reason: we ask for too much before we've earned it. Cut the first three steps. Show one win fast. Everything else can wait until they actually care.