I want to rewrite some of my blog posts to sound more like a specific style of author (e.g., punchy, minimalist, Hemingway-esque). I've tried to just say 'write like X,' but it doesn't work well. Is there a better prompt approach? Should I provide snippets of the author's work in the prompt, or is there a 'style guide' template I can use to define these variables more clearly? I want a repeatable, consistent style output.
To replicate a specific author's style, provide the model with a reference dataset of the author's work and prompt it to analyze and adhere to defined quantitative metrics such as sentence structure, complexity, and vocabulary usage.
3 answers
To achieve a consistent output, you need to decompose the writing style into a modular template that Claude can process as a fixed instruction set.
- Create a reference document containing 5 distinct paragraphs of the target author.
- Identify 3 core metrics like sentence length, verb-to-adjective ratio, and transition word frequency.
- Build a prompt that requires Claude to evaluate the draft against these specific constraints before generating the final version.
- Use a chain-of-thought process where the model performs a style analysis on its own output to ensure compliance with the prescribed constraints.
Shahistha Sullad, your approach sounds very thorough. I’m just a bit worried I might mess up the metric analysis, but I suppose I should try following your template step-by-step to be sure.
Thanks for this, Shahistha Sullad! I’ve been trying to get Claude to mimic styles without much luck. Using a chain-of-thought process makes perfect sense, even if I'm still a little fuzzy on the specifics.
You should stop relying on generic style labels and instead leverage few-shot prompting by feeding the model a curated set of high-fidelity examples that define the specific syntactic patterns of your target author. A single descriptive prompt is insufficient because it lacks the granular context required to align the latent probability distribution of the model with the cadence you are looking for.
I remember back when I was fine-tuning models for legal compliance documentation, I struggled with the same issue until I stopped trying to force a general persona and started building a library of stylistic fragments.
I eventually realized that showing the model exactly how a specific author handles sentence length and subordinate clauses worked much better than asking for an abstract style like Hemingway. Providing these concrete samples allowed the model to internalize the rhythm of the text rather than just mimicking the surface vocabulary.
I appreciate the breakdown, Shahistha Sullad. I’ve been struggling with consistent outputs for weeks, and while this seems like a lot of extra work, it’s honestly the most logical path forward I've seen yet.