Train an AI to Write Like You: Fine-Tune a Local LLM (2026)

You can train an AI to write like you on the Mac you already own, for $0, with less of your writing than you'd think. Four routes exist to fine-tune an LLM on your own writing: tune the weights with LoRA, install a packaged app that tunes for you, write a style prompt, or index your notes with retrieval. This post maps all four with real corpus numbers and the failure cases most guides skip. It builds on our local LLM writing stack: that post assembled the gear, this one makes it sound like you.

The pain is specific. Writekin's developer got here after a prompt-based fix failed: he stripped the usual AI tells with a style rule, and "Pangram still came back as ~100% AI written" on text that was mainly his own writing with AI tweaks (Show HN). Scrubbing the surface didn't change the voice underneath. Changing the weights does.

Key Takeaways

• Four routes: LoRA fine-tuning, a packaged app (Writekin), a style prompt, or RAG. Voice comes from the first two; facts from RAG; everything else is a prompt.
• Your style is a small signal: stylometry fingerprints a voice in 150 to 500 words, and Unsloth's floor is 100 rows (antirez, Unsloth).
• A 7B model trains in about 6GB of memory with 4-bit QLoRA, so a 16GB Mac works (LLaMA-Factory).
• Overfitting is the real risk: training loss below 0.2 signals memorization (Unsloth).
• OpenAI's fine-tuning platform is winding down as of September 2026; local stacks are the open door (OpenAI docs).

What Does "Writing Like You" Actually Mean?

Your style is a small, learnable signal, not your whole mind. Stylometry, the forensic science of identifying authors by their prose, can fingerprint a style from remarkably little text. In his 2025 reproduction, antirez cites research that "you need even as little as 150 words" and that "the range from 150 to 500 is considered to be optimal" (antirez).

That reframes the project. A voice is mostly word choice, sentence rhythm, punctuation habits, and the structures you lean on; those are statistical properties, and they compress small. When a Hacker News commenter asked what LoRA actually does, the answer that stuck: people expect knowledge addition from fine-tuning, but with LoRA "it's about style transfer" (HN). LoRA changes a low-rank slice of the model, exactly the shape of a personal style.

Steve Krouse's 2026 experiment makes it concrete. He built a rubric-based eval scoring drafts against his voice, iterating until his posts scored above 90 and everyone else's below (Steve Krouse). His rules included instant fails for wrong dash styles. A style is a list of measurable habits, and lists are learnable. The ceiling: the AI imitates your habits, not your judgment, which is why the route and corpus questions matter more than the model question.

What Are the Four Routes to an AI That Writes Like You?

Four routes, four different problems. Route A tunes model weights on your corpus. Route B installs an app that does Route A for you. Route C writes your style into a prompt. Route D retrieves from your past writing. Practitioner consensus across the threads cited throughout: knowledge and facts go to RAG, voice and style go to LoRA, everything else goes to a good prompt.

Route A: Tune the weights yourself

Route A is the direct path: pick a small open model, prepare a dataset of your writing, run a LoRA or QLoRA fine-tune. The 2026 field, all open source:

• MLX and mlx-lm (6.9k stars, MIT) is Apple's Mac-native framework and the default for Mac users: one command fine-tunes on Apple Silicon, and QLoRA kicks in automatically on quantized models (mlx-lm). MLX shipped v0.32.2 in late August, per its releases page (MLX releases).
• Unsloth (75.6k stars, Apache 2.0 core) is the performance option, claiming 2x faster training with 70% less memory, exporting GGUF files you can run in Ollama (Unsloth).
• LLaMA-Factory (74.6k stars, Apache 2.0) fine-tunes 100+ models through a web UI with zero code, and publishes the clearest memory table of the four (LLaMA-Factory).
• Axolotl (12.4k stars, Apache 2.0) is the power-user CLI, one YAML file, but needs a modern NVIDIA or AMD GPU (Axolotl).

That memory table answers "can I do this on my machine?" From LLaMA-Factory's README, 7B column: full 32-bit tuning needs 120GB, LoRA at 16-bit needs 16GB, QLoRA at 8-bit needs 10GB, and QLoRA at 4-bit needs about 6GB (LLaMA-Factory). A 4-bit QLoRA of a 7B model fits a 16GB Mac with room to spare. For base-model picks and running the result, our local LLM writing stack covers the runner side.

Route B: Let Writekin do it

Writekin, v0.9.0 on July 28, 2026, is the packaged Route A: a native Mac app that ingests your Mail, Messages, documents, and chat exports, curates them into a corpus, then QLoRA-fine-tunes a local model via MLX (GitHub). It wants an Apple Silicon Mac with "16 GB unified memory minimum; 32 GB+ recommended for training larger models," and nothing leaves your machine. Its founder confirmed an M4 with 16GB works, possibly with a smaller model (Show HN).

Its positioning is the clearest explanation of why tuning beats prompting for voice: it "changes the model's actual weights rather than just its instructions," a "different and more durable change" (Writekin site). Prompts get truncated or half-followed. Weights don't.

Two safeguards are worth copying into any homebrew pipeline: a verbatim-overlap check that flags generated text matching your corpus too closely, catching memorization instead of style, and a self-ingestion guard that keeps its own drafts out of your training data.

One catch: the license is PolyForm Noncommercial, free for personal use, commercial use needs a separate license. And the category is mortal. Mimicr, a 2025 "AI that writes like you" launch, is already DNS-dead while its HN thread survives (HN).

Route C: Write a style prompt

The cheapest route. Claude's personalization features include account-wide instructions, per-project instructions, and Skills, which can "Apply communication patterns based on your own writing or preferences" (Anthropic). OpenAI's Prompt Personalities cookbook teaches the same lever: define the personality explicitly, because it "improves consistency, reduces drift," and personality "should not be treated as aesthetic polish, but as an operational lever" (OpenAI cookbook). Paste your best 1,000 words plus a rule list, and a frontier model does a passable imitation.

Practitioners split on how far that gets you. On the Mimicr thread, skeptrune argued frontier models work if you paste a sample and ask them to emulate the style; the founder countered that the ones he tried "honestly aren't great at truly emulating style, even with context" (HN). The decisive view came from gamegoblin on OpenAI's GPT-4o fine-tuning launch: prompt caching and few-shot prompting get you most of the way "except for cases where you are really trying to perfectly match a particular writing style" (HN).

Steve Krouse's eval is Route C's upgrade: a scored, measurable definition of your voice, no training. His verdict after living with it: "The AI-produced results weren't great, so I spent a lot of time editing," and the loop was "mostly not very helpful for me yet" (Steve Krouse).

Route D: RAG on your own writing

Retrieval-augmented generation puts your actual notes in the model's context instead of baking them into weights. In a 138-point thread on training a personal LLM, consensus formed around one line: fine-tuning is "lossy compression of your notes into the model weights," while RAG retrieves the notes losslessly (HN). Tooling is mature: Khoj (37.1k stars, AGPL) turns any local or online LLM into a searchable second brain (Khoj), and GPT4All's LocalDocs does private document chat with no GPU.

The catch: RAG retrieves facts, not voice. It answers "what did I write about X" in your themes, but it doesn't make fresh prose sound like you. One diarist in that thread indexed 153 days of daily writing, roughly 300,000 to 450,000 words by our arithmetic, into GPT4All on an 8GB MacBook Air and got "remarkable, almost spooky" answers about his own thinking (HN). That's a great knowledge tool. It isn't a voice tool.

Route	Hardware floor	Corpus	Best for
A: LoRA or QLoRA	~6GB (7B at 4-bit)	100 rows to start	Durable voice on new topics
B: Writekin	16GB Apple Silicon	No hard minimum	Hands-off tuning on a Mac
C: Style prompt	Any device	A pasted sample	Quick drafts in your rough tone
D: RAG	8GB works	Your notes, indexed	Facts from your past writing

How Much of Your Own Writing Do You Need?

Less than you think, and the honest answer is a range. Unsloth's dataset guide sets the vendor floor: "at least 100 rows" as a bare minimum for reasonable results, with "over 1,000 rows" preferable (Unsloth). Writekin's site says "There's no hard minimum," with more and varied writing yielding better results (Writekin site). Classic practitioner guidance from the 2023 OpenPipe wave: "50-1000 examples" (HN).

Rows need translating into words. A row for style training is one passage of your prose, and the practical floor is the stylometry floor: if 150 to 500 words can fingerprint a style, the signal in a 50,000-word blog archive is already dense. Our ladder: 5,000 to 20,000 words of clean, self-written prose gets a style shift; 50,000 to 150,000 across varied topics gets something that survives real use.

Quality beats quantity: as Unsloth puts it, effectiveness depends heavily on dataset quality, so clean, deduplicate, and diversify narrow sets to prevent overfitting. Strip other people's text before training; quotes and forwarded mail teach the wrong voice. Writekin's insight: your Messages app is bulk prose, already on your Mac, already in your voice.

How Does the Fine-Tune Actually Work?

Mechanically, a style fine-tune is an afternoon of work. The dataset is a JSONL file, one example per line. For pure prose style, mlx-lm's raw text format is the right shape: each line is simply {"text": ...} with a passage of your writing. If you want instruction-following, pair old outlines with finished posts using the {"prompt": ..., "completion": ...} format. One warning from the docs: "Longer examples require more memory" (mlx-lm LoRA guide).

The train command, verbatim: mlx_lm.lora --model <path> --train --data <path> --iters 600. QLoRA is automatic if the model path points to a quantized model. Adapters land in adapters/, you fuse them with mlx_lm.fuse, and you can export GGUF to run the tuned model in Ollama, the runner from our local stack. Speed reference from the doc's own example: a Mistral-7B QLoRA run hit about 250 tokens per second on an M1 Max with 32GB.

Overfitting is the failure you'll actually hit, and the guardrails are numeric. Unsloth's guide: training beyond 3 epochs offers diminishing returns and increases overfitting risk, and a training loss below 0.2 means likely overfitting (Unsloth). Fixes: stop at 1 to 3 epochs, raise weight decay, add LoRA dropout around 0.1, expand the dataset rather than epochs. Hold out a validation split so you can watch eval loss alongside training loss. If generated text matches your corpus verbatim, the model memorized instead of learned; that's the overlap check Writekin automates.

Why Is This Worth Doing Now?

Three forces converged. First, the research caught up with the annoyance: AI assistance measurably flattens voice. Wenger and Kenett's 2025 study found that "LLM responses are much more similar to other LLM responses than human responses are to each other" (arXiv). Anderson et al.'s 2024 36-participant study found ChatGPT users produced more ideas but "less semantically distinct" ones, and "felt less responsible for the ideas they generated" (arXiv). Cornell's April 2025 CHI study put 118 participants through an autocomplete assistant and watched their writing converge, "primarily eroding Indian writing styles" toward Western norms (Cornell).

Wan and Kalman's 2025 study adds the encouraging part: homogenization "may emerge from uniform deployment practices rather than from an inherent limitation of GenAI" (arXiv). Uniform defaults flatten you; deliberate personalization is the countermeasure.

Second, the cloud door is closing. As of September 2026, OpenAI's docs state its fine-tuning platform is "winding down," closed to new users (OpenAI docs). The local LLM vs API calculus now has a hard edge: the mainstream cloud option is exiting just as laptop QLoRA got good. Together AI still sells LoRA fine-tuning at $0.48 per million tokens for models up to 16B with a $4 job minimum (Together AI), cheap in dollars, expensive in a different coin: your corpus, Mail and Messages included, on someone else's GPUs.

Third, the cycle matured. The 2023 fine-tune-your-own wave crested at 955 points on HN (HN); 2024-25 brought the disillusion, a 193-point thread titled "Fine-tuning LLMs is a waste of time" (HN); 2026 brought the narrow revival: voice, on hardware people already own, with Writekin and MLX shipping monthly.

When Should You NOT Fine-Tune?

Fine-tuning for voice works; fine-tuning for other things mostly doesn't, and the failure catalog is documented.

Catastrophic forgetting is real. Improving fit on your fine-tuning data decreases fit on the original corpus, and sota_pop's verdict on his runs: "My experience has been that it always hurts generalization" (HN). Your tuned 7B may write like you and get worse at everything else.

Bad data breaks models fast. jondwillis fine-tuned GPT-3.5 on his texts and emails and it "pretty much generated schizophrenic output" (HN). He blamed poor curation, correctly. An uncleaned corpus is worse than no corpus.

Prompting covers most needs: "prompt caching gets you 99% of the benefits for 1% the effort" (HN). Small fine-tuned models are also brittle, "superficially good" with cracks that show on anything not tightly tied to the training data (HN).

Even the believers admit unevenness. Writekin's author: "It's v0.9 and the output is uneven," sometimes nailing your voice, sometimes "completely off" (Show HN). Steve Krouse's verdict on his voice-matching loop: "mostly not very helpful for me yet" (Steve Krouse). The realistic win is drafts needing less editing, not a ghostwriter. If a style prompt is 80% there, the last 20% may not be worth an afternoon of QLoRA.

On detectors: don't fine-tune to evade them; the goal is voice, not evasion. Detectors misfire both ways; GPTZero itself says no AI detector is 100% accurate, and false positives against human writers are documented, as we cover in our AI writing assistants roundup. A model trained on your real prose writes like you because it learned you.

Related Tools & Further Reading

• The foundation: our local LLM writing stack covers Ollama, GGUF exports, and the hardware tiers these fine-tunes plug into.
• Before you upload a corpus anywhere, the local LLM vs API comparison frames the privacy tradeoff.
• Cloud-side style options and their detector false-positive problems: best AI writing assistants.
• The full map: text writing tools.