Opus 4 is the tone-matching model. Stop using it like a generalist.

The legal tech crowd has been quietly running the most demanding tone-matching workloads in the AI industry. A contract drafted by partner A reads nothing like one drafted by partner B, and getting that wrong is a malpractice problem, not an aesthetic one. The teams shipping into that space settled on Claude Opus for the final drafting layer. That’s a signal worth borrowing.

Marketers keep asking why their AI copy sounds generic. The answer is partly prompt design, partly model selection. Most of us are using a single model for ideation, structure, and final voice. Those are three different jobs.

What “fine drafting” actually means

Fine drafting is the last-mile work where you already know what you want to say and you need it to sound like a specific person or fit inside an existing document. It’s not generation. It’s interpolation against a tight style constraint.

This is where Opus pulls ahead of cheaper models in a way that’s hard to see on benchmarks. Sonnet will write you a clean paragraph. Opus will write you a paragraph that sounds like the three paragraphs you pasted above it. The difference shows up most when the source voice has quirks: a specific cadence, an unusual punctuation habit, a tendency to bury the verb.

For legal teams that means matching how a particular partner phrases an indemnity clause. For a marketing team it means matching how a founder writes on LinkedIn, or how a brand handles transitions between sentences, or whether a product page leans declarative or hedged.

The workflow worth stealing

The pattern I’ve seen work, and that I’m now running for two clients, looks like this:

A cheaper model (Haiku, or a mid-tier from another provider) handles research, outline, and a rough first pass. That draft gets the ideas in roughly the right order. It will sound like AI. That’s fine.

Then Opus gets a second prompt with three things: the rough draft, 1500-3000 words of authentic brand or founder writing, and an instruction to rewrite the draft so it reads as if written by the same author. No style guide, no adjective list, no “be punchy.” Just the source material and the task.

The output is noticeably closer to the target voice than anything you get from prompting Opus with a written style guide. Style guides describe voice. Examples are voice. Opus seems to treat the examples as a distribution to sample from rather than a rubric to satisfy, and that’s the right behavior for this job.

Why this breaks if you skip the cheap-model step

You can do the whole thing in Opus. It works. It also costs roughly 5x what the two-step version costs, and the quality gain on the ideation pass is minimal.

The reason to split is that ideation rewards breadth and tone-matching rewards precision. Different jobs, different models. The bill matters when you’re producing 40 pieces a week for a client, not four.

There’s a second reason. When Opus does both jobs in one shot, it tends to compromise on voice to preserve structure. When you hand it a finished structure and ask only for a voice pass, it commits harder to the style. The constraint is narrower, the output is sharper.

What doesn’t transfer from legal

The legal use case has one thing marketing doesn’t: a clear ground truth. A contract clause either covers the risk or it doesn’t. A reviewer can mark up the draft.

Brand voice has no ground truth. You’ll think the output sounds right, the founder will read it and say “that’s not how I’d say it,” and you’ll go back and forth. The fix is to treat the source examples as the contract. If you can’t point to three sentences in the reference material that justify a phrasing choice, the model is improvising and you should reroll.

The other gap: legal teams have stable voice targets. Senior partners write the same way for years. Brand voices drift quarterly, founders change how they sound after a podcast tour, and product pages get rewritten by whoever’s on deck. The reference corpus rots. Rebuild it every quarter or the model starts matching a voice that no longer exists.

The catch

If you try this on a brand whose existing copy is itself AI-generated, you’ll get a model matching a model. The output will be technically on-voice and completely hollow. The technique only works when the reference material was written by an actual human with consistent habits. Audit your reference corpus before you start, and if half of it came out of ChatGPT last year, you’re tone-matching a ghost. Find the original human writing, the Slack messages, the early blog posts, the founder’s old tweets, and build the corpus from that. That’s the work. The model is the easy part.