Hey, Mark here.

I lost almost an hour last Tuesday before I'd written a single word. The hour went to ramping back into a client account I hadn't touched in eleven days. I re-read the brief, scrolled through old emails, and got their voice back in my head before I could start.

Multiply that by the number of clients you're juggling and the number of times you switch between them in a week. That's the cost most working copywriters pay without noticing.

Another previous issue argued that voice match against the client's brand is part of what keeps premium work premium. This is the system that makes voice match repeatable, and that took my ramp-in cost to near zero.

Here's the build.

You're juggling five clients. You sit down to write for Client B. You haven't touched their account in eleven days.

You re-read the brief. There's an email thread from March that has the founder using a phrase you want to remember the rhythm of, so you scroll back through that. Then you pull up the landing page you wrote in February and re-read your own old copy to find the cadence. An hour goes by. You haven't typed anything for the new project.

Multiply that by five clients. Multiply that by every context switch in a week.

That's the cost. Most working copywriters pay it without measuring it, because they've never seen what their week would look like without it.

Claude can hold a client's voice as well as you can, and it can hold five voices better than you can. It just needs to be set up properly, and almost nobody sets it up properly.

This is that setup.

You build a Claude Project for each client. The project holds a voice profile you extracted from their samples, plus the samples themselves, plus a verification protocol you run any draft through before sending. Once it's built, you never re-ramp. You open the project, you write, and the voice is already loaded.

The system has three parts: voice DNA extraction, project encoding, and verification protocol. I'll walk through each.

The first thing most copywriters get wrong is briefing Claude on voice with adjectives. "Warm, professional, slightly playful." That tells the model nothing useful. It pattern-matches against generic training data and gives you back generic copy.

Voice doesn't live in adjectives. It lives in specific patterns: sentence length distribution, vocabulary register, what they say repeatedly, what they specifically don't say, how they address the reader. You can't write those down off the top of your head, so you have to pull them from existing samples.

You need three to five samples. Mix the mediums — a landing page, a few emails, maybe a founder's personal post if you can get one. Range matters more than volume. Three samples spanning landing/email/social will give you a sharper profile than ten samples that are all product pages.

Here's the extraction prompt I use. It's long because it has to be. Short voice prompts get back vague nonsense. This one forces specifics.

Here's what this prompt is doing.

Every claim has to point back to a specific quote from the samples. That removes the model's ability to default to generic descriptors and forces it to find the patterns in the actual writing.

Half of voice is what a brand never says, and most extraction prompts skip the negative space entirely. The list of words and phrases the brand avoids becomes the rule set that prevents drift later.

And the final paragraph the model produces goes straight into your project instructions in Step 2. That's voice guidance built from evidence rather than assembled from adjectives.

A few patterns keep showing up when I watch working copywriters do this step.

One sample is too little. The model can't see patterns from a single piece. It pattern-matches against its training and hands you back something generic.

Volume alone won't fix it. Ten product page sections will produce a worse profile than three samples spanning different mediums. Range across formats is what surfaces the voice underneath.

The "what they don't do" section is the one most people skip. Most voice errors come from saying things the brand wouldn't say, and without that list of forbidden phrases, you can't enforce the negative space later.

If the profile that comes back feels generic, re-run with better samples. The fix is almost always upstream.

Now you're going to drop the voice profile and the samples into a Claude Project. If you've never built one, this is the moment.

A Claude Project is a persistent workspace with its own custom instructions and its own attached files. Every conversation you start inside the project picks up the same context, so you don't re-paste the voice profile and you don't re-ramp.

For each client, build one project. Name it something obvious like the client's name plus what it's for, such as "Acme Co — Copywriting." Don't get clever with the naming. You'll be opening these fast on busy days.

Inside the project, two things matter: custom instructions and attached files.

Custom instructions go in the project's instructions field. Here's what mine look like:

Attached files are where the samples themselves go. Paste two or three originals in as separate files or as one combined reference document. The model needs the raw material to compare against. The instructions tell it what to do; the samples show it what the result should sound like.

The "flag and ask" line in the rules took the longest to figure out. Without it, the model silently drifts when you give it a brief that doesn't fit the voice. With it, you get a check at the brief stage instead of a bad draft you have to rewrite.

I should add a note on tools. I run this in Claude Projects because that's where I work. The same setup runs in ChatGPT's Custom GPTs, in the Claude API with a system prompt, or in any environment with persistent context and attached files. The principle is the same. Find the place where you can write custom instructions and attach reference material, then use it.

Voice drift is invisible until someone points it out. You'll write inside the project, the model will match the voice for the first six paragraphs, and then somewhere around paragraph seven it'll slip back into default Claude tone. You won't catch it because you've been reading your own work for forty minutes and your ear is tired.

The verification protocol is a separate pass with a different framing. You start a new conversation rather than continuing the writing one, and the model's job here is to grade the draft against the references.

Here's the prompt:

The structure of this prompt is the whole game.

"Read the references first" calibrates the model before judgment. Without that line, the model often scores the draft against its own internal sense of "good copy" instead of against the brand. With it, the standard becomes the brand's existing work.

The five axes prevent vague feedback. Without the axes, you get back "the draft mostly matches the voice with some minor inconsistencies." That's useless. With the axes, you get a specific score on each dimension and you know where to fix.

"Quote three specific phrases" forces actionable output. The model has to point at the exact words that don't fit and suggest a replacement. You can use those suggestions directly or write better ones — either way, the problem is no longer abstract.

The final verdict is a binary: publish the draft as the brand, or send it back. No hedging is allowed. That forces a decision instead of a vague "this looks pretty good."

You should expect the first verification pass on any draft to flag two or three real issues. That's normal. Iterate on those, run it again, see what's left. Two passes is usually enough. Three passes means the draft has a structural problem rather than a voice problem, and you need to look at the brief instead.

I keep the verification prompt as a saved snippet rather than inside the client project itself. The judgment needs to come from a fresh evaluation rather than from a continuation of the writing conversation. A new context gives you clear eyes on the draft when you need them most.

It's worth being honest about the limits of this system before you build it.

The system doesn't work for brands without a consistent existing voice. If the client has three founders who all write differently and the existing samples are stylistically scattered, you can't extract DNA that isn't there. You'll have to do the harder work of building the voice first. That's a different project, and the model won't help you skip it.

Strategic judgment still belongs to you. The model can match a voice. It can't tell you what the client should be saying. The brief, the angle, the offer, and the structural moves on the page are all still your work. The system buys you time on voice, but it doesn't give you back judgment.

Profiles go stale. Brands evolve, founders change, the company hires a new content lead and the tone moves a few degrees. I refresh mine for active clients every few months, which takes twenty minutes or so per client. The refresh is faster than the original extraction because most of the work is checking what changed rather than building from scratch.

Here's what the system does well.

Ramp-in time goes near zero. You open the project, you have the brief, you start writing. The voice is already loaded. The hour I used to lose on Tuesday mornings is gone.

Output quality stops being a function of your day. Variance between your best days and your worst days shrinks, because the voice isn't something you're holding in your head while you're also holding the strategy.

And the drafts get harder to spot as LLM-assisted. The Floor Rose issue pointed at trust premiums going up for human-written work. The harder it is to tell, the more of that premium you keep.

For a client I see weekly, I built the project in about an hour and have used it for thousands of words of output since. The setup compounds. Every session pays back the time you spent up front.

Voice match used to be a setup cost. Build the project once, and it stops being one.

The hour you'd spend re-ramping is the hour you can spend on the part of the work the system can't do.

Try it on one client this week. Pick the one you switch into most often.

More clicks, cash, and clients,
Mark Masters