
[ founder voice specification ]
Your voice, written down precisely enough that a machine can execute it.
I turn the way you write into a specification, a set of rules, and worked examples that Claude, ChatGPT, Copilot, or a contractor can follow without you rewriting the result. The work runs two weeks at a price I quote before we start, and it needs three to four hours of your team's time.

[ the problem ]
You write better than your company does, but you can't write often enough to keep up.
You can sit down and produce something that sounds exactly like you, and it lands. Then three weeks pass before the next one, because a product issue and a hiring problem came first, and writing is always the first thing that slips.
​
So you hand it off to a contractor, a marketing hire or ChatGPT. What comes back is fluent, on topic and not yours. Fixing it takes longer than writing it would have, so you rewrite it yourself and end up owning the work again.
​
Nobody has written down what "sounds like you" means, so nobody can execute it, human or machine.
This is a specification problem
wearing a writing problem's clothes.
[ the objection ]
The models will keep improving,
and they still won't sound like you.
The fair objection is that the next model will fix this on its own, and I understand why people believe it. Fluency has improved fast and will keep improving, but fluency was never what was missing. A better model hands you more plausible options to choose from, and it still has no way of knowing which one you would have approved.
​
Words like direct, warm and no-nonsense work as shorthand between two people who already share context. A model shares context with no one, so it returns the most statistically likely version of the adjective, which is roughly the average of everyone who has ever used the word.
​
A specification works the way a newspaper's style guide works, because it lets an editor who has never met you make the same call you would have made, especially on the close calls.
[ what you get ]
Six documents that let someone else write like you.
Voice specification
This is your voice stated as rules about behavior instead of adjectives, including how you build sentences, the moves you always make and the ones you save for specific situations.
Banned-pattern list
This lists the words, openings and constructions you never use, so a model stops reaching for them by default.
Approved examples
These are drafts you would sign without edits, and each one is marked with the rule it shows.
Rejected examples
These are fluent drafts you would send back, each with a plain explanation of why it missed.
Scoring rubic
The rubric runs named checks that return a number, so you can judge a draft before anyone reads it closely. Right now you're the only person who can tell whether something is right, and this is the piece that changes that. Truth be told, it will still flag the occasional draft that is fine, which is why week two tests it against your real drafts.
Configured workspace
I load the specification into the tool you already use, whether that's Claude, ChatGPT, Copilot or a brief for a contractor. It stays a portable document on purpose, so switching tools later doesn't cost you the work.
[ how it runs ]
Two weeks from the first conversation to handoff, with no discovery phase.
1
Extraction session
week one
2
week two
3
4
Drafted specification
week three
Revision session
week four
Handoff
I interview you and read at least twenty pieces you have already published. That session doubles as the diagnostic, so the work starts on day one.
I write the rules, the banned patterns, both example sets and the rubric from what you said and what you published.
We run the specification against real drafts together and tighten every rule that missed.
You receive all six documents and a configured workspace, and the engagement is finished without a retainer attached.
Your time across the two weeks comes to three to four hours in total, most of it in the two working sessions.
[ pricing ]
One fixed price, quoted before anything starts.
[
$4,500
]
Voice specification
The work takes two weeks, with half paid at signature and half at handoff.
Other options
engagement
Multi-writer variant
fee
This is for teams with three or more people writing, and it adds rules for holding every writer to the voice, plus a guide to who reviews what.
$7,500
Quarterly refresh
This is optional, and each I re-score your recent drafts and update any rule that has drifted.
$750 / quarter
Founding clients
I'm pricing my first three engagement below this rate, in exchange for using the results and your name as a case study. If that trade works for you, say so when you email.
[ fit ]
This works best when three things are already true.
You have published at least twenty pieces of content
Delegated writing to AI has already disappointed you
You can give me three to four hours
Below twenty there isn't enough to pull rules from, so the work turns into invention, which is a different and worse product.
If you have handed writing off and rewritten what came back, you have felt the exact problem this solves.
That time is spread across two weeks, mostly in the extraction session and the revision session.
[ scope ]
I build the specification, and I don't write your content.
The point of the specification is that you stop needing me, or anyone with my particular judgment, to produce work that sounds like you.
​
I also don't take on social media management, full brand strategy or general AI training, because this practice is narrow on purpose. To be frank, I don't have a software engineering background either, so I won't maintain running code inside your systems; the work is specification and configuration, and it's built to run without me.
[ background ]
I learned this work by writing for a Fortune 500 chief executive.
I drafted speeches, talking points and social media posts for a Fortune 500 chief executive, which meant learning one person’s voice well enough that my drafts came back with few edits. The demand for those words always outran the hours we had to write them, so I shifted from drafting each piece myself to building an executive voice agent that could produce the first drafts.
​
That shift taught me two things I now bring to every engagement. I learned how to spell out exactly what a model needs to be fed, and how to build a structure inside it that cuts down review time and gets writers to a usable first draft faster. It works like a good brief for a new speechwriter, because the clearer the brief, the less the editor has to rewrite.
​
Before that, I worked in communications strategy at Padilla, ICF Next and Bellmont, and in the Washington State Attorney General's office. Getting an organization to sound like itself was a discipline problem long before it was a technology problem, and while the systems are new, the standard is not.
​
Sam Zender
Founder & Principal, Carrawright

[ writing ]
I write about the mechanics of brand voice on Substack.
[ for larger teams ]
Marketing and communications teams need a different engagement.
If your company already has a communications function and AI drafting is happening across it, the problem is drift rather than volume. Vendor model updates shift the voice overnight, and shared prompts get quietly edited until nobody is working from the same source.
​
For those teams I build voice systems inside Copilot, Gemini or a Claude tenant, with review tiers, a voice fidelity rubric and a re-test after every model update. Those engagements start with a fixed-fee diagnostic, and I'll send the full brief if you ask for it.