Good morning.
An agency owner forwarded me a LinkedIn post last month and asked why it wasn't getting him traffic. It had a clean hook, tidy structure, and a real point underneath it. The post also read like it could have come from any of four hundred other people in his category.
He'd written it with AI in about six minutes. That was the point, and it was also the problem. He'd done what everyone tells you to do: pasted a few of his old posts into the chat, told the model to study his voice, and asked it to write the next one. What came back was competent and anonymous. The model had averaged him into the middle of his market.
The reason is simple once you see it. A handful of examples in a prompt window gives the model a vibe to chase, and a vibe regresses to the mean the moment the model starts generating.
Today's issue is the build that fixes it. The Imprint and Generate architecture, the platform decision, full prompts, and worked examples for the four business types I see most often on this list.
— Sam
IN TODAY’S ISSUE 🤖
Why build the Ghostwriter Agent
The two platform paths, ranked
Why "write like me" fails every time
The Imprint: the three layers of a voice profile
Seven build steps, every prompt inside
The Ghostwriter by business type, side by side
Hand it to your team without flattening your voice
Let’s get into it.
Why Build The Ghostwriter Agent
The Ghostwriter Agent is the build that lets you publish in your own voice at the volume your business needs without writing every word yourself.
Most owner-operators are the voice of their business whether they planned to be or not. The newsletter sounds like you. So do the sales emails that close and the LinkedIn posts that bring in inbound, and all of it routes through the one person who cannot be cloned. That caps the output at whatever you can personally write between everything else you do.
So operators reach for AI, and they hit one of two walls:
The first wall is the anonymous draft: technically fine, recognizably no one.
The second wall is the editing tax, where the draft needs so much rewriting to sound like you that you'd have been faster writing it yourself.
The Ghostwriter Agent clears both walls by doing the one step almost everyone skips. It builds a structured profile of how you write, holds that profile on disk, and references it on every generation. The output starts much closer to your voice, and the edit becomes a pass instead of a rewrite.
Three reasons to build it:
It removes you as the bottleneck on published volume without removing your voice from it. The constraint on most operators' content isn't ideas. It's the hours between having the idea and having it written the way only they would write it.
It compounds every time you publish. Each piece you approve becomes calibration material for the next one. The profile sharpens as you use it, so the agent that wrote your posts in March writes them better in June.
It captures something you can't hand off any other way. A new hire can learn your process in a month. Learning your voice takes a year, if it happens at all. The voice profile makes the tacit explicit, which is the only form your voice can travel in.
This is where the Digital Employee frame applies. An agent that has studied your writing, holds a profile of your voice, and drafts against it on demand is a staff writer on your bench who already knows how you sound.
You manage it the way you'd manage a strong ghostwriter you hired:
A clear brief on how you write (the voice profile)
A body of your real work to learn from (the corpus)
A review on what it produced (you read the draft, you mark what's off, you feed the correction back)
A widening remit over time (more formats, more channels, more of your publishing load)
That mental model keeps you honest about what the agent is for. It drafts. You still decide what's worth saying and you still own the final read. The Ghostwriter writes the version you would have written if you'd had the afternoon. You're the one who didn't have the afternoon.
The Platform Decision (Read It Once, Decide, Move On)
Two serious paths for building the Ghostwriter Agent. Pick one. Run it for a month before considering the other.
Path | Best for | What you get | What to know |
|---|---|---|---|
Claude Code in Cowork mode | Operators already in Claude | Folder-based voice profile you can read and edit; a corpus folder the agent reads on every run; drafts written straight to disk | The path I run on. Lowest setup if you already have a Claude workflow. |
The Codex app | Operators in Microsoft 365, Slack, or Notion | Persistent memory of your voice profile across sessions, plugins for the apps your drafts and notes live in | A real candidate as of the recent release. The profile persists between sessions, so you calibrate once and generate for months. |
Quick Decision Rule
Already in Claude? Use Cowork mode.
Already in Microsoft 365, Slack, or Notion? Use the Codex app.
Why "Write Like Me" Fails
The instruction almost everyone gives the model is some version of "here are three of my posts, now write like me." It produces a draft that's close enough to feel like progress and wrong enough to need a rewrite. It’s worth understanding why, because the failure points straight at the fix.
A few examples in a prompt window give the model a target it can only approximate. It pattern-matches the surface (you write short paragraphs, you use a dash, you ask a question near the top) and fills everything underneath with its own defaults. Its defaults are the average of everything it was trained on, which is the most generic prose in existence.
The closer your real voice sits to that average, the more the output passes. The further your voice sits from it, which is the whole reason your writing works, the more the model sands you down toward the middle.
The fix is to stop asking the model to infer your voice on the fly and instead calibrate it once, deliberately, into a profile you can read, correct, and reuse. That's the architecture.
The Imprint and Generate Architecture
The Ghostwriter Agent runs in two modes, and the entire build hinges on getting the first one right.
The Imprint is the calibration. The agent reads a real corpus of your best writing and produces a structured voice profile: a document that describes how you write across three layers.
A vague imprint produces a vague ghostwriter. A precise imprint produces drafts you can ship with a light edit. Calibration quality is the whole game.
The Generate is the writing. With the profile built, the agent drafts any new piece against it, layer by layer, checking its own output against the profile before it hands you anything.
The reason to separate them is that calibration is hard and generation is easy, and bundling them hides the hard part. When you do the imprint as its own deliberate step, you can inspect it, argue with it, and fix the places where the agent misread you. You can't inspect a vibe.
The Three Layers Of A Voice Profile
The imprint describes your writing across three layers, from the surface down to the part that's truly you.
Layer | What it captures | How the agent derives it | Why it matters |
|---|---|---|---|
Mechanics | Sentence length and rhythm, paragraph shape, punctuation habits, how you open and close, formatting tics | Measured from the corpus: average and range of sentence length, paragraph length distribution, frequency of questions, lists, one-line paragraphs | This is the layer the model gets roughly right on its own. The profile pins it down so it stops drifting back to default cadence mid-draft. |
Lexicon | The words you reach for, the words you never use, your signature phrases, the register you write in | Pulled from the corpus as a kept list and a banned list, plus the recurring constructions that are yours | This is where contamination happens. The model leaks its own favorite words in. An explicit banned list is the only thing that holds the line. |
Stance | How you argue, your relationship to the reader, the moves you make (analogy, confession, the contrarian turn), what you believe | Inferred from the corpus and then confirmed by you, because this is the layer the agent is least able to read on its own | This is the layer that's truly you. It's also the one the agent gets least right unsupervised, which is why the profile names it explicitly and why you edit it by hand. |
The Mechanics are measurable, the Lexicon is listable. The Stance has to be drawn out and then checked by the one person who knows it, which is the work of Step 5 below.
What Generation Does With The Profile
When the agent drafts, it doesn't write and hope. It writes against the profile and checks itself:
It drafts the piece holding the Stance layer as the brief: the move you'd make, the position you'd take, the way you'd come at the reader.
It shapes the draft to the Mechanics: your sentence rhythm, your paragraph length, your way in and your way out.
It runs the Lexicon as a filter: it strips its own default vocabulary and checks the banned list before it hands you the draft.
You read the result, mark what's off, and feed the correction back into the profile. The imprint sharpens. The next draft starts closer.
The Protocol
Seven steps. Every prompt below is usable as written.
Step 1: Pick The Platform And Confirm Access
Pick Claude Code in Cowork mode or the Codex app based on the decision section above. Confirm the agent has access to:
A folder or workspace where the voice profile and drafts will live
A corpus folder where you'll put samples of your writing
Web browsing, if you want it to reference your published work directly
You don't need anything exotic. A folder with a dozen of your best pieces in it and a place to write drafts is the whole setup.
Step 2: Assemble The Corpus
This step sets the ceiling on everything downstream. The agent can only imprint on what you give it, so give it your best and your most you.
Gather twelve to twenty samples of your writing into the corpus folder. The rules that matter:
Pick your strongest work, not your most recent. The corpus is the target the agent aims at. Fill it with the pieces that sound the most like you at your best, even if some are two years old.
Match the corpus to the job. If you're building this to write LinkedIn posts and sales emails, the corpus should be LinkedIn posts and sales emails. Don't calibrate on your newsletter and expect it to nail your cold outreach. They're different registers.
Use only work you wrote yourself. Anything ghostwritten by someone else, or already AI-assisted, poisons the imprint. The agent will faithfully learn a voice that isn't yours.
Strip the noise. Pull out the boilerplate, the legal footers, the signatures, the quoted material. You want your prose, not the scaffolding around it.
Name the folder voice-corpus/ and drop the files in as plain text or markdown. Twelve strong samples beat thirty mixed ones.
Step 3: Install The System Prompt
This is the agent's job description, and it governs both modes. Put it in a file called ghostwriter-system-prompt.md in the same folder.
You are the Ghostwriter Agent for [Your Name]. You write in their voice, and you do not write in any other voice, including your own defaults.
You operate in two modes: CALIBRATE and GENERATE.
Always read voice-profile.md before generating, if it exists. Everything in that file takes precedence over your defaults. When the profile and your instinct disagree, the profile wins.
CALIBRATE MODE
When asked to calibrate, read every file in voice-corpus/ and produce voice-profile.md describing how this person writes across three layers. Work from the corpus, not from assumptions. Where the corpus does not give you enough to judge, say so rather than guessing.
THE MECHANICS. Measure and report: average sentence length and the range; paragraph length distribution; how often single-sentence paragraphs appear; frequency of questions, lists, and dashes; the typical way pieces open and the typical way they close. Give numbers where you have them.
THE LEXICON. Produce two lists. KEPT: the words, phrases, and constructions that recur across the corpus and read as this person's own. BANNED: words this person never uses that you, as a language model, would be tempted to insert. Infer the banned list from the gap between the corpus and your own defaults.
THE STANCE. Describe how this person argues, the relationship they take to the reader, the recurring moves they make (analogy, anecdote, confession, contrarian turn, direct instruction), and the beliefs or positions that show up across pieces. Flag this layer as DRAFT, because the human will correct it.
Output voice-profile.md in the template below. Do not generate any new writing in this mode.
GENERATE MODE
When asked to write something, draft it against voice-profile.md in this order:
1. Hold the STANCE as the brief: what move would this person make, what position would they take, how would they come at the reader on this topic.
2. Shape the draft to the MECHANICS: match the sentence rhythm, paragraph length, opening and closing habits in the profile.
3. Run the LEXICON as a final filter: remove your own default vocabulary, check every word against the BANNED list, and confirm the piece could plausibly sit in the corpus.
Before you hand over a draft, state in one line which parts you are least confident match the voice, so the human knows where to look. Never claim the draft is finished. It is a draft.
VOICE PROFILE TEMPLATE:
# Voice Profile: [Your Name]
Calibrated from [N] samples on [date]. Job: [what this profile is for, e.g. LinkedIn posts and sales emails].
## The Mechanics
[Measured cadence, paragraph shape, punctuation habits, opening and closing patterns, with numbers.]
## The Lexicon
### Kept
[Recurring words, phrases, and constructions that are this person's own.]
### Banned
[Words and phrases this person never uses that the model is tempted to insert.]
## The Stance (DRAFT, human to confirm)
[How they argue, their relationship to the reader, the recurring moves, the positions that recur.]
## Known Gaps
[Anything the corpus did not give enough signal to judge.]
A couple of things this prompt is doing under the hood:
It forbids the agent's own voice in plain language. The model's default is to help by smoothing prose toward the average. The line "you do not write in any other voice, including your own defaults" is doing real work, and it's why the banned list exists as its own layer.
It makes the agent report its low-confidence spots before you read the draft. That one line turns your review from a hunt into a check. You go straight to the places the agent already flagged.
Step 4: Run The Imprint
With the corpus assembled and the system prompt installed, run calibration:
Run CALIBRATE mode. Read every file in voice-corpus/ and produce voice-profile.md using the template in your system prompt.
Work only from the corpus. Where you're inferring rather than measuring, say so. Mark the Stance layer as DRAFT. Do not write any new content in this pass.The agent reads the corpus and writes the profile. The Mechanics will be close, because that's the measurable layer. The Lexicon will be mostly right with a few misses. The Stance will be a rough first read that needs your hand. That's expected, and it's the whole reason the next step exists.
Step 5: Inspect And Correct The Imprint
This is the step that separates a Ghostwriter Agent that works from one you abandon after a week. Open voice-profile.md and read it as if someone described your writing to you. Three passes:
Read the Mechanics for accuracy. These are measured, so they're usually right. If the numbers feel off, your corpus is probably mixed. Go back and tighten it.
Read the Lexicon for theft and gaps. Check the Kept list for anything that isn't really yours, and the Banned list for the model's tells it missed. Add the words you know you never use. This is your highest-leverage edit, because the Lexicon is where the agent's voice leaks back in.
Rewrite the Stance in your own words. The agent's draft of this layer is a guess. This is the layer that's truly you, so it's the layer you should write. Describe how you argue, what you believe, the moves you make, in language that sounds like you describing yourself. Five sharp sentences you wrote beat two paragraphs the agent inferred.
The profile is a living document. You'll keep editing it for the first month and lightly after that. The half hour you spend here is worth more than any prompt engineering you'll do later.
Step 6: Run Your First Generation
Pick something real you need to publish this week. Send it:
Run GENERATE mode. Draft a [LinkedIn post / sales email / newsletter section] about [topic].
Context: [the point you want to make, any specifics that have to be in it].
Draft it against voice-profile.md. Hold the Stance as the brief, shape it to the Mechanics, run the Lexicon as a final filter. Before the draft, give me one line on where you're least confident it matches my voice.The first draft will be 70 to 80% there. Read it against your own ear:
Layer | What to check |
|---|---|
Mechanics | Does it move like you? Right sentence rhythm, right paragraph length, your kind of opening |
Lexicon | Any word in here you'd never use? Any model tell that slipped the banned list? |
Stance | Is the underlying move yours, or did it take the safe, average angle? |
Where it's off, don't just fix the draft. Note why it's off and feed it back into the profile, so the next draft starts from the corrected version. The first ten drafts are as much calibration as they are output.
Step 7: Tune After Three Drafts
Three real drafts in, the failure patterns show themselves. Update the profile to close each one. The four you'll almost certainly hit:
Corporate creep. The drafts get more buttoned-up than you are, reaching for "leverage," "streamline," "robust." Add to the Banned list and add a Stance note: "Register is a sharp practitioner talking to a peer, not a brand talking to a market. When in doubt, less polished."
Vocabulary contamination. The model's signature words keep appearing: "delve," "navigate," "underscore," "tapestry," the dash-heavy cadence. Every time you catch one, it goes on the Banned list. This list never stops growing, and that's correct.
Length bloat. The drafts run longer than you write, because the model pads to feel thorough. Add a Mechanics constraint with your real numbers: "Target [X] words. Paragraphs run [Y] sentences. When a sentence can be cut without losing the point, cut it."
Stance flattening. The hardest one. The draft is mechanically perfect and says nothing you wouldn't have softened. The move is too safe, the angle too agreeable. Sharpen the Stance layer with the specific positions you hold and the moves you make, and tell the agent: "Take the position the corpus takes. Do not hedge a view this person states plainly."
After three rounds of this, the profile is yours and the drafts arrive most of the way home.
What This Looks Like By Business Type
The Ghostwriter Agent stays the same. The corpus you feed it and the work you point it at change with the business. Here's the build by business type, side by side:
Agency | SaaS | Ecommerce | Expert / Creator | |
|---|---|---|---|---|
Corpus to gather | Founder's best LinkedIn posts, sales emails, proposal intros, client-facing notes | Founder's product announcements, changelog narratives, thought-leadership posts, investor updates | Founder's brand emails, product-launch copy, the lines customers quote back, social captions | Newsletter issues, best posts, sales-page sections, talk transcripts |
What it writes | Inbound posts, outbound sequences, proposal openers in the founder's voice | Launch posts, lifecycle email copy, positioning narrative that doesn't sound like every other SaaS | Campaign emails and product copy that hold the brand voice at volume | Newsletters, posts, and sales copy at publishing cadence without the founder writing every word |
Voice that anchors the profile | The founder's personal voice (the agency is the founder) | Often a house voice the team writes in, not one person | The brand voice, which may be a persona the founder built | The founder's personal voice, the entire asset |
Watch-Out | Sales register vs. content register. Calibrate them separately or the outbound sounds like a blog post. | House voice drift. Without a profile, every writer reinvents the voice and it splinters. | Persona consistency. The brand voice has to survive being written by an agent and three team members. | Stance flattening. The expert's whole value is the take, and that's the layer the agent softens first. |
The pattern across all four: the imprint is only as good as the corpus, and the Stance is the layer that needs you. Get those two right and the agent writes in your voice at a volume you can't hit by hand.
A Note On The House Voice
Two of these builds (most SaaS teams, many ecommerce brands) calibrate a shared voice rather than one person's. The architecture doesn't change but the corpus does. Instead of one person's best work, you assemble the pieces that best represent the brand at its sharpest, from whoever wrote them, and the profile becomes a shared asset the whole content team generates against.
That's the version that keeps a brand voice from splintering the moment you add a second writer, and it's covered in the Bonus below.
Deploying This Across Your Team
The Ghostwriter Agent works as a solo build. It changes your whole content operation when the team writes against the same profile you do.
Four pieces of ownership keep it clean:
You own the profile. The voice profile is the asset. It lives in a shared folder the team can read but not edit, so every draft anyone generates aims at the same target. The whole team writes toward one defined voice instead of five private reads on it.
The team generates, you calibrate. Your writers run GENERATE mode against the profile for first drafts. You're the only one who runs CALIBRATE and the only one who edits the profile, because the profile is your judgment about the voice and that judgment shouldn't fork.
Corrections route back to one place. When a draft misses, the writer notes why and the correction goes to you, not into a private prompt tweak. The profile improves once, for everyone, instead of drifting into five local versions.
You keep the final read on what publishes in your name. The agent drafts and the team edits, but anything going out under your name gets your eye. The Ghostwriter buys back the writing hours. It doesn't buy back the judgment about what's worth saying.
The compound effect: your published voice stops being capped at your personal writing speed and stops splintering across whoever happens to be drafting. The newsletter, the posts, and the emails all sound like the business at its sharpest, and the profile that makes that happen gets better every time someone uses it.
Bonus: The House Voice Variant
Some operators on this list are the entire voice of the business. Others run a brand voice a team writes in, or want to. The Imprint and Generate architecture works for both. The build changes in two places.
#1: The corpus changes from one person's best work to the brand's best work, regardless of who wrote each piece. Pull the emails, posts, and pages that sound the most like the brand at full strength, and leave out the ones that were already drifting. The agent imprints on the brand, not on a person.
#2: The Stance layer changes from "how this person argues" to "what this brand believes and how it talks to its market." This is the layer to write by hand, the same as in the personal build, because a brand's stance is a decision, not a measurement. Name the positions the brand takes, the things it refuses to say, and the relationship it keeps with the reader.
Run it and you get a voice profile that's a shared asset rather than a personal one. A new writer on the team generates against it on day one and sounds like the brand instead of spending six months learning it by osmosis. The profile becomes the thing that holds a voice steady as the team that writes it changes.
The discipline is identical to the personal build: the corpus decides the ceiling, and the Stance is the layer that needs a human. A brand voice splinters the same way a personal one flattens, and the profile is what keeps it whole.
Want The Full Agent Build Pack?
You can build the Ghostwriter Agent from the prompts above. If you'd rather skip the setup work, Cortex subscribers get the full Ghostwriter Agent Build Pack as a download:
The installable Claude Skill that sets up the agent in Cowork mode in one command, calibration step included
The Codex configuration pack with system prompt, plugin list, and persistent-profile setup
Four pre-filled voice profiles (agency, SaaS, ecommerce, and expert / creator), each with a worked Mechanics, Lexicon, and Stance you can read as a model before you build your own
The Tune Pack: the most common drift patterns I've seen for this agent (corporate creep, vocabulary contamination, length bloat, stance flattening), with the exact profile edits that fix each one
Three worked drafts with the operator's edits visible, so you can see what a light edit looks like versus a rewrite
The Channel Variant Pack: an advanced extension that adapts a single calibration to LinkedIn, X, newsletter, and client-comm tones, so one voice profile drives every channel without recalibrating from scratch
This is only available to paying Cortex subscribers.
What is Cortex?
Cortex is the premium monthly membership version of Bionic Business, helping entrepreneurs and business owners grow and scale with AI and Agents.
In addition to 3-4 full, unrestricted regular issues per month, you also get:
1x SIGNALS Strategic Briefing issue (strategic moves you must make now to secure your business revenue, market share, and profits).
1x CIRCUITS Tactical Guide issue (workflows, how-to’s, prompts, very practical implementation inside a self-driving business).
Access to all Skills issues (full details, download files, tutorials).
Claude Cowork as your marketing team special issue (full guide on how to run Cowork as an agent marketing team for your business).
50% OFF workshops on prompt engineering, agent operating systems, and more. Only active Cortex subscribers get any kind of discount, no one else.
So, if you’d like to receive full issues and the monthly special Signals and Circuits issues, then you should sign up for Cortex, which is $50/month (and you have full archives of previous regular issues for as long as you’re a paying member).
One note on Signals and Circuits: You only get these special issues for the months you’re subscribed—which means if you skip a ~30 day period, about a month, you won’t get them for that month. There are no back-issues for past months, so now is your chance to lock-in these upcoming issues while you can.
After you’ve upgraded to Cortex, log in to the website and come back to this issue, and the Agent Build Pack will be unlocked below.
Ready? Get the Agent Build Pack, full regular issues, and the upcoming Signals and Circuits issues:
Cortex is a monthly premium version of Bionic Business, helping entrepreneurs grow and scale with AI and Agents. You can cancel at any time. But there are no refunds, for any reason. All sales are final.
Build the Ghostwriter Agent and the thing you stop doing is staring at a blank draft trying to sound like yourself on a day you don't feel like it.
The voice profile holds your voice on disk, the agent drafts against it, and you go from blank page to a light edit instead of from blank page to a finished piece.
The operators who win the next two years on attention are the ones who turn their voice into an asset their business can run on at volume, instead of a bottleneck only they can clear.
Your voice is the one thing AI should not average into the middle, as long as you do the calibration that keeps it yours.
Build it this week, and feed it your best.
Talk soon,
Sam Woods
The Editor
.
