Voice DNA and Content

This is the web edition of Chapter 09 from the GTM Coding Agent Playbook. Expanded with deeper process detail, real before/after examples, and the compound voice effect.

---

A friend of mine runs a SaaS company. Good product, growing revenue, real customers. Last year he decided to "scale content" by giving Claude his blog topics and letting it write. He published twelve posts in a month. Traffic went up. But something weird happened to his pipeline.

Inbound leads dropped.

People were visiting the blog, reading a paragraph or two, and leaving. The content was technically correct. It covered the right topics. It hit the right keywords. But it didn't sound like him. It sounded like everyone else's AI-generated content. The same cadence. The same word choices. The same rhythm of setup, explanation, conclusion. His prospects could smell it, even if they couldn't name exactly what was wrong.

He came to me frustrated. "AI content doesn't work for B2B."

No. AI content without voice context doesn't work. AI content with your actual voice baked in works better than anything you could produce manually, because it's you at scale, writing at a pace no human can sustain.

The difference is Voice DNA.

The Slop Problem

You've seen it. We've all seen it. LinkedIn posts that start with "In today's rapidly evolving landscape..." Blog posts packed with "leverage," "synergy," and "game-changer." Emails that read like they were assembled from a corporate phrase book by someone who has never had a real conversation.

This is slop. And slop is the new spam.

Here's why it matters for GTM specifically. Your buyers read 50 to 100 emails a week. They scroll past hundreds of LinkedIn posts. Their filter for "this was written by a human who actually thinks about this stuff" versus "this was generated by AI and blasted to 10,000 people" is razor-sharp and getting sharper by the month.

The irony is that AI-generated content isn't bad because it's AI-generated. It's bad because most people use AI with zero context about how they actually write. So the output defaults to generic corporate voice. The statistical median of all the professional writing the model has ever seen. It's competent, inoffensive, and utterly forgettable.

You don't sound like the median. Nobody does. That's exactly why your voice matters. Voice DNA is how you extract what makes your writing yours and inject it into every piece of content your agents create.

What Voice DNA Actually Is

Voice DNA is a structured profile that captures how you write. Not how you want to write. Not how you think a "thought leader" should sound. How you actually sound when you're writing well and not thinking about it.

It includes:

Identity. Who you are, what you do, what you've earned the right to have opinions about. A first-time founder writing about product-market fit reads differently than a three-time founder writing about the same thing. The identity section gives the agent the context to match your level of authority without overstating it.

Tone markers. Are you dry? Blunt? Self-deprecating? Technical and precise? Warm and conversational? Most people have five to seven tone descriptors that capture their writing personality. The combination is what makes you sound like you and not like someone else who shares two of those traits.

Vocabulary patterns. The words you reach for naturally. The phrases that recur across your writing. The technical terms you use without explaining because your audience knows them. Also, critically, the words you never use. If you never write "leverage" in real life, seeing it in AI-generated content immediately breaks the illusion.

Sentence structure. Short punchy sentences? Long flowing ones? A mix with a particular rhythm? How long are your paragraphs? Do you use lists frequently or rarely? Do you ask your reader questions? These structural patterns are some of the most distinctive fingerprints in any writer's voice.

Anti-patterns. This is the kill list. Words, phrases, and structural patterns that are banned from any content generated in your workspace. This is arguably the highest-value piece of the entire Voice DNA system, and we'll spend a lot of time on it.

Platform rules. How your voice adapts across LinkedIn, email, blog, and other channels. The voice stays the same. The format and intensity change.

When this profile lives in your workspace as structured markdown files, a coding agent can reference it every time it writes content for you. The output goes from "generic AI" to "sounds like me on a good day."

How to Extract Your Voice

This isn't theoretical. This is a practical process you can finish in about 30 minutes.

Step 1: Gather your samples

Find 5 to 10 pieces of your best writing. Not your most polished. Your most you. The distinction matters. Polished writing is often edited toward a norm. You want the raw stuff where your personality comes through.

Places to look:

• LinkedIn posts that got real engagement. Not just impressions. Actual comments from people who resonated with the content.
• Emails you sent to prospects or customers that got genuine replies. Not "thanks for sending this" replies. "This really clicked, let me introduce you to our CTO" replies.
• Blog posts or newsletter issues where you wrote from experience rather than researching and summarizing.
• Slack messages or internal docs where you explained something complex in your own words. These are gold because you weren't performing for an audience. You were just communicating.
• Twitter/X threads if you write there. Thread voice tends to be the most distilled version of how someone thinks.

Copy the raw text. Don't edit it. Don't fix the typos. Don't clean up the casual tone. The agent needs to see your natural voice, warts and all, not a sanitized version.

Step 2: Feed them to your coding agent

Open Claude Code in your workspace and give it the samples along with a prompt like this:

I'm going to paste 5-10 writing samples. These are my actual voice.
LinkedIn posts, emails, and blog content I've written.

Analyze these samples and create a voice DNA profile with:

1. Identity: who I am based on what I write about and how I write
2. Tone markers: 5-7 adjectives that describe my writing voice
3. Vocabulary patterns: words and phrases I use often, natural to my style
4. Sentence structure: short/long/mixed, paragraph length, use of lists
5. What I never say: words, phrases, and patterns absent from my writing
6. Platform differences: any differences you notice between channels

Be specific. Don't give me generic advice. Ground everything in the
actual samples. Cite specific phrases and patterns from the text.

Then paste your samples.

Step 3: Review and edit ruthlessly

The agent will produce a voice profile. Read it critically. It should make you nod and think "yeah, that's how I write" at least 80% of the time. If a tone marker doesn't ring true, cut it. If it missed a phrase you use constantly, add it. If it attributed a pattern to you that was actually just one sample, remove it.

This is collaborative extraction, not automated analysis. The agent does the heavy lifting of finding patterns. You do the quality control of confirming which patterns are real.

Step 4: Save it to your workspace

Save the profile to templates/voice/core-voice.md in your GTM workspace. This becomes the reference file that your CLAUDE.md points to whenever content gets created.

Here's what the finished structure looks like:

# Voice DNA

## Identity
- Role: [your actual role, not a title you wish you had]
- Domain: [what you sell, who you sell to, what you know cold]
- Authority: [what gives you the right to write about this]

## Tone
- [5-7 tone markers, e.g.: direct, technical, dry humor, zero patience for buzzwords]
- Default register: [casual / professional-casual / formal]

## Vocabulary
- Words I use naturally: [list from your actual writing]
- Phrases I reach for: [list]
- Technical terms I use without explaining: [list]

## Structure
- Sentence length: [short / mixed / long]
- Paragraph length: [1-2 sentences / 3-4 sentences]
- Lists: [frequent / occasional / rare]
- Questions to the reader: [yes, often / sometimes / never]

## What I Sound Like
[2-3 sentences from your best writing that capture your voice perfectly.
These serve as a calibration reference.]

The Anti-Slop Rules

This is the highest-ROI section in this chapter. I mean that literally. If you do nothing else from this entire book, create an anti-slop file and reference it in your CLAUDE.md. It will improve every single piece of content your agents generate, immediately.

Why slop happens

Language models generate text by predicting the most likely next token. In a professional writing context, the most likely tokens are the most common professional phrases. "Leverage" is more statistically probable than "use" in business writing. "In today's rapidly evolving landscape" is a high-probability opening because thousands of business blog posts start that way.

Anti-slop rules override those probabilities. When you tell the model "never use the word leverage, use 'use' instead," you're redirecting it away from the generic toward the specific. Do this across 15 to 20 words and phrases and the cumulative effect is dramatic.

Words to ban

Instead of	Write
leverage	use
utilize	use
streamline	speed up, simplify
game-changer	[be specific about what changed and by how much]
cutting-edge	new, latest
revolutionary	[just describe what it does]
synergy	[delete the sentence]
elevate	improve, raise
robust	solid, strong
unlock	get, access, enable
deep dive	look at, dig into
empower	help, enable, let
landscape	market, world, space
ecosystem	market, community, stack
at the end of the day	ultimately, in practice

Patterns to ban

These are more insidious than individual words because they shape the entire feel of a piece.

"In today's [anything]..." Delete the whole sentence. Start with the point. If your content needs a "today's landscape" warmup sentence before getting to the actual idea, the idea isn't strong enough to lead with. Fix the idea.

"Let me be clear:" Just be clear. You don't need to announce it. This is the written equivalent of someone saying "honestly" before every sentence. It signals that you're usually unclear.

"Here's the thing:" Overused as a transition. If you genuinely use this phrase in real conversation, keep it. If not, cut it.

Opening with a rhetorical question you immediately answer. "What if I told you that 73% of GTM teams are wasting money on tools?" Just tell me. Don't What If me.

The LinkedIn one-liner power close. You know the pattern. A whole post building to a final line that sits alone for dramatic effect. "The best time to start was yesterday. The second best time is now." This is the single most detectable AI pattern on LinkedIn right now. If your content ends this way, everyone who reads it will assume a robot wrote it.

Triple emphasis. "This is not just important. It's critical. It's essential." Pick one word. Say it once. Move on.

The coffee test

Read your generated content out loud. If you wouldn't say it to a colleague over coffee, rewrite it. That's the whole test. No framework needed. No scoring rubric. Just: would you actually talk like this?

If the answer is no, you haven't extracted your voice well enough or your anti-slop list needs more entries. Go back to your samples and look for the gaps.

Before and After

This is where it gets concrete. Here's what the same content looks like without Voice DNA versus with it.

Cold email, no voice profile:

Subject: Streamlining Your Outbound Strategy

Hi Sarah,

In today's competitive B2B landscape, companies like Acme Corp are increasingly leveraging AI-powered tools to revolutionize their go-to-market strategy. I wanted to reach out because I believe our solution could help elevate your outbound efforts and unlock new pipeline opportunities.

Would you be open to a quick call to explore how we might be able to help?

Every alarm bell is ringing. "Streamlining." "Landscape." "Leveraging." "Revolutionize." "Elevate." "Unlock." Six slop words in four sentences. This email gets deleted in two seconds.

Same email, with Voice DNA applied:

Subject: Your Apollo setup might be leaving data on the table

Hi Sarah,

I noticed Acme just posted three SDR roles on LinkedIn. That usually means outbound is working well enough to scale, or not working well enough and you're throwing bodies at it.

Either way, we've helped two companies your size cut enrichment costs by 60% by replacing their Clay/Apollo overlap with a single Python pipeline. Takes about a week to set up.

Worth a 15-minute look?

Same sender. Same product. Completely different email. The difference isn't talent or copywriting skill. It's that the second version was generated with a voice profile that said "direct, specific, reference real signals, one ask per email, no buzzwords."

LinkedIn post, no voice profile:

I'm thrilled to share that we've just launched our new AI-powered GTM platform! After months of hard work by our incredible team, we're excited to bring this game-changing solution to market. In today's rapidly evolving landscape, B2B companies need to leverage cutting-edge tools to stay competitive. Our platform streamlines the entire outbound process, from enrichment to sequencing.

The future of GTM is here. Are you ready?

Reading this physically hurts. But scroll LinkedIn for five minutes and you'll find a dozen posts exactly like it.

Same announcement, with Voice DNA:

We shipped the GTM pipeline tool last week. Here's what it actually does:

You give it a list of target accounts. It enriches them through Apollo, scores them against your ICP, writes personalized first lines, and loads them into your sequencer. The whole thing runs in about 4 minutes for 500 accounts.

We built it because we were doing this manually and it took our team 6 hours a week. Now it takes 4 minutes. That's not marketing copy. That's a stopwatch.

If you're running outbound and spending hours on list building and enrichment every week, I'll show you how we set it up.

Same announcement. One version sounds like a press release written by committee. The other sounds like a person talking about something they built and why it matters.

The Voice DNA didn't write the second version. It prevented the first version from happening.

Building a Content Pipeline That Scales

Voice DNA isn't just about individual pieces of content. It's about building a system that produces consistently good content without you writing every word.

Here's how the pipeline works:

Layer 1: Voice DNA files. Your core voice profile, your anti-slop list, and your platform playbook. These live in your workspace and get loaded before any content generation.

Layer 2: Content templates. Structural templates for each content type. A LinkedIn post template that specifies your preferred structure (hook, body, close, no power-liner). A cold email template that specifies length constraints and the "one ask" rule. A blog post template that specifies your section style and how you handle transitions.

Layer 3: Source material. The raw inputs for content. Your ICP data. Your customer stories. Your product updates. Your market observations. Industry news you have opinions about.

Layer 4: Agent generation. Claude Code reads your voice DNA, loads the template for the content type, incorporates the source material, and generates a draft. Because it has all three layers of context, the output sounds like you wrote it.

Layer 5: Human review. You read it. You edit the 10-20% that doesn't land. Over time, as your voice profile gets more refined and your anti-slop list gets more comprehensive, that percentage drops. Some people get it down to 5% edits after a month of iteration.

The compound effect here is real. Every piece of content your agent generates and you edit is training data for the next generation. Not literally (the model doesn't learn from your edits), but practically. Each time you edit something and notice a pattern ("the agent keeps using 'however' as a transition and I never do"), you add that to the anti-slop list. The next piece of content is better. And the next. And the next.

After a month of this, your voice profile is dialed in tight. After three months, people who follow you on LinkedIn can't tell which posts you wrote manually and which ones your agent drafted. That's the goal. Not "hide that AI was involved." It's "the AI output is indistinguishable from my best writing because it has deep context about how I write."

Platform Adaptation

Your voice stays the same across platforms. The format changes. This is an important distinction. You don't become a different person on LinkedIn versus email versus your blog. You adjust the packaging for the medium.

LinkedIn

Short paragraphs. One to two sentences per block, max. The feed is dense and your post needs visual breathing room.

The hook in the first line matters more than anything else you'll write. Not clickbait. A real hook that earns the next line. "I wasted $2,400/month on GTM tools" works because it's specific and raises a question. "5 tips for better GTM" doesn't work because it sounds like everything else in the feed.

No links in the post body. LinkedIn's algorithm punishes posts with outbound links. Put links in the first comment. This isn't a hack or a trick. It's just how the platform works.

Personal experience beats abstract advice every time. "I did X and learned Y" outperforms "5 tips for Z" by a wide margin. Your voice profile already captures this if you extracted it from real writing. Reinforce it in your platform playbook.

Email (outbound)

One ask per email. Not two. Not "and also." One. If the prospect needs to make two decisions to respond, you've cut your reply rate in half.

Under 100 words for cold outbound. Under 150 for warm follow-ups. Every word past 150 is actively working against you. The prospect glances at the email, sees a wall of text, and moves on.

Sound like a person, not a company. First person singular. "I" not "we." Even if you work at a company with a thousand employees, the email is from you as a human, not from the corporate entity.

Reference something specific about them. Not "I saw your company is growing." That's generic and everyone says it. "I saw you hired 3 AEs last month and your Apollo enrichment setup is still manual." That's a signal that says I actually did the research.

Blog and long-form

More room to breathe. You can be detailed and technical here. Blog readers chose to click on your article. You've already earned their attention. You don't need to fight for every line the way you do on LinkedIn.

Subheads matter because readers scan before they commit to reading. If your subheads tell a story on their own, the scanner converts to a reader.

Real numbers, real examples, real specifics. "We reduced enrichment time by 97%" is a claim. "We went from 6 hours of manual enrichment per week to 4 minutes with a Python pipeline" is proof. The second one is harder to write because you need the actual numbers. That's why it works.

Why This Matters for Pipeline

Two specific reasons this is worth the effort, both measurable.

Outbound reply rates. Personalized emails that sound like a real human get 3 to 5x more replies than templated slop. When your coding agent writes outbound using your voice profile plus your ICP data plus your positioning, the output is a personalized email that reads like you sat down and wrote it for that specific person. Because structurally, that is what happened. The agent just did the sitting-down part faster.

At 500 outbound emails a month, going from a 2% reply rate to an 8% reply rate is the difference between 10 conversations and 40 conversations. Same list. Same product. Same total effort. Just better content.

Brand compounding. Consistent voice builds trust. Trust builds pipeline. When someone reads your LinkedIn post, clicks your profile, reads three more posts, and they all sound like the same person with the same perspective and the same specific-not-generic style, that's brand equity you can't buy with ads.

Voice DNA is what makes that consistency possible at scale. Without it, your Monday post sounds different from your Thursday post because the AI defaulted to different phrasings each time. With it, everything sounds like you, because the agent has a detailed map of what "you" sounds like.

The math works out simply. Time invested in Voice DNA extraction: 30 minutes. Time saved per piece of content: 10-15 minutes of editing. Break-even point: your third piece of content. Everything after that is pure leverage.

And unlike the word "leverage" itself, this one you actually want.