Most AI-generated content sounds like AI-generated content.
You know it when you see it: the hollow enthusiasm, the "best-in-class solutions that empower teams to unlock value," the way every sentence could have been written by anyone, for anyone, about anything.
The models are fine. Your prompts, however, are context-starved. You're asking a pattern-matching machine to write like you without showing it any patterns, so of course it defaults to the median. The median is all you gave it.
The fix Is boring (and it works)
One plain-text markdown file (.md) that lives in your context window (the text the model sees each time you prompt it). It tells the model how you actually write. You build it once and use it forever.
I've been running a version of this for over a year, and the difference is night and day.
Anatomy of a voice file
A good voice file has five parts. None of them are complicated.
1. Tone as contrasts
LLMs understand positioning better than absolutes. "Write confidently" is vague, but "confident without being certain, casual without being sloppy" gives the model something to calibrate against.
From my own file:
Conversational surface, rigorous underneath. Use contractions naturally. Direct address. Occasional dry asides. But always grounded in real data and careful sourcing. The casual tone earns trust; the rigor keeps it.
You're defining a range, not a point, and the model can locate itself within that range.
2. Sentence-level patterns
How long are your sentences? When do you use fragments? How do you handle transitions? This stuff matters more than most people think, because voice isn't just word choice. It's rhythm.
Short paragraphs. Rarely more than 4-5 sentences. Single-sentence paragraphs for emphasis. Use sparingly.
Vary length. Punchy declaratives mixed with longer explanatory ones. Avoid monotony.
Read your last ten pieces and count the sentences per paragraph. Notice when you break rhythm and why. Write it down.
3. Words and phrases you actually use
These are your verbal fingerprints, the stuff that sounds like you and nobody else.
I use "yeesh." I make dry parenthetical asides and cite specific numbers instead of "millions of users." When I'm being precise about what technology can and can't do, I'll say something like "the tools have made X diagnosable." Small things, but they add up to a voice. My voice.
4. The banned list
This matters more than the positive guidance. LLMs love certain constructions and reach for them constantly, so if you don't explicitly ban them, they'll show up in everything.
My list has about 40 words that are instant tells: delve, unlock, embrace, beacon, unprecedented, elevate, supercharge. If a word could appear in anyone's marketing (or if it smacks of LLM slop), it's out. But the structural bans matter even more.
I banned the "It's not X, it's Y" construction entirely. ("This isn't about visibility. It's about revenue.") LLMs love this move because it sounds punchy, but it's a crutch, and once you notice it you can't stop noticing it.
Also banned: filler phrases ("It's important to note..."), rhetorical tics ("Here's the thing"), hollow enthusiasm without evidence, em dashes (colons and parentheses do the same work with less drama). The banned list is where you exorcise the slop, so be aggressive.
5. Sample transformations
Before/after pairs showing generic writing versus your actual voice. The model pattern-matches against these, so three or four good examples go a long way.
Too generic:
Organizations are increasingly leveraging AI to enhance their go-to-market capabilities.
Better:
91% of GTM teams are using AI tools. 53% report seeing no meaningful impact.
Too hedged:
It might be worth considering that some buyers may be starting their research in AI tools.
Better:
A growing share of discovery now happens inside closed AI systems. Among tech industry buyers, 80% use GenAI as much or more than traditional search when researching vendors.
Too breathless:
This is an incredibly exciting time for marketers as AI revolutionizes everything!
Better:
The tools have made visibility diagnosable. What they haven't solved is what comes next.
How to Build Yours
This takes 30-60 minutes, maybe less if you know your writing well.
Step 1: Pull your last 10 pieces. Blog posts, LinkedIn, emails to clients. Whatever represents your actual voice, not your aspirational voice.
Step 2: Read them out loud. Notice what sounds like you and what sounds like filler. Mark both.
Step 3: Answer these questions in writing:
How do I open a piece? (Anecdote? Question? Provocation? Statement of fact?)
How long are my paragraphs? Sentences?
What words do I use that other people don't?
What words do I hate seeing in my drafts?
When do I use humor? What kind?
How do I handle uncertainty? Do I hedge or commit?
Step 4: Write the banned list first. Seriously, this is where the leverage is. What do you delete every time you edit? What makes you cringe in other people's writing? Those are your bans.
Step 5: Add 3-4 before/after transformations. Take generic sentences and rewrite them in your voice. The model will extrapolate from there.
Step 6: Test it. Prompt the model with your file in context, ask it to write something short, see what's off, and revise the file. You're not going to nail it on the first try, and that's fine. Revise as you go.
A skeleton to start from
## Tone[Describe your voice as contrasts: "X but not Y, A without being B"]## Rhythm[Sentence length, paragraph length, when you use fragments]## Verbal fingerprints[Words, phrases, and moves that are distinctly yours]## Banned words and structures[The stuff you delete on sight. Be specific.]## Before/after examples[3-4 transformations showing generic vs. your voice]This is a starting point. You'll add to it as you notice patterns.
Beyond voice: Other context files worth building
Once you see how this works, you'll find other uses.
Brief files. Campaign context, audience specs, brand guardrails, competitive positioning. Everything a freelancer would need to write for you, except you give it to the model.
Process files. How you want a task done, step by step. I have one for running GEO audits and another for building content briefs. The file encodes decisions so I don't have to remake them every time.
Project files. Ongoing context that accumulates over the life of a project. Working on a six-month campaign? Keep a file with key decisions, approved directions, and stakeholder feedback. The model remembers what you'd otherwise have to re-explain.
Some tools (Claude's "skills" feature, for instance) let you build context files that also execute: they can run code, pull data, and generate outputs in specific formats.
That's a deeper topic for another time, but the principle is the same. You're compressing decisions you've already made so the model can apply them without being told.
Three ways this breaks
If your file is too long, the model starts ignoring parts of it. Keep it under a page or two.
If your rules contradict each other, the model averages them. "Be concise" plus ten paragraphs of instructions sends mixed signals.
If you describe your aspirational voice instead of your actual voice, the model writes like your LinkedIn alter ego. Use real samples, not fantasy.
The point
A voice file is a compression of decisions you've already made about how you sound. The LLM just needs them made explicit.
You're teaching the model to write like you already do, and it takes under an hour to build. Saves every hour after that.