How to Use AI to Accelerate Content Creation Without Sounding Like Our Robotic Overlords
If you know me, you know I’ve been publishing content since the late 90s. Back then, "tweaking" meant messing with your Windows 95 registry. So it’s no surprise that I am dropping content all over the place these days. I could probably keep this pace up on my own, but I’ve relied on external editing for a long time to make sure I show up the best way I can.
The difference is that now, more often than not, AI provides that help instead of another human.
But let's be real for a second. It took me (more than) a few rounds to get it right. I had to learn how to avoid a bad case of "AI voice." I think that’s the written equivalent to Ozempic butt, or maybe a form of digital COVID. Whatever it is, it’s gross. You may not be able to describe it, but we all know it when we see it, right?
We're watching a massive shift in the enterprise. Spending on generative AI has tripled to $37 billion in 2025. 60% of organizations have deployed these tools. But throwing money at an LLM to generate content doesn’t make you a thought leader... it usually just makes you a spammer. I thought I'd share my workflow so you can scale your output too without sacrificing the authenticity that got you a seat at the table.
The Art of the Prompt
Smart people make a basic mistake here. They treat their AI interface like a single, monolithic oracle. They type a massive, sprawling prompt into a fresh window and hope for the best. That’s a rookie move. One context is bad. Many contexts are good. You need to treat your AI interactions like you’re building a team of specialist, genius ninja toddlers.
(Yes I'm referring to another article I wrote there. No way an AI would think to do that, right?)
I create custom spaces (I'm a big Perplexity user). Depending on your AI environment of choice you might call them Gems, GPTs, or Projects. Each instance is configured to act in a different way. This separation of duties prevents the "muddying" of instructions. If you ask the same bot to be a creative writer, a fact-checker, and a compliance officer all in one thread, it defaults to the lowest common denominator. You get safe, boring, robotic text that reads about as original as... I can't think of a good example of the something more artificial than AI writing. Damn.
Anyhow, here's the "org chart" of AI spaces I keep on my sidebar.
The Prompt Builder
This is the meta-tool. I use this space solely to build prompts for other spaces. It’s trained on chain-of-thought reasoning. I tell it, "I need a prompt that will get an LLM to critique a technical whitepaper from the perspective of a cynic." It generates the precise instructions I need and asks me clarifying questions if my ask isn't clear. This cuts down my iteration time.
The Researcher
This thing has one job, it finds information. I explicitly instruct it *not* to be creative. I want it to scour the web, parse technical documentation, and find sources that agree or disagree with position X. If I’m writing about the shift from ETL to ELT data pipelines, I’ll ask the Researcher to find benchmarks from 2024 and 2025 that contradict my assumption. This prevents confirmation bias. AI is notorious for reinforcing that bias if you aren't careful. This is also where I dump PDFs of industry reports so I can query them specifically.
The Ghostwriter
This is the heavy lifter. The Ghostwriter takes a detailed outline – that I wrote MYSELF – and writes a rough draft. But here is the secret sauce. It is trained on *my* voice, *my* style guide, and *my* persona. I don’t just say "write this." I have uploaded samples of my writing from the last 20 years. It has my articles from 25 years ago. It has my current Field CTO thought leadership. I’ve given it a "Do Not Use" list of words. When I feed it an outline, it knows not to sound like a generic marketing bot. While it's not perfect, it's at least better.
The Editor
I never let the Ghostwriter edit its own work. The Editor is a separate space with different instructions. Its prompt is purely critical. I tell it, "You are a ruthless editor at a top-tier tech publication. Cut the fluff. Sharpen the arguments. Point out where the voice sounds fake." It takes the draft from the Ghostwriter and tears it apart. It gives feedback on sentence structure, voice, and document flow. Sometimes, it’s humbling. But it is necessary so I don't miss something.
The Manager
Finally, I have the Manager. This is my personal social media strategist. Once the piece is polished, I feed it to the Manager. I ask, "How do we promote this?" It breaks the post down into LinkedIn threads and pithy tweets. It knows the difference between my professional LinkedIn tone and my casual no-longer-Twitter voice.
As an aside, what's replaced formerly Twitter? Threads and Bluesky both kind of suck.
For what it's worth, I don’t use all of these spaces every time. I find that less technical content works better when I write most, if not all of the rough draft myself. I just use the Editor to polish those (it's hard to edit your own words without putting it away for two weeks). But for deep, technical content where I need to bridge the gap between "data engineer" and "CIO," I build a detailed outline. I let the Ghostwriter do its thing. Then I take the wheel for the final edit... the Editor ends up critiquing my draft and not the rough cut.
Training AI to Be Less Robotic
You can't expect an off-the-shelf model to sound like you. It’s trained on the internet. By default, it sounds like the average of the internet. That means bland, corporate, and safe - or maybe academic. Occasionally, it sounds like Reddit. To fix this, you need a voice training corpus.
This isn’t hard to build, and you don’t need to be a machine learning engineer. Just have your Research bot find a list of your publications. If you don’t have a lot of published writing, just dig up your last 5-10 high-quality pieces... even long emails can work. Feed them into the context window as attachments. Tell it, "Analyze the writing style, tone, sentence length, and vocabulary of these texts. Create a style guide based on this analysis." I even fed it my Linkedin feed and some longer emails I wrote.
If you don’t have a large body of prior work, that’s fine. Find a few authors you aspire to sound like. Maybe you like the punchiness of a specific tech journalist. Or you might prefer the strategic clarity of a certain CEO. Feed a few of their pieces into the machine for training. Don’t plagiarize their ideas. Just steal their cadence. How do you think I learned how to write, anyway?
Once you have that analysis, have the training corpus build you a few profiles. I keep a "Casual Keith" profile for blog posts and a "Professional Keith" profile for whitepapers and other more formal pieces. "Casual Keith" uses fragments, he starts sentences with "And" or "But.," and he might even throw in a mild "damnit" here and there. "Professional Keith" sticks to the data and the strategic implications, and is more based on my professionally edited pieces.
Iterative Refinement
The biggest failure mode in AI writing is the "one-shot" attempt. You give it a vague as hell prompt. It gives you 1,000 words of borderline nonsense. You hate it. The problem is usually ambiguity. AI is good at drafting. It is bad at creating. If your prompt is ambiguous, you get garbage.
My favorite trick is to tell the AI to interview you before it starts writing. My prompt looks like this: "I want to write a blog post about X. Here is my rough thinking. Before you generate any text, ask me 5-10 questions. Fill in the gaps in your understanding. Capture my specific perspective."
It will come back with questions. "What is your contrarian take on this?" "Can you provide a specific example from your time at a startup?" Answering these questions forces you to clarify your thoughts. It gives the AI the specific "meat" it needs to write a draft that sounds like you. Otherwise, you just get a Wikipedia summary.
Telltale Signs of Ozempic Butt (I Mean, AI Voice)
Even with all this training, "digital COVID" can sneak in. LLMs are trained using Reinforcement Learning from Human Feedback (RLHF). This rewards text that is helpful and safe. Unfortunately, safe usually means boring and repetitive. Here are (some of) the telltale signs that your post was written by a robot.
Overuse of Em-Dashes
I’ve used em-dashes for a long time. They add a punchy aside and can signify a hard break between related thoughts. But AI is obsessed with them. It uses them to replace soft breaks or commas. If you replace a comma with a dash, you are probably using it wrong (okay, not always... but it's still a good rule of thumb). LLMs use dashes to extend a sentence without committing to a full stop. So if you see a bunch of em-dashes in one paragraph, replace them with more appropriate punctuation.
You can even get wild and use fancy brackets. Tildes are nice too. I mean, tildes aren't grammatically useful replacements for em-dashes but it'll be clear they weren't AI generated at least.
Context Framing
AI loves to say things like "It’s not X, it’s Y." It does this in complex, compound sentences. "It’s not just about data storage; it’s about unlocking the potential of your intelligence." Even leaving that sentence in this piece is giving me the heeby jeebies. Models use it to sound profound without saying anything. Humans just say what the thing is, so do that.
Overuse of Formal Phrasing Instead of Casual
If your piece reads like a book report, you're doing it wrong. AI loves formal transitional phrases like "Moreover," "Furthermore," and "In conclusion." It almost never uses "But," "And," or "So" to start a sentence. But that is how people talk, and write.
I have a specific instruction in my style guide: "Banned words: Moreover, Furthermore, Delve, Verisimilitude" (I haven't really banned verisimilitude, I just thought it was funny - if you don't get why, look it up). When in doubt, replace them with more casual replacements. If you can delete the offending transition word entirely and the sentence still makes sense, do it.
Sentence Structure and Complexity
AI likes to write at a college reading level, using long, complex, compound sentences. It defaults to a "noun-heavy" style. Humans vary their sentence length... intermittently using short sentences between long winding ones when we try to explain a complex concept. Then we follow them up with something more direct. For most writing, you should target an 8th-10th grade reading level. Make sure you have varying sentence lengths.
Not sure how to do that? Check out the Hemingway editor [https://hemingwayapp.com/], it's the secret to a lot of great writing. I use the downloaded version on my laptop but the web version is great too. It will brutally show you where your "AI voice" is showing.
Vocabulary
I love me some $5 words. "Bifurcate." "Orthogonal." "Ephemeral." My favorite is "Corpus" (in fact, I used that one earlier). But if you can express your idea without them, ditch 'em. Oh, and if AI wrote this, it would have written "five-dollar words" instead of "$5 words." It lacks the nuance of idiom and will pretty much always avoid things like contractions. You want your writing to be perfectly imperfect.
Don’t Let AI Write the Conclusion
This is my golden rule. Even when you use a Ghostwriter for the rough draft, write the conclusion yourself. AI sucks at making a point. It loves to summarize by restating everything it just said. It usually starts with "In conclusion, the landscape of..." Don't let it do that unless you want it to sound like the essay you got a C on in 8th grade english class. Write the final paragraph yourself. Pepper in some real anecdotes. Don't let AI make that shit up. End with a point that feels like it came from a person.
As it stands, we are in a weird transition period. We can (mostly) spot AI writing from a mile away, but that window is closing with each new version of a model. That means the tells will get subtler and harder to detect. For now, if you want to be read, you have to sound like a human. So do the work by training your tools, editing your drafts, and injecting your own weird, messy, human perspective into everything you publish.
Oh, and add a picture of your cat (or your dog, or your watch, or whatever). That humanizes even the most robotic post.