I Fed My AI My Worst Writing Tics to Make It Sound Like Me

I have a confession to make. My name is Rohan Mehta, and I have a problem with the word ‘leverage’. For years, it has been my crutch. It’s in my old emails, my draft proposals, my performance reviews. It’s a word that sounds important while saying very little, the perfect sin of corporate speak. I was painfully aware of this tic, but I hit a new low a few months ago. I asked my AI assistant to help me outline an article, and there it was, in the second bullet point: ‘Leverage AI capabilities to enhance user workflow.’ The machine was mocking me. It was holding up a digital mirror to my own laziest habits.

That was my breaking point. As an editor at Pulse AI, I live and breathe this technology. The promise was always that it would save us time, freeing us up for more creative work. But in reality, I was spending a significant part of my day on a new kind of drudgery: de-botting text. I was constantly editing out soulless, generic phrases, trimming sentences that started with ‘In the digital age we live in…’, and fighting a tide of bland consensus that AI models are so good at producing. It felt like I was just a glorified copyeditor for a very fast, very prolific intern with no personality.

Then came the realisation. The AI wasn’t just a content generator; it was a mirror. A flawed one, yes, but it was learning from the vast ocean of text on the internet, which is full of my writing, your writing, and everyone’s cringiest habits. The problem wasn’t just the AI. The problem was also me. And instead of just correcting the output every single time, I wondered: what if I could teach it to be better? What if I could teach it to be more like me—or rather, the better version of me who doesn’t say ‘leverage’ every other sentence?

This led me down a path from being a simple AI user to becoming an AI trainer. I stopped thinking in terms of one-off prompts and started thinking about creating a permanent ‘personal style profile’. It’s a document, a set of instructions, that goes beyond telling the AI what to do. Crucially, it tells the AI what *not* to do.

This concept of a ‘negative style guide’ is, I believe, the key to unlocking the next level of human-AI collaboration. You're not just giving it a fish; you're teaching it how to fish, while also providing a detailed list of poisonous fish to avoid. This moves you from getting a generic draft to co-creating a piece of work that has your unique fingerprint on it.

So, how do you do it? The first step is the most humbling: you must conduct a ruthless self-audit. You need to become an archaeologist of your own linguistic flaws. I created a folder on my desktop and dumped in everything I’d written over the past two years—articles, major reports, even some blog posts from my early career days in Mumbai that I thought were long buried. I copied and pasted thousands of words into a single document.

Then came the moment of truth. I used the simple ‘Find’ function. ‘Leverage’: 112 instances. Sentences starting with ‘In fact…’: 48. ‘Deep dive’: 31. The list went on. I saw my pet phrases, my favourite sentence structures, my go-to jargon that I used to try and sound smart for clients in San Francisco and London. Reading it all together was excruciating. But the cringe is your compass; it points you toward the things you need to fix.

Once you have your list of shame, the next step is to codify your anti-rules. You must be brutally specific. An AI doesn’t understand nuance unless you spell it out. I created a new document that became my master style guide, divided into sections.

First, Vocabulary. I made two columns: ‘Avoid’ and ‘Prefer’. Under ‘Avoid,’ I listed: leverage, synergy, impactful, touching base, futuristic, paradigm shift, and a dozen others. Under ‘Prefer,’ I wrote their simpler, more direct alternatives: use, collaboration, effective, checking in, forward-looking, fundamental change. This isn't just about single words; it's about shifting the entire texture of the language from jargon-filled to human-centric.

Second, Syntax and Structure. This is where you can get really granular. My rules included: ‘Do not start more than one sentence per paragraph with a transition word like However, Therefore, or Moreover.’ ‘Keep the average sentence length below 20 words.’ ‘The passive voice should only be used when the actor is unknown or irrelevant; aim for active voice 95% of the time.’ ‘Limit the use of nested clauses. A sentence should rarely have more than one comma.’

Third, and perhaps most importantly, Tone and Personality. This is where you inject your soul. My instructions read: ‘Write with a tone of cautious optimism. Be analytical but not cold. Use analogies, but draw them from everyday life (like cooking, gardening, or travel) and avoid clichés from sports or war. Be direct and confident, but not arrogant.’ I even added a rule reflecting my own background: ‘When discussing Indian business or culture, it’s acceptable to use a common Hinglish word like ‘jugaad’, but you must briefly explain its meaning for a global audience.’

With my style guide drafted, the next logical step was implementation. Most modern AI chat interfaces, from ChatGPT to Claude, have a feature for this. It’s usually called ‘Custom Instructions’ or something similar. It’s a persistent memory for the AI. You copy your entire style guide—your banned words, your syntax rules, your tonal preferences—and paste it into this section. This becomes the AI’s foundational document for every single interaction it has with you. It’s the master brief that says, ‘Before you write a single word for Rohan, read this.’

Of course, this is not a one-and-done solution. It’s an iterative process. The first time I tested my new instructions, I asked the AI to draft an email about a project delay. It brilliantly avoided ‘leverage,’ but it replaced it with ‘utilize’ three times in one paragraph. I let out a deep sigh. But it wasn’t a failure; it was feedback. I went straight back into my custom instructions and added ‘utilize’ to my banned list. This is the new workflow. It’s a conversation. You provide instructions, the AI attempts to follow them, you observe the output, and you refine the instructions. Each loop closes the gap between its voice and your intended voice.

Over the last few months, this process has had an unexpected side effect. In teaching my AI to be a better writer, I have become a more self-aware writer myself. The act of codifying my anti-habits has made me intensely conscious of them in my own raw, unassisted typing. I now physically recoil when I find myself about to type ‘let’s touch base offline’. By defining my voice so explicitly for a machine, I have clarified it for myself. It’s a strange but powerful form of digital therapy for bad writers.

We are at an inflection point in our relationship with artificial intelligence. The novelty of generating instant text is wearing off, and the frustration with its generic nature is setting in. The way forward isn’t just about crafting cleverer prompts. It's about deep personalization. It’s about taking the time to teach the tool, to shape it in your own image—or, more accurately, in the image of the writer you aspire to be.

Creating a personal style profile turns the AI from a clever parrot into a true apprentice. It’s an assistant that learns your preferences, understands your weaknesses, and adapts to your voice. My AI is still learning, and so am I. But its drafts no longer feel alien. They feel like a starting point I can actually build on. They sound, more and more, like me. And for that, I am willing to leverage all the tools at my disposal. Wait, no. I mean, I am happy to use them.