You can spot AI-written copy from across the room. It “delves”. It “navigates the evolving landscape”. Every paragraph ends in a neat little summary, and everything comes in threes. It’s grammatically fine and completely forgettable.
The problem isn’t that ChatGPT can’t write. The problem is that most people give it nothing to work with, so it falls back to its default voice — a sort of polite corporate beige. If you want drafts that sound like you and need a light edit rather than a rewrite, you have to set it up properly.
A recent guide from The Rundown AI lays out a four-part approach. I’ve worked through it, added some practical notes, and flagged where you need to be careful. Here’s what actually matters.
Step 1: Write a Voice Guide Before You Write Anything Else
This is the step most people skip, and it’s the one that does the most work.
A voice guide is a short document that tells the model how you write. It doesn’t need to be long. It should cover:
- Patterns you use. Short sentences? Plain language? Do you use “you” and “I”? Australian spelling?
- Before-and-after examples. Take a generic AI paragraph and rewrite it the way you’d say it. Two or three of these teach the model more than a page of adjectives.
- Banned words and phrases. Make a list of the things that make you cringe. “Leverage”, “seamless”, “game-changer”, “in today’s fast-paced world” — whatever your personal blacklist is.
Put this in your Project instructions in ChatGPT so it applies to every chat inside that project, rather than pasting it in each time. If you use other tools — Claude Projects, Copilot agents, or coding agents that read an AGENTS.md file — the same document works there too. Write it once, reuse it everywhere.
Step 2: Feed It Your Own Words, Not Just a Topic
Ask ChatGPT to “write a memo about X” and you’ll get a memo about X in ChatGPT’s voice. There’s no source material, so it has nothing to copy but itself.
The better approach is to give it your raw language as the starting point. The Rundown’s suggested prompt structure is:
“Use this transcript to draft a [memo/email/tutorial]. Preserve every source fact and the way I naturally explain the topic.”
Good raw material includes:
- Dictation. Talk through the topic for five minutes on your phone and paste the transcript in.
- Meeting transcripts. If you’ve already explained something well to a client or your team, that explanation is gold.
- Voice notes you’ve recorded in the car or on a walk.
This works because the model is now editing and structuring your thinking rather than inventing it. The phrasing, examples and opinions stay yours.
A caution from the IT side: be deliberate about what transcripts you paste in. A meeting recording with a client may contain commercially sensitive or personal information. If you’re on a personal ChatGPT plan, check your data controls, and if it’s client material, use a business-tier account with appropriate data protections — or strip out the identifying details first. Your voice isn’t worth a privacy breach.
Step 3: Run a “Humaniser” Pass
Even with good inputs, AI drafts tend to carry tell-tale habits: inflated significance, forced rule-of-three lists, vague attributions, and em dashes everywhere. A second, dedicated clean-up pass catches these.
The guide recommends Humanizer, an open-source skill by developer Blader that’s designed to strip out common AI writing patterns. It installs via the terminal with:
npx skills add blader/humanizer --global
Once installed, you run it against a copy of your draft and it rewrites the robotic bits into something more specific and natural.
Know the limitations before you dive in:
- Skills in ChatGPT are currently aimed at Business, Enterprise and Edu workspaces, and availability depends on your workspace settings. If you’re on a personal plan, you may not have this option yet.
- It’s a community-built tool, not an OpenAI product. Treat it like any third-party code: review what you’re installing, particularly on a work machine.
- The same skill format works in other agent tools such as Claude Code, Codex and Cursor, so it’s not locked to one platform.
If terminal installs aren’t your thing, you can get a decent chunk of the benefit simply by adding a “patterns to avoid” section to your voice guide and asking the model to self-check its draft against it.
Step 4: Borrow a Technical Writing Standard
This is the most interesting tip in the guide, and the one I’d recommend for anyone writing procedures, documentation or client-facing technical content.
Add this to your prompt or project instructions:
“Write in Technical English. Use ASD-STE100 as a style reference.”
ASD-STE100 is Simplified Technical English — a controlled-language standard from the aerospace industry, created so maintenance manuals could be understood clearly by people working in English as a second language, often under pressure. It favours short sentences, one instruction per sentence, active voice and a restricted vocabulary.
You’re not trying to formally comply with the standard. You’re borrowing its discipline. The result is noticeably more direct: less padding, fewer hedges, and no fluffy transitions. For IT runbooks, how-to guides and policy documents, it’s a genuine upgrade.
It’s less suited to marketing copy or anything where personality is the point, so use it selectively.
Test One Change at a Time
The final piece of advice is the most practical: change one thing at a time.
Pick a single writing task — say, a 300-word client update — and run it through each method separately. Voice guide only. Then voice guide plus transcript. Then add the humaniser pass. Then try the technical English rule. Compare the outputs side by side.
If you change everything at once, you won’t know what’s actually helping. And you’ll likely find that one or two of these techniques do 80% of the work for your particular style.
The Honest Take
None of this is magic. AI still won’t have your opinions, your client history or your judgement — and it shouldn’t pretend to. What these techniques do is cut down the editing time between “first draft” and “something I’d put my name to”.
The biggest wins, in order:
- A proper voice guide in your project instructions (15 minutes, pays off forever)
- Starting from your own words instead of a blank prompt
- The ASD-STE100 trick for technical and procedural content
- A humaniser pass if you have access to skills and are comfortable with terminal installs
If you do nothing else, do the first two. You’ll notice the difference on the very next draft.
Getting individuals using AI well is one thing. Getting a whole team producing consistent, on-brand, secure output — with the right data protections in place — is another. Get in touch to talk about AI setup for your business.