💬
ChatGPT Tutorials
How I Use ChatGPT Custom Instructions to Clone My Voice
After 14 months of testing, here's the exact Custom Instructions setup that makes ChatGPT write in my voice well enough to fool my own editor.
Last March, my editor approved a 1,200-word column without a single change. I had written exactly 84 words of it. The rest came from ChatGPT, running on a Custom Instructions file I'd been refining for nine months.
This isn't a magic-prompt story. It's a calibration story. ChatGPT can mimic almost any writing voice if you stop treating Custom Instructions like a bio field and start treating them like a style guide written for a junior copywriter who has never met you.
Why most voice prompts fail
The default approach goes something like: "Write in a casual, conversational tone with humor." That instruction is useless. It describes roughly 60% of the internet.
Your voice isn't a tone. It's a collection of micro-habits: the sentence lengths you favor, the transitions you avoid, the punctuation tics nobody else has. When I audited 40 of my own published pieces, I found I open 70% of paragraphs with a subject-verb construction under eight words. I almost never use semicolons. I use em dashes constantly. I rarely end sentences with prepositions, but I'll start them with "And" or "But" without flinching.
Those aren't preferences. They're fingerprints. And fingerprints are what Custom Instructions need to capture.
The two-box framework
ChatGPT (as of the 2025 update on Plus and Team plans, $20/month) gives you two Custom Instruction fields: "What would you like ChatGPT to know about you?" and "How would you like ChatGPT to respond?" Most people stuff personality into the first box. That's the wrong move.
Box one should contain context — who you write for, what platforms, what you sell. Box two should contain mechanics — the actual rules of your prose.
| Field | Wrong approach | Better approach |
|---|---|---|
| About you | "I'm a friendly writer who loves tech." | "I write for B2B SaaS founders. Audience: 30-45, technical, skeptical of hype. I publish on Substack and LinkedIn." |
| How to respond | "Be casual and engaging." | "Avg sentence: 14 words. Vary 4-25. Use em dashes, not semicolons. Open paragraphs with short declaratives. Never use 'leverage', 'utilize', 'in order to'." |
Pro tip: Paste three of your best published paragraphs into ChatGPT and ask it to extract 15 stylistic rules. It catches patterns you can't see in your own writing — like how often you use parenthetical asides.
The exact build process
- Gather 8-12 samples. Pick pieces you'd happily republish. Mix formats: a newsletter, a long essay, a product page, a tweet thread. Diversity teaches range.
- Run a style audit. Prompt: "Analyze these samples and give me 20 specific, mechanical rules about sentence structure, word choice, punctuation, and rhythm. No vague descriptors like 'engaging' or 'witty'."
- Cut to 15 rules. Keep the ones that feel like you and would feel wrong if violated. Drop generic advice.
- Add banned phrases. List 10-20 words you'd never write. Mine includes "delve", "robust", "seamless", "in today's world", and "game-changer".
- Add required habits. I require: at least one em dash per 200 words, no sentence over 30 words, one specific number or named example per section.
- Test on a known piece. Ask ChatGPT to rewrite an article you've already published. Compare line by line. Edit your instructions based on the gaps.
- Re-test monthly. Your voice drifts. So should the file.
The output rules that actually move the needle
Three instructions did more for my output quality than the other twelve combined.
One: forbid soft transitions. "Never use 'Moreover', 'Furthermore', 'Additionally', or 'In conclusion'." These are tells. Removing them forces ChatGPT to write transitions that actually carry meaning.
Two: demand specificity. "Every claim must include a number, a named tool, or a concrete example. No abstract generalities." This single rule killed about 80% of the AI-flavored mush.
Three: control the opener. "Never begin with 'In' or 'The'. Start with a specific story, a counterintuitive claim, or a number." Openings are where AI writing announces itself. Lock the opener and the rest follows.
Pro tip: Add a "voice check" instruction: "Before finalizing, scan the draft and flag any sentence that sounds generic AI. Rewrite those." ChatGPT will self-correct about 60% of the time. It's not perfect, but it cuts editing time roughly in half.
What this still can't do
Custom Instructions handle mechanics. They don't handle judgment. ChatGPT still can't decide what's worth saying — which anecdote lands, which argument is too obvious, which joke is actually funny versus merely structurally correct.
I write the spine. I draft the hook, the thesis, and the closing line by hand. Then I let ChatGPT fill the connective tissue against my style file. The ratio is roughly 15% me, 85% machine — but the 15% is the part that matters.
FAQ
Does this work on the free version of ChatGPT?
Yes. Custom Instructions are available on free, Plus, and Team tiers. The free model (GPT-4o mini as of 2025) is less consist