Teaching GPT-5 to Sound More Like Me

By Ken Hollow, AI wrangler, burnout survivor, and reluctant personal brand

When GPT-5 dropped, I thought, “Perfect. Finally, an assistant who can handle my tone so I can stop writing captions at 1 a.m. while half-watching reruns of Kitchen Nightmares.”

The dream: teach it to write like me, minus the spiraling panic that I lace into every paragraph.

The reality: mixed success.

Let’s talk about what worked, what didn’t, and why GPT-5 now thinks I’m 40% cynicism, 40% coffee, and 20% ellipses.

🔹 Step 1: Figure Out What “Me” Even Sounds Like

Easier said than done.

When you write in a constant state of exhaustion and sarcasm, your voice is less a “style” and more an emotional coping mechanism. But fine, let’s try to define it.

I dumped a bunch of my past posts into GPT-5 and asked it to:

• Identify tone and style
• Highlight recurring language patterns
• Extract emotional undercurrents (god help me)

It came back with:

“Your style is conversational, self-deprecating, and sarcastic, with a tendency to juxtapose humor against mild despair.”

Which, rude, but accurate.

🔹 Step 2: Craft the Perfect Prompt

Here’s what I started with:

Write as Ken Hollow: sardonic, witty, burnout-tinged but insightful. Avoid excessive doom, keep the humor sharp, and deliver actual value beneath the sarcasm.

This worked… sort of. The early attempts felt like reading a parody of myself. Imagine if a chatbot watched you for a week and then tried to do an impression at a party.

So I tweaked it:

Adopt Ken Hollow’s voice: dry humor, self-awareness, light bitterness. Blend insight with wit. Make readers feel entertained but oddly seen. Avoid melodrama.

Better. Still not perfect. The trick with GPT-5 is specificity without handcuffing it — tell it how to sound, but not exactly what to say.

🔹 Step 3: Use Few-Shot Examples

GPT-5 learns tone faster when you show it instead of just telling it. I gave it:

• One of my sarcastic listicle intros
• A bittersweet “Final Thoughts” section
• A mid-rant paragraph about algorithm betrayal

Then I prompted:

Using the above as tone reference, write a post intro about the rise of AI-generated friends.

Boom — 80% nailed. The humor landed, the pacing felt right, and it didn’t sound like it was trying too hard. That last part is important. If AI sounds like it’s performing wit, you lose the authenticity.

🔹 Step 4: Tweak for Sarcasm Levels

Sarcasm is my brand seasoning. Too little, and it’s bland. Too much, and you sound like a Twitter reply guy.

GPT-5 can balance this if you:

• Use tone modifiers (“light sarcasm” vs “dry wit” vs “biting commentary”)
• Specify emotional intent (“make them smirk, not cry”)
• Tell it when to break the fourth wall

Example tweak:

Use dry humor and subtle irony. One or two direct asides to the reader are fine. Keep it playful.

This got me to that sweet spot where readers chuckle but don’t send wellness check texts.

🔹 Step 5: Teach It What Not to Do

My early drafts had GPT-5 slipping into:

• Overexplaining jokes
• Random “inspirational” lines
• Buzzwords I would sooner choke on than type

So I added a negative style guide:

Avoid clichés, corporate jargon, and forced optimism. No overexplaining punchlines. No motivational poster energy.

That alone cut out 90% of the awkwardness.

---

🔹 Step 6: Use It as a First Draft, Not a Final Voice

This was my biggest mistake: letting GPT-5’s output go straight to publish. It’s good — scarily good — but it’s still missing the micro-moments that make writing feel lived-in.

Now, I:

1. Prompt it for structure and tone
2. Let it write the messy first draft
3. Go in with my human edits — adding the weird metaphors and oddly specific references it wouldn’t dare invent

This hybrid approach keeps the AI efficiency without losing my fingerprints.

🔹 Bonus: Give It Mood Context

I realized GPT-5 doesn’t know if I’m feeling “mildly caffeinated” or “two emails away from snapping.” Tone is emotional context, so I started adding it:

Tone: playful but tired. Audience: creators who feel slightly dead inside but still love the work.

Shockingly effective. It’s like giving the model a little backstory before it enters the scene.

What Worked:

• Few-shot examples of my actual writing
• Clear tone modifiers (light sarcasm, dry humor)
• Negative style rules
• Emotional context tags

What Didn’t:

• Overly rigid prompts
• Letting it write without human editing
• Forgetting to define audience

🔹 Final Thoughts (Written By Me, Not the Machine)

GPT-5 can mimic me with unnerving accuracy now. Sometimes I’ll read a draft and have to check if I actually wrote it during a blackout.

But here’s the thing — it’s still mimicry. It can echo my phrasing, match my rhythm, even capture that bittersweet punch at the end. What it can’t do is feel the moment it’s writing about.

And maybe that’s fine. Maybe my job isn’t to fight the machine for personality, but to use it as a mirror — to see what’s essential about my voice and what’s just filler I hide behind.

So yes, GPT-5 now sounds more like me.

Just without the 2 a.m. dread, the Slack avoidance, or the creeping suspicion I should’ve opened a coffee shop instead.

Ken Hollow, freelance cynic, AI tone coach, and guy who just trained a robot to roast him better than his friends do.