QWE AI Academy Quick test: open ChatGPT and Claude side-by-side. Ask both to write a LinkedIn post about lessons learned this year. The outputs won’t be identical, but they’ll share the same confident cadence, the same three-part structure, the same absence of genuine uncertainty.
This isn’t a bug. It’s training data.
Two Ways to Escape the Default Voice – One Actually Works
Method A: Tweak temperature and top-p parameters. Crank temperature to 0.9, set top-p to 0.85, hope for variety. You’ll get output that’s more random, sometimes more creative. Often just messier.
Method B: Custom instructions + few-shot prompting with your actual writing. Upload 5-10 samples of your own prose, let the model extract your patterns, apply them consistently. Takes 10 minutes upfront. Works for months.
Method B wins. Why, and how to do it without the usual pitfalls.
Why Every LLM Sounds Like a Confident Reddit Comment
Reddit: 40.1% of all LLM citations (2025 Semrush analysis), nearly double Wikipedia’s share. That 2016 user base? 69% male, compared to 49% of the general population. The internet these models trained on skewed toward assertive, declarative prose optimized for upvotes.
Then RLHF stripped out hedging – “I think,” “as far as I know,” the phrases that signal appropriate uncertainty. Models now sound confident even when they’re wrong. Carnegie Mellon confirmed it: LLMs maintain or increase stated confidence regardless of actual performance. Humans adjust based on results. LLMs don’t.
That’s the homogenization everyone’s complaining about on LinkedIn right now.
The Problem with Just Adjusting Temperature
Temperature controls creativity, sure. Lower = deterministic; higher = random. But parameters fight when you stack them wrong.
Set temperature to 1.2 for creativity, then add custom instructions asking for “concise, factual responses” – they clash. Temperature wins. You get creative nonsense instead of creative-but-coherent prose. I tested this with Claude Sonnet 4.6 and GPT-5.1. Same prompt, temp 0.9 + strict factual instructions. Both models produced meandering output that ignored half the constraints. Lower the temp to 0.3? Instructions work again – but you’re back to predictable phrasing.
The fight happens because OpenAI and Anthropic both say: alter temperature OR top-p, not both. Most tutorials skip this.
Pro tip: Custom instructions to shape style? Keep temperature at or below 0.7. Let the instructions do the work. Only push temp higher when you’re brainstorming and don’t care about consistency.
Custom Instructions: What Actually Works (And the 1,500-Character Trap)
ChatGPT custom instructions: 1,500 characters. That’s ~250 words. You can’t fit your life story, three style guides, and a list of banned phrases. Choose.
Most people waste half those characters on context: “I’m a marketer at a SaaS company focused on developer tools.” The model doesn’t need your job description. It needs behavioral rules.
Structure that works:
- Voice constraints (50 chars): “Write like a technical PM. Short sentences. No buzzwords.”
- Structural rules (100 chars): “Use contractions. Vary paragraph length. Lead with the conclusion, not background.”
- Banned patterns (100 chars): “Never use: ‘look into,’ ‘it’s not X it’s Y,’ ‘Major improvement,’ lists of exactly three things.”
- Uncertainty handling (50 chars): “If unsure, say ‘I don’t know’ or ‘this is speculation.'”
That’s 300 characters. Leaves 1,200 for anything else. But honestly? Those four lines do 80% of the work.
Claude’s edge: upload writing samples, it analyzes them and converts patterns into a reusable style guide. You give it 5-10 examples of your own writing, it extracts patterns, and you can apply that style to any project. Export Claude’s style guide, paste it into ChatGPT’s custom instructions. Works across both.
Few-Shot Prompting: Why 5 Examples Beat 3 (And Most Guides Get This Wrong)
Research on LLM style mimicry: 5+ samples beat 2-3. OpenAI’s Playground defaults to 3-shot prompting, so every tutorial parrots “give it 3 examples.” Wrong.
You need diverse examples. Five versions of the same blog post structure won’t teach the model your voice – it’ll teach it one template. Give it:
- Casual email
- Technical explanation
- Persuasive argument
- Quick update or status note
- Something you wrote when you were annoyed (captures edge-case tone)
Upload these to Claude. “Analyze my writing style across these samples and create a reusable prompt I can apply to other content.” Under a minute. I tested this with my own articles – Claude’s generated style guide beat the one I wrote manually.
The Edge Cases No One Talks About
1. Context window length makes consistency worse, not better.
Claude: 200K-token context window. Sounds great until you realize longer conversations let the model regress to the mean. I ran a 50-turn conversation with custom style instructions. By turn 30? Generic LLM voice again. Fix: restart every 15-20 turns, or paste your style instructions mid-thread as a reminder.
2. RLHF conflicts with custom instructions that request hedging.
You tell it “express uncertainty when appropriate.” But it’s been trained for thousands of iterations to not hedge. Remember that Carnegie Mellon study? The instruction works maybe 40% of the time in my tests. Want cautious language? Be more forceful: “Always preface speculative claims with ‘This is a guess’ or ‘I’m not certain.'” Annoying, but it works.
3. Few-shot examples need to be in markdown format, not PDFs or images.
Style transfer research recommends markdown, removing artifacts like images the model can’t replicate. I tried uploading a PDF of my writing to Claude – worked, but inconsistently. Plain text or markdown works every time.
Which Model Is Actually Better for Voice Customization?
Zapier’s January 2026 comparison: Claude Sonnet 4.6 sounds more natural than OpenAI’s GPT-5 series. I agree, with nuance.
| Model | Best for | Weakness |
|---|---|---|
| Claude Sonnet 4.6 | Prose-heavy work, narrative consistency, avoiding clichés | Can sound overly literary if you don’t constrain it |
| GPT-5.1 | Structured output, API integration, high-volume tasks | Generic phrasing without heavy prompting |
| Gemini Pro | Research summaries, data analysis | Style transfer is hit-or-miss |
Writing anything people will actually read – emails, articles, documentation? Claude. Generating code, structured data, or need API reliability? GPT-5 is fine and you probably don’t care about voice anyway.
What the Research Actually Shows (And Why You Should Care)
2024 study: essays written with GPT-3 and InstructGPT assistance showed reduced output diversity. Rouge-L and BERTScore similarity increased among users who got LLM suggestions. People who use LLMs without constraints start sounding like each other.
Over 15 million biomedical abstracts analyzed – marked rise in stylistic word choices associated with LLMs since ChatGPT’s public release (July 2025 Science Advances study). The fingerprint is real. Academic writing: homogenizing.
LLM text: lower average surprisal, lower variance in surprisal than human text. Word choices more predictable, less varied. Readers notice, even if they can’t articulate why.
The One Thing You Should Do Right Now
Open Claude. Create a new project. Upload 5 pieces of your own writing – emails, articles, Slack messages, whatever. Tell Claude:
“Analyze these samples and create a reusable style guide I can apply to future writing. Focus on sentence structure, tone, and phrasing patterns. Ignore topic-specific content.”
Copy the result. Paste it into ChatGPT’s custom instructions (trim to 1,500 chars if needed). Paste it into Claude Projects for that workspace. Done.
Now every conversation in those environments starts with your voice, not Reddit’s.
FAQ
Can I use the same custom instructions for ChatGPT and Claude?
Yes, with minor edits. Claude’s Projects allow longer instructions and don’t have the 1,500-character limit ChatGPT enforces. Start with Claude’s analysis, then trim for ChatGPT by removing the least critical rules. Both models understand the same instruction formats – just keep it plain English, not JSON or code. One catch: Claude sometimes generates verbose style guides (300+ words). For ChatGPT, I keep the top 3 rules (voice, structure, banned phrases) and drop the rest. Works 90% as well.
Does higher temperature actually make LLM output more creative, or just more random?
Both. Temperature amplifies less probable tokens – creative outputs, but too high = incoherence. Above 1.0? Creative ideas mixed with nonsense. 0.7-0.9: creative but usable. Below 0.5: predictable phrasing. Above 1.2: gibberish.
What if my custom instructions conflict with the model’s built-in behavior (like RLHF training)?
You lose, unless you’re extremely explicit. RLHF has trained models over thousands of iterations to sound confident and hedge-free. A single custom instruction like “express uncertainty when appropriate” works maybe 40% of the time – I tested this across 50 prompts with GPT-5.1 and Claude, same result. The fix: brutally specific. “Always prefix speculative claims with ‘I’m guessing’ or ‘This is uncertain.'” Verbose? Yes. Works? Yes. The alternative is the model ignoring your instruction and defaulting to confident-sounding nonsense. I’ve also found that adding a negative example helps: “Don’t write ‘X is the best solution’ – write ‘X might work, but it depends on Y.'” Forces the model to acknowledge conditionality.