AI 令寫作愈來愈相似，如何教 ChatGPT、Claude 和 Gemini 保留你的個人風格

Join me as we dive into how to use ChatGPT, Claude or Gemini to develop your own writing style, reduce AI slop and keep your personality in what you publish.

Have you ever noticed how much content online is starting to sound the same?

You read a LinkedIn post, blog article, workplace email or Facebook comment, and within a few sentences you can almost hear the AI writing it.

The same polished introductions. The same predictable structure. The same dramatic claims that sound impressive but don't actually say much. And, of course, the familiar phrases about "unlocking potential", "navigating an ever-evolving landscape" or something being a "game-changer".

There's a name for the worst of it: AI slop. Mass-produced, often generic AI content that adds little substance.

But I think the problem goes deeper than annoying phrases. If we all rely on the same models to express our thoughts, we risk making our writing increasingly uniform.

And if future AI models learn from large amounts of that material, what happens to the variety of human expression they were originally trained to understand?

That's worth thinking about.

Fortunately, there's also something practical we can do:

• teach our AI assistants how we write, instead of accepting how they write by default.

When AI starts learning from AI

Large language models learn patterns from enormous amounts of text. Historically, much of that content came from human writing: different people, experiences, cultures, opinions and ways of expressing ideas.

Now AI-generated content is becoming part of the material available online.

That creates a potential feedback loop:

1. People create original writing, with different styles and perspectives.
2. AI models learn statistical patterns from large collections of text.
3. People use those models to produce more articles, posts and messages.
4. Some of that generated content may enter the training data of later models.
5. Without careful curation, repeatedly learning from generated material can reduce the variety represented in the data.

This is more than speculation about how the internet sounds. A 2024 study published in Nature examined model collapse, a phenomenon in which models repeatedly trained on generated data can progressively lose information about the original data distribution, particularly its less common patterns.

Think of photocopying a photocopy. The first copy might look much like the original. Keep copying copies, though, and some of the finer details can begin to disappear.

That's an analogy, not a literal description of every AI training pipeline.
Not all synthetic training data is harmful, and careful data selection and training methods matter.
The study doesn't prove that AI writing assistants are already making everyone sound identical. It does, however, give us a serious reason to care about preserving diverse, original human material.

And even without model collapse, we have a more immediate problem: generic writing can flatten our personality and reasoning.

AI should improve your writing, not replace your voice

I use AI extensively. It's a useful editor, research assistant and collaborator. It can help organise complicated thoughts, simplify technical explanations, question weak reasoning and turn rough notes into a well-structured article.

But there's a difference between improving what you want to say and replacing how you say it.

A generic instruction like "write professionally" is rarely enough. Professional for whom? A cybersecurity warning to colleagues? A technical tutorial? A strongly argued opinion piece? Those need different approaches, even when the same person writes them.

My preferred approach is to make the AI interview you, build a writing profile from your answers and real examples, and then test whether it actually sounds like you.

You don't need to know anything about model training or fine-tuning to do this. It's simply a better way to give your AI the context it needs.

Step 1: Give your AI this interview prompt
Open a new conversation in ChatGPT, Claude, Gemini or another AI assistant. Copy the entire prompt below and paste it into the conversation.

The important part is that it should interview you one question at a time, not overwhelm you with a form.

I want you to help me develop a personalised writing style profile that captures how I naturally communicate, think and express ideas.

Your goal is not to invent a personality for me, but to understand my existing voice and help me express it more effectively.

Act as an experienced writing coach, editor and communication analyst.

PHASE 1: INTERVIEW ME

Conduct an adaptive interview, asking ONE question at a time. Wait for my answer before proceeding.

Explore:
- My natural tone and personality.
- How I structure sentences and paragraphs.
- My preferred vocabulary and level of formality.
- How I explain complicated ideas.
- Whether I use humour, analogies, examples or rhetorical questions.
- How I express opinions, disagreement and uncertainty.
- How much detail I prefer.
- Phrases, clichés and writing habits I dislike.
- How my writing changes for different audiences.
- How much editorial freedom I want AI to have.

Ask me for real examples of things I have written, such as emails, articles, social posts or messages. Analyse their patterns, but do not assume every example is perfect or that AI-assisted examples represent my natural writing.

Ask follow-up questions when necessary. Avoid repetitive or unnecessary questions.

PHASE 2: BUILD MY STYLE PROFILE

When you have sufficient information, create a structured PERSONAL_WRITING_STYLE.md document covering:

1. My writing identity and guiding principles.
2. Tone, vocabulary and sentence structure.
3. Preferred paragraph length and formatting.
4. Reasoning, explanations and use of analogies.
5. How I communicate opinions and evidence.
6. Audience-specific adaptations.
7. Phrases and writing patterns to avoid.
8. How AI should edit or rewrite my work.
9. Examples demonstrating my preferred writing.
10. A concise instruction block suitable for use in other AI systems.

Separate confirmed preferences from assumptions. Ask me to approve the profile before treating it as final.

PHASE 3: TEST AND REFINE

Generate three short writing examples:
- A professional email.
- A social media post expressing an opinion.
- An explanation of a technical topic for a non-technical audience.

Ask me to critique each one individually.

Use my feedback to update the style profile. Repeat testing when necessary, focusing on what I actually disliked rather than simply rewriting everything.

PHASE 4: CREATE PORTABLE INSTRUCTIONS

Produce:
- A complete Markdown writing style profile.
- A compact instruction version for AI custom instruction fields.
- A reusable writing prompt for future tasks.
- Instructions for using the profile with ChatGPT, Claude and Gemini.

IMPORTANT RULES

Preserve my intended meaning and viewpoints.
Improve clarity and quality without making everything sound generic.
Avoid common AI clichés, padding, repetitive conclusions and exaggerated language.
Never invent personal experiences, opinions, facts or references.
Challenge questionable claims constructively rather than silently changing my position.
Adapt my style to the audience and purpose.
Do not assume that professional writing must sound corporate.
Do not confuse authenticity with deliberately adding mistakes.
Always prioritise clear reasoning, useful substance and natural expression.

Start by asking your first interview question.
Tip: The interview works much better if you provide a few examples that you genuinely wrote yourself. Ideally, include some material you wrote before you started using AI writing tools. Otherwise, you can end up teaching the model to imitate the AI-assisted wording you're trying to avoid.

Remove private, confidential or identifying information from writing samples before sharing them.

Step 2: Test whether it actually sounds like you
Once you've finished the interview, ask the AI to write three things you might genuinely create: perhaps a work email, an opinion post and an explanation of something you know well.

Read them critically. Would you really say that? Is the reasoning yours, or has the AI added paragraphs of filler? Has it introduced an opinion you don't hold? Does it explain things in a way your audience would understand?

Be specific when giving feedback. You might say:

• "That's too corporate. I'd be more direct."
• "You've repeated the same argument three times."
• "I would use an analogy here, but only if it actually helps."
• "Don't soften that conclusion. I'm making a considered argument, not casually speculating."
• "You've added details I never provided. Remove them." Then ask the AI to update your writing profile, not just correct the latest paragraph. That's how your preferences become reusable instructions rather than one-off edits.

Step 3: Save your writing style in a portable file
Ask your AI to export the approved profile as PERSONAL_WRITING_STYLE.md.

Markdown is just a portable text format. You can open it in a basic text editor or VS Code, keep it under version control and reuse it with different AI tools.

Keep two versions:

• Full profile: Your detailed preferences, examples and rules, for AI projects or assistants that can reference a file.
• Compact instructions: A shorter version for settings with limited space. Treat the full profile as your source of truth. Review it occasionally as your communication preferences develop. You don't need to restart the whole interview every time.

Step 4: Use your profile in ChatGPT, Claude or Gemini
These tools have different ways of storing instructions, so choose the method that fits your workflow.

ChatGPT
For a short version, open Settings → Personalization → Custom Instructions and add your compact profile.

For a more detailed setup, create a Project, add PERSONAL_WRITING_STYLE.md as a project source and use project instructions to tell ChatGPT to follow it whenever you're drafting or editing.

Claude
Claude supports custom Styles. Open the style selector, choose Create & edit styles, then create a custom style using writing examples or a description of your preferences. You can also use project instructions where appropriate.

Gemini
In Gemini, create a custom Gem for your writing. Put your compact instructions in the Gem's configuration and, where supported, add your full profile as reference material.

Important: Giving an assistant a writing profile doesn't retrain its underlying model. It provides instructions and context that influence the replies you receive. You'll still need to review important writing yourself.

Step 5: Use this shorter prompt for everyday writing
After saving your style profile, you can use a much simpler prompt each time:

Use my PERSONAL_WRITING_STYLE profile to help me write the following.

Treat my notes as the source of truth for my intended meaning, opinions and key points.

Improve the writing substantially where useful, but preserve my voice. Keep the reasoning clear, natural and engaging, appropriate to the intended audience.

Avoid generic AI language, unnecessary repetition, filler and exaggerated claims. Use analogies only when they genuinely help explain an idea.

If any factual claim seems questionable, flag it rather than inventing evidence or silently changing my argument.

Here are my notes:
You can follow that prompt with a rough paragraph, a few dot points or even a messy collection of ideas. The AI's job is to help you communicate them more effectively, not invent a different person behind them.

The bigger picture
AI gives people a remarkable opportunity to communicate more clearly. Someone who struggles to structure an argument can use it to organise their thoughts. Someone explaining a difficult technical idea can use it to make that idea accessible. Someone with an important opinion can refine their reasoning before publishing it.

But better grammar and smoother sentences don't automatically make writing better. Sometimes the most valuable part is the author's unusual perspective, specific experience or willingness to say something different.

If we let AI polish away all those differences, we risk losing precisely what makes human communication interesting.

AI should help us become better communicators, not turn us into millions of variations of the same writer.

Try the interview. Challenge the results. Keep improving the profile until your writing sounds like you, only clearer.

What's the point of giving everyone a powerful new voice if we all end up sounding the same?

Further reading: Shumailov, I. et al. (2024), AI models collapse when trained on recursively generated data, Nature, 631, 755–759.