Verses Over Variables: How To

How to actually train ChatGPT to write like you (without losing your mind)

OpenAI released GPT-5.1 on November 13th, and buried in the model improvements is something that actually matters: it follows instructions now. I don't mean "sort of follows instructions if you remind it three times." I mean you can tell it "never use em dashes" and it won't use em dashes, finally. This sounds basic. It's not. Previous versions of ChatGPT treated custom instructions like polite suggestions. You'd set up elaborate rules, then watch the model ignore half of them in the next conversation. GPT-5.1 changed that. Better instruction-following means the personalization system actually works reliably now. ChatGPT effectively comes with a control panel: a three-layer personalization system sitting in ChatGPT's settings that most people never configure properly. Custom instructions, Projects, and prompt patterns that actually stick.

Let me show you how this works.

Why most custom instructions fail

You've probably tried custom instructions. You might have given up after ChatGPT ignored your specifications and wrote in generic corporate-speak anyway. The problem isn't that custom instructions don't work. The problem is that most people write them like polite suggestions instead of hard rules. I see this constantly. People write things like "I prefer a conversational tone" or "Keep responses concise." These are vibes, not instructions. ChatGPT doesn't know what "conversational" means to you specifically, and "concise" could mean anything.

What actually works: specific, testable rules. Instead of "conversational tone," I tell ChatGPT: "Write in short paragraphs, 2-4 sentences each. Use second person when teaching. Never start sentences with 'Moreover,' 'Furthermore,' or 'Additionally.'" Instead of "keep responses concise," I specify: "Default to 200-300 word responses unless I ask for more detail." This precision matters because ChatGPT is a prediction machine, not a mind reader. You need rules a machine can follow.

Step 1: Building your style blueprint

Before you configure anything in Settings, you need to know what your voice actually is. This is the foundation everything else builds on. Gather 5-7 pieces you've written that sound distinctly like you. Mix formats if you have them: articles, emails, social posts. Feed them to ChatGPT in a new chat and ask:

"Analyze my writing style across these samples. Extract specific, testable rules I can use as custom instructions. Focus on sentence structure, paragraph length, word choice, what I avoid, and how I transition between ideas. Give me rules in the format 'Always...' or 'Never...' so I can verify them."

ChatGPT's first pass will be too general. It'll say things like "conversational and engaging" or "uses vivid language." Push back: "Those descriptions are too vague. Give me 10 specific rules I can test. Include things like 'Never start sentences with X' or 'Keep paragraphs under Y words' or 'Use analogies when explaining technical concepts.' I need instructions a machine can follow." You'll need to iterate 2-3 times before you get something usable. Keep pushing for specificity. Once you have concrete rules, you're ready to set up the control panel.

Step 2: The Three Layer Control Panel

I structure my ChatGPT setup like nested contexts. Each layer adds specificity without overriding the previous one. (FYI, these layers aren’t just for ChatGPT, you can use this system in most other LLMs.)

Layer 1: Your Global Voice

Head to Settings > Personalization, and

• Choose the Base Style and Tone: these are preset personalities. Choose the one that is closest to your natural voice. Would you be surprised that I chose Nerdy?

• Fill out your Custom Instructions: use your style blueprint to explain your voice.

• Fill out About You: write context that helps ChatGPT understand your work, who you are, and who your audience is.

Once you save these settings, every new chat starts with these rules. You never have to re-explain yourself. That is, until you change careers.

Layer 2: Project Specific Context

Projects are ChatGPT's workspace feature. Each gets its own file uploads, custom instructions, and isolated chat history. Think of them as specialized versions of ChatGPT for different types of work. I have four main projects: Content, Clients, Quick Research, and Sandbox.

When I work inside my Content Project, I have specific custom instructions that I adjust by content type and audience. I have also uploaded 20 examples of my work, showing my actual patterns for this content. (The catch: you need good examples. I spent an afternoon selecting my best work, not just recent work. In projects, quality matters more than quantity.) The difference with using a Project and just a chat window is stark, because ChatGPT can reference my past conversations, and it matches my rhythm.

Layer 3: Prompt Patterns

Even with global voice and Projects configured, I use specific prompts to keep output on track. These are tactical patterns I've developed through trial and error.

• Lock in the voice: When working on something complex or outside my normal patterns, I remind ChatGPT explicitly: "Use my saved style guide. Reference the newsletter from March 15th. Match that level of specificity and opening structure."

• Self-Critique: After first draft: "List three specific ways this draft doesn't match my voice. Then rewrite fixing those issues." ChatGPT will flag things like "I used passive voice in paragraph 3" or "The opening is too formal."

• Two-pass generation: For anything over 500 words, I split structure from style. "First: outline this piece with section breaks and key points. Show me the outline before writing." Then after approval: "Now write the full piece using my voice and the uploaded examples."

• Comparison versions: When calibrating a new Project or testing voice changes: "Write this in two versions. Version A: generic ChatGPT voice. Version B: my saved voice. Put them side by side so I can see the differences."

• Concrete anchoring: When starting something new: "Reference the newsletter from October 3rd where I covered Chinese digital painters. Match that balance of technical detail and accessibility. Use a similar structure."

What this actually gets you

Using this three layer system, I can tell you what has changed and what stayed the same. What improved: volume without voice loss. I can draft 8-10 pieces daily and they all sound like me. The model remembers my patterns, references my past work, and applies my rules consistently. First drafts went from 30% usable to 70% usable. That is real time saved. What didn't change: I still make all judgment calls. ChatGPT can't decide what's interesting, what angle to take, or what deserves coverage. It can't feel out whether an argument is compelling or whether an example resonates. That's still my job.

Use the model for scale and consistency. I handle taste and direction.

This system is useful but not magical. I'm not suddenly writing twice as fast. I'm writing at the same speed with less friction. The AI stops fighting me on basics like sentence structure, so I can focus on harder problems like "is this actually interesting?"

The set up tax

This system only works if you do the initial configuration properly. That takes time. If you skip building your style guide, your custom instructions will be too vague. ChatGPT will default to generic. If you don't upload examples to Projects, the model won't have enough context to match your voice. If you don't use prompt-level tactics, you'll still get drift. The setup is about 30 minutes for your style blueprint and 10 minutes per Project setup. But once it's done, it's done. I haven't touched my settings in two months. The system just runs.

If you take nothing else from this: write your custom instructions as testable rules, not vibes. For writers, creatives, and anyone who works in public, this 3 layered system has a specific payoff: you can scale your volume without flattening your personality.

Share your results in the comments and let me know what worked.

— Lauren Eve Cantor

thanks for reading!

if someone sent this to you or you haven’t done so yet, please sign up so you never miss an issue.

we’ve also started publishing more frequently on LinkedIn, and you can follow us here

if you’d like to chat further about opportunities or interest in AI, please feel free to reply.

if you have any feedback or want to engage with any of the topics discussed in Verses Over Variables, please feel free to reply to this email.

banner images created with Midjourney.