You typed "write in my voice, casual but professional" into the prompt box again. You got the same clean, slightly hollow paragraph you always get. It reads fine. It also reads like nobody, which is the problem: you wanted it to read like you.
This happens because "sound like me" is not an instruction ChatGPT can act on. It is a feeling you have about your own writing, and a model can't read a feeling. It can read patterns. So the fix isn't a better adjective in your prompt. It's giving the model something concrete to pattern-match against: your actual words, structured in a way it can reuse.
Here's the method, in order. Do all four steps once and you'll never write "sound more like me" into a prompt box again.
Step 1: Write a Voice Charter (this is the part everyone skips)
Before you touch a prompt, write one sentence, 25 words or fewer, that says who you are, who you talk to, and how you sound. Not your bio. Not your mission statement. A sentence you'd actually say out loud.
We're the coach who asks the question you've been avoiding. Direct, warm, zero corporate speak.
That's a Voice Charter. It does two jobs a vague instruction can't: it names two or three personality traits specific enough to defend in an argument, and it names one thing you never sound like. "Never corporate" tells ChatGPT more than "professional" ever will, because it rules things out instead of leaving them all technically allowed.
Skip this step and every later step just sounds smarter without sounding more like you. This is the C in CAFECITO: Context. It's the bedrock everything else sits on.
Step 2: Stop describing your voice. Show it examples.
"Write casually" means something different to every writer alive. "Write like this" means the same thing to everyone, including the model.
Pull three to five real samples of your best writing: an email you're proud of, a post that actually sounded like you, a paragraph a client complimented. Paste two or three of them directly into your prompt, above the request, labeled as reference:
Here's how I actually write. Match this voice exactly: [paste sample 1] [paste sample 2] Now write [the actual task] in this same voice.
This single change does more than any adjective stack. Models are pattern-matchers first and instruction-followers second. Give them the pattern and the instructions become a formality.
Step 3: Turn it into a Brand Voice JSON so it's reusable
Pasting three samples into every prompt works, but it's slow, and you'll skip it on a Tuesday when you're rushing. The fix is building it once as a structured reference you paste at the top of any chat: your Voice Charter, your do/don't word list, two or three example lines, and your channel notes (how you sound on email versus social versus a sales page).
That structured file is what CAFECITO calls the Brand Voice JSON, the Cohesion pillar. It's not magic formatting. It's the difference between re-explaining your voice from scratch every session and having a file any AI tool can read in one paste. Build yours with the free Voice JSON Builder: answer a handful of questions, get a structured file back, paste it at the top of every new chat.
Step 4: Build a Red-Pen loop so it gets better, not worse
Here's what most people miss: the first output from a well-built voice JSON is good, not perfect. The mistake is either accepting the near-miss because rewriting feels slower than starting over, or scrapping the whole approach because "AI still doesn't get it."
Neither is right. What you want is a Red-Pen habit: when you edit AI output, don't just fix it and move on. Note why you changed it. "Too formal, cut the throat-clearing intro" or "good bones, wrong CTA verb." After ten or fifteen edits, you'll see the same two or three notes repeating. Those become permanent additions to your voice file. That's the Feedback pillar, and it's what separates a voice system that improves from one that plateaus at "pretty close."
Why "just add adjectives" doesn't work
If you've tried prompts like "write in a friendly, confident, approachable tone" and gotten generic output anyway, you've hit the ceiling of adjective-only prompting. Every brand claims to be friendly and confident. Those words don't discriminate between you and your competitor, so the model has nothing to lean on except its own defaults, which is exactly the flat, safe, everyone-brand voice you're trying to escape.
Specificity is the whole trick. "Direct, warm, never corporate" beats "professional yet approachable" because it's a real editorial choice, not a compliment. A model can act on a choice. It can't act on a compliment.
Putting it all together
The order matters: charter first (so you know what you're aiming for), examples second (so the model has a pattern), structured file third (so it's reusable), feedback loop fourth (so it improves). Most people try to skip straight to a clever prompt and wonder why the output still sounds like a stranger wrote it. The prompt was never the problem. The missing voice system was.
If you want the full framework behind this, including the Tone & Mood Matrix and channel-specific prompt templates, that's what the CAFECITO Miami Method walks through end to end. And if you're not sure how off your current AI output actually is, the free voice check will tell you in about three minutes.
Start here
Don't overthink the first move. Write your Voice Charter today, in one sitting, out loud if it helps. Then turn it into a structured file you can paste anywhere.
Build yours now with the Voice JSON Builder. It's free, it takes about ten minutes, and it's the single most useful thing you can do before your next ChatGPT session.