I’ve got an amazing set of custom instructions to share today.
These are not theoretical.
I’ve been editing, refining, and actively using these instructions inside ChatGPT for over a year. Across articles. Emails. Longform writing. Notes that actually mattered.
They exist because I got tired of AI writing sounding like AI. Also, just the existence of em-dashes (—) really annoy me.
So I stopped trying to “prompt better” and started forcing the model to behave differently.
This article explains how and why that works.
The real problem with AI writing
AI does not sound robotic because it lacks intelligence.
It sounds robotic because it lacks constraints.
Human writing is irregular.
Sentence length varies.
Thoughts interrupt themselves.
Emotion leaks between ideas.
Most people strip all of that out when they prompt.
They ask for clarity, structure, professionalism.
What they really ask for is sameness.
So the model gives them the safest possible average voice.
Readable. Polished. Forgettable.
What these instructions actually do
The attached instructions force the AI to behave less like a system and more like a narrator.
I use them to do four things consistently.
First, they lock in a persona.
The model has to maintain a stable emotional bias and point of view across the entire piece.
Second, they enforce rhythmic variation.
Sentence length shifts.
Paragraph size changes.
Pacing becomes intentional instead of uniform.
Third, they allow imperfection.
Micro hesitations.
Slightly awkward phrasing.
Incomplete thoughts that feel real.
Fourth, they anchor language in sensation and memory.
The writing stays connected to texture, movement, emotion, and lived experience.
These rules do not make the writing flashy.
They make it believable.
How to use the instructions (do this exactly)
Upload the instructions as a .pdf or .txt file to your Project or Custom GPT.
Name it clearly.
For example: Writing Instructions.pdf.
Then, inside your project instructions or system instructions, add this line verbatim:
Writing Instructions.pdf is to be used when writing books, articles, anything media-related, and the like. Always refer to it for crafting these types of works: books, stories, articles, etc.
This step matters more than people think.
Uploading the document gives the model a persistent reference.
Explicitly pointing to it tells the model when to apply it.
Do not paste the full text into every prompt.
Do not restate the rules each time.
Set the rule once.
Let it run.
If the output feels wrong, adjust the document.
Do not argue with the model sentence by sentence.
That is how I’ve been using this for over a year.
One thing most people miss
If you have any examples of your own writing, upload those too.
Emails.
Drafts.
Old blog posts.
Notes you never published.
Polished samples are optional.
Authentic samples are not.
Instructions define how the AI should behave.
Your writing defines who it should sound like.
Without samples, the model converges toward an average voice.
With samples, it starts orbiting yours.
This single step fixes most complaints people have about AI tone.
What these instructions are not
They are not a shortcut to creativity.
They will not make bad ideas interesting.
They will not replace judgment.
They are for people who already know what they want to say, but want help saying it faster and cleaner.
If you rely on AI to think for you, this will not help.
If you use AI to express your thinking, it will.
Final note
AI does not need to sound human.
It needs to sound specific.
Specificity is what readers trust.
Specificity comes from rules and reference, not clever wording.
Upload the instructions.
Reference them properly.
Upload your writing.
Then let the model meet you halfway.
