Steno / ChatGPT

How to make ChatGPT sound human

Two problems, and most people only solve the second. Knowing what to write is one. Knowing where to put it so you never write it again is the other.

Updated 6 August 2026 · About seven minutes

Why "write like a human" does nothing

The instruction fails because the model has no idea which of its own habits read as machine written. It has been trained on human text, so from the inside, everything it produces is already human writing. Ask for a human tone and you get the same output with a friendlier opener.

What does work is naming the pattern. Do not open with a sentence about the modern world is actionable. Sound natural is not. Every instruction below is specific for that reason.

Where the instruction goes, tool by tool

ChatGPT

Settings, then Personalisation, then Custom instructions. There are two boxes. The first is about you, the second is about how ChatGPT should respond. Style rules go in the second box. Whatever you put there applies to every new conversation until you edit it, which is the entire point: you write it once instead of at the top of every chat.

The box has a character limit, so terse rules beat explanations. Write no em dashes, no emoji, no bold on the first words of bullets rather than a paragraph about why.

Claude

Two options, and they behave differently. Projects carry instructions for everything inside that project, which suits a body of work with a consistent voice. Styles apply across conversations and are the closer match to ChatGPT custom instructions. If you write in more than one register, make a style per register rather than one compromise.

Gemini

Saved info holds standing preferences. Gems are the equivalent of a project, useful when a particular kind of writing needs its own rules.

Cursor, Copilot and coding agents

These read a rules file from the repository. The style rules matter more than people expect, because the prose these tools generate ends up in commit messages, pull request descriptions, changelogs and documentation, and that prose is what your colleagues actually read.

A block you can paste right now

This is a short version. It is genuinely useful on its own and it is not the full product, which is longer and handles the cases where the rules conflict with each other.

Write as a person would. Never state a fact, number or citation
you cannot support.

Do not use: inflated significance (stands as a testament,
pivotal, evolving landscape), participial padding (highlighting,
ensuring, fostering, showcasing), promotional adjectives
(nestled, boasts, vibrant, renowned), vague attribution
(experts argue, studies suggest), negative parallelism (not
only X but Y), signposting (let us dive in, here is what you
need to know), or filler (in order to, it is worth noting that).

Use is and has. Do not write serves as, represents or features
when is or has will do.

Do not force ideas into groups of three. Two is fine.

Vary sentence and paragraph length. Uniform rhythm is the
loudest tell.

No emoji. No em dashes. No boldface applied by rule.

Bullets only for real lists, never for prose.

Have a view where the genre allows it.

Paste it into the second custom instructions box, start a new chat, and ask for something you have asked for before. The difference shows up in the first paragraph.

What to expect, honestly

Three things are worth knowing before you judge the result.

The rules decay over a long conversation. Instructions sit at the top of the context and compete with everything said since. After twenty exchanges the openers start creeping back. A short reminder mid conversation fixes it.

The model overcorrects at first. Told to avoid ornament, it sometimes produces prose that is clipped to the point of rudeness. Adding one line about keeping a warm tone usually settles it.

And instructions cannot repair text that already exists. Anything you wrote before pasting them stays exactly as it was. That is a different job, an editing pass over finished text, and it needs its own set of rules: rewrite the prose, keep every claim, invent nothing.

The words are not the problem

Every list of banned phrases goes stale, because the words rotate and the habits do not. A model that cannot say delve says explore in exactly the same sentence shape. The shape is the tell.

Which is why the useful instructions are about structure: sentence length variation, no forced triplets, no participial clause bolted onto the end, no conclusion stated on the reader's behalf. The full list of twelve habits is here, with a before and after for each.