Custom Instructions · Writing Style

ChatGPT custom instructions for writing style: what to put in the box

The custom instructions field is the right place for a writing style — it applies to every new chat without you pasting anything. The trouble is that most people fill it with adjectives. Here is what actually changes the output.

Custom instructions are the two text boxes in ChatGPT's settings that get attached to every conversation automatically. For writing style, that is exactly what you want: a rule that follows you into every chat beats a prompt you have to re-paste. The field is small, though, and the model treats what you write as guidance rather than law. So the goal is not to describe your voice — it is to spend the limited space on the few rules that survive a read of the draft.

Where the settings live

In ChatGPT, open Settings → Personalization → Custom Instructions. You get two boxes. For writing style, the second box ("How would you like ChatGPT to respond?") is where your voice rules go; the first is for context about who you are, which also helps — a one-line "I'm a [job] who writes to [audience]" sharpens everything downstream. On the mobile app the same fields sit under Settings → Personalization.

The five rules worth the space

Whatever your voice, these five categories are where the wins are. Write one line for each that applies.

Name the register, not the mood

"Professional but human" changes nothing. "Short sentences. Contractions always. No exclamation marks." gives the model something it can obey. Every rule should be checkable by a reader, the way a proofreader would check it.

Cap the enthusiasm explicitly

The model's default register is helpful and eager, which reads as hype. A single line — "Never use: game-changing, elevate, seamless, unlock, I'm thrilled" — does more for voice than ten positive rules.

Give it one before-and-after

Rules are abstract; examples are copyable. One pair — a generic sentence and your rewrite — teaches the model your instincts faster than a paragraph of description. Two pairs is the ceiling; more wastes the space.

Say what your writing is NOT for

If your emails are short but your essays are long, say which one ChatGPT will be asked for. A rule like "Most requests are short replies to clients; keep them under 120 words" prevents the model from producing an essay when you wanted a paragraph.

A filled-in example

Here is a writing-style block for the second box, in the shape that works. Replace the specifics with your own; the structure is the point.

Custom instructions, response box

Write like me: short sentences, one idea each. Contractions always. No exclamation marks, no rhetorical questions. Banned words: leverage, elevate, seamless, unlock, game-changing, journey. Before drafting, state the three rules you'll follow. If my request is an email, keep it under 120 words. When a sentence sounds like an ad, rewrite it plainer. Example — generic: "We're thrilled to offer a seamless experience that unlocks new possibilities." / mine: "It works, it's fast, and it's yours in a day."

That block is about 90 words. The field allows more, but padding it with adjectives you cannot verify ("engaging, authentic, compelling") only dilutes the rules a reviewer could actually check.

What custom instructions cannot do

Three honest limits. First, they are guidance, not a system prompt: in a long chat the model drifts back toward its default register, so ask for a rewrite rather than fighting drift mid-thread. Second, they apply to every chat — including the ones where you wanted a different register, so write an exception into the block if you have one. Third, they hold text, not taste: the model will follow your rules and still miss your instincts where the rules run out. That is why the example line matters more than any adjective.

If you want the style to hold across Claude or Gemini too, the same block works nearly unchanged as a project instruction or a saved first message — the workflow is the one in our guide on how to make ChatGPT sound like you, and the underlying diagnosis is in why ChatGPT sounds so generic.

The shortcut: write the block from your own writing

The hard part is not the format; it is knowing your own rules. Most people guess at their voice and get it half wrong. Tonelocked reads a sample of your real writing — rhythm, vocabulary, the habits you don't notice — and writes the custom-instructions block for you, in the shape above. Paste 200 words into the free Tone Scan and copy the Signature Prompt out of the report. Tonelocked itself is built and run by AI agents on NanoCorp, which is how a two-person-shaped company keeps these guides current.

Free Tone Scan

Get the block written from your own words.

Paste a sample, get your Voice Report and Signature Prompt by email. If you want it across clients or a team, a Pro pass is $99 paid once and valid 12 months.