ChatGPT custom instructions let you set response preferences once, such as your usual audience, voice and format.

They do not pick a fresh angle for every post or guarantee the model will follow each rule.

Where are ChatGPT custom instructions, and what is the character limit?

On web and desktop, open Settings → Personalization, turn on Enable customization, and enter your instructions in Custom Instructions.

On iOS and Android, open Settings → Customize ChatGPT, turn on Enable customization, then enter them in the Custom Instructions field. We checked the exact path in November 2026 against the OpenAI Help Center instructions.

OpenAI says custom instructions "are available on all plans on Web, Desktop, iOS, and Android." The current character limit is 1,500 for Free and Go, and 5,000 for Plus, Pro, Enterprise, Business and Education. Keep a copy before editing, then save and test the preference in a new chat.

What ChatGPT custom instructions examples can I copy for writing?

Keep durable preferences here: audience, voice, format and boundaries. Replace the bracketed details, then trim any rule that does not improve a real draft.

Instruction typeCopy and adaptWhat it usefully specifiesWhat it cannot decide
Voice“Write in plain English for [audience]. Use concrete nouns and verbs. Keep my meaning and characteristic wording.”A default tone and level of editingWhich idea or stance fits this particular post
Phrases to avoid“Avoid ‘I hope this finds you well,’ ‘game-changer’ and ‘delve.’ Keep a phrase if it appears in source text I ask you to quote.”Words you repeatedly delete, with a clear exceptionEvery phrase that may sound wrong in a new context
Format“For a draft, return the draft first, then a short list of factual questions. Use paragraphs under four sentences.”A predictable shape for the same kind of taskA different structure for each distinct piece
Fact boundaries“Use only facts in my prompt or attached sources. Mark missing details as [CHECK]; do not invent names, dates, results or quotes.”How to handle unsupported detailsWhether a source or claim is true without checking it

Specific instructions are easier to check than “make it sound natural.” OpenAI’s personalization guide suggests using custom instructions for recurring preferences and the chat prompt for the task in front of you. For example, keep “use short paragraphs” as a default, then ask for a thread, email or product announcement in the current prompt.

How do custom instructions work with memory, Projects and GPTs?

Treat them as separate layers with different scopes. Custom GPTs are on their way out.

FeatureScopeUse it for
Custom instructionsAccount-level personalization in ordinary chatsStable defaults such as your writing voice
MemoryDetails ChatGPT may remember from chats and other available sourcesContext you want it to retain and update over time
Project instructionsChats inside that ProjectA client, publication or campaign with its own rules and files
GPT instructionsChats with that GPTExisting GPTs only: OpenAI is retiring custom GPTs, and personal accounts can no longer create new ones

Project instructions apply only inside that Project and override global custom instructions there, according to OpenAI’s Projects guide.

OpenAI’s Memory guide says memory does not retain every detail, so do not use it as the only home for a rule you need to state explicitly.

Don't build a new writing setup on a custom GPT: OpenAI's GPT editor guide says it is "planning to retire custom GPTs" in favour of plugins, with standard retirement scheduled for December 11, 2026. New GPT creation is not available on Free, Go, Plus or Pro (checked November 2026).

If one kind of writing needs a different editorial voice, put that rule in its Project's instructions.

Why does ChatGPT still give every post the same shape?

Custom instructions can set a default, and a default does not choose a different approach for each writing job. A rule like “be concise” changes the editing target; it does not tell the model whether this post should be a story, a lesson or a question.

We checked one saved X-post example in the homepage showcase. The task was “why we stopped writing a changelog entry for every small fix.” Each of three plain Claude runs and three plain Codex runs started with the same three words: “We used to.” The voice, Always and Never rules were already present in both arms.

Across three different protocol draws, each model instead used three different openings: “A longer changelog,” “1 entry per,” and “Who reads a changelog” for Claude; “Writing up every,” “1 small fix,” and “Who was helped” for Codex.

That is a small sample, not a general rate or proof that every output will differ.

How can I vary the approach for each piece?

Keep your global writing rules, then give each task an angle, audience, evidence and form. A one-off prompt can do this manually. For a repeated writing job, a playbook can choose a different direction for each piece.

We make Unsame. It returns a short protocol before your model writes: the situation, points to engage with, one approach, a form and your fixed voice rules. The model still writes the words.

You can connect ChatGPT from Settings → Connectors; developer mode may be required. See Connect for the setup details.

Free options such as a saved prompt or Project instructions can also work if you are willing to choose the angle yourself each time. For a repeated X writing job, see the X post workflow.

If your drafts already sound like you but keep using the same opener or structure, start by adding one task-specific direction to the prompt. For a repeated stream of posts, compare a few approaches instead of adding more global voice rules. Try Unsame if you want the protocol selected for each piece.

— Don