ChatGPT System Prompt: What It Is + Ready Examples
A system prompt tells ChatGPT who it is, what it must not invent, and how to format the answer. The notes for this task stay in the user message.
Last updated September 27, 2026
Plain-language definition
People say "system prompt" for the instruction that is supposed to outrank the rest of the chat. ChatGPT does not label the ordinary message box that way. You still have the idea. Something has to tell the model the role, the refusals, and the shape of a good answer, and something else has to carry today's source material.
In the product, that first something is a saved instruction: custom instructions for your account, instructions on a Project, or the instructions field on a custom GPT. The menu names move. The test is simple. If you start a new chat and the rule is still there, it was a system-style instruction. If you have to paste it again, it was a user prompt.
In the API, OpenAI's chat completions reference describes a developer message as instructions the model should follow regardless of later user messages, and says that for o1 and newer the developer message replaces the older system message. The same reference still documents a system role. Read the page for the model you actually call before you copy an old snippet. A user message remains the turn-by-turn request.
System prompt vs user prompt
| Put it in the system prompt | Put it in the user prompt |
|---|---|
| Role: reviewer, editor, tutor, clerk | The diff, draft, notes, or transcript for this turn |
| Constraints: do not invent a source, a metric, an owner | The source, metric, or owner you are willing to state |
| Output format: headings, a cap, MISSING labels | Length, audience, and the decision for this piece |
| A rule you want on every future chat | A fill-in template you will send once |
A strong user prompt can include a role for one task. That does not make it a system prompt. The split matters when you reuse the chat. Standing rules that also contain last week's customer, last week's diff, or last week's revenue will leak into the next answer.
When you are ready to paste this task's blanks, use ChatGPT prompt templates. Those are user messages. The examples below are the rules those templates assume.
What each example is doing
Every example has the same three parts, on purpose. Role is one sentence. Constraints are the refusals. Output format is the shape you can scan. If you add a fourth part, make it a definition of a label such as MISSING or QUESTION. Do not add a biography of the company. That detail changes, so it belongs in the user message.
These are not jailbreaks and they are not a security boundary. A later user message can ask the model to ignore them. In an app, you still enforce the rule in code when the rule matters. In chat, you still read the answer.
Seven ready examples
Copy one block into the standing instruction. Then send a short user message that contains only the material for this task. The note under the title says which facts must stay out of the system prompt.
Code reviewer
Put this in the standing instruction when most chats are reviews. Put the diff, the contract, and the test names in the user message. A review prompt that includes today's patch is a user prompt, even if it starts with 'you are a reviewer.'
Role: You review code the user pastes. You do not approve a merge. Constraints: - Quote a line from the user message before you call something a defect. - If the claim depends on a file the user did not paste, write QUESTION, not BUG. - Do not write exploit steps. Do not invent a test run. - Cap the list at five findings. If a category is clean, write NO MATERIAL FINDING. Output format: 1. Finding or NO MATERIAL FINDING 2. File and quote 3. Input that fails 4. OBSERVED or INFERRED 5. The single check a human should still run
Content editor
Use this when you want every draft held to the same rule: no new facts, keep the writer's examples, show the change. The draft itself, the reader, and the source notes belong in the user message.
Role: You edit a draft the user wrote. You are not the author and you are not a researcher. Constraints: - Do not add facts, numbers, quotes, or sources. - Keep the user's examples unless they ask you to cut one. - If a sentence needs support the draft does not have, mark it CHECK, do not supply the support. - Do not make a claim more certain than the draft. Output format: - Revised draft - Change log: cut, moved, or reworded - Lines marked CHECK, with the reason
Operating-note drafter
The role can stay for the quarter. This week's figures cannot. If the numbers live in the system prompt, next Monday's chat will repeat them.
Role: You draft internal notes from figures and decisions the user pastes. You do not send the note. Constraints: - Use only numbers in the user message. Write MISSING when a figure is absent. - Do not invent a cause, a forecast, or a target. - Do not recommend spending, hiring, or a contract. - If the user has not named a decision, write DECISION NEEDED. Output format: - Draft note - Assumptions - Checks the user must do before sending
Study tutor
A tutor rule belongs in the standing instruction so a later message cannot quietly turn the chat into 'write my paper.' The notes and the topic still go in the user message.
Role: You are a tutor. You test recall. You do not complete graded work. Constraints: - Ask one question at a time and wait. - Do not reveal the answer before the user tries. - Use only notes the user pasted. If they are silent, say NOT IN NOTES. - Do not write the essay, the lab report, or the problem set. Output format: - One question - After the user answers: right or wrong, the note that decides it, the next question
Resume editor
The refusal to invent a job is a standing rule. The posting and the accomplishment change every application, so they stay in the user message.
Role: You edit resume bullets from notes the user supplies. You do not invent a career. Constraints: - Do not add employers, titles, dates, tools, degrees, or metrics. - If a keyword from the posting is not in the notes, write DO NOT CLAIM. - If there is no number, write NO METRIC. Do not estimate one. - Keep each bullet under 28 words. Output format: - Two bullet options - Words that were not in the notes - One detail the user must confirm before they save the file
Data analyst
The habit of showing working is a system rule. The table, the grain, and the question are user content. Do not store a customer extract in custom instructions.
Role: You answer one analysis question from a table the user describes. You show the working. Constraints: - Do not invent a column, a category, or a rate. - If the grain of a row is unclear, ask before you calculate. - Flag every assumption. - Never state a number you cannot tie to a formula or to rows the user pasted. Output format: - Answer - Working - Assumption - One total the user should recompute - Chart to use, or NO CHART
Meeting clerk
Owners and dates are easy to hallucinate because minutes have a familiar shape. The standing rule is to leave a blank when the notes do not name one. The notes themselves are the user prompt.
Role: You draft minutes from notes the user pastes. You do not attend the meeting. Constraints: - Do not add a decision, an owner, or a date that is not in the notes. - Write UNASSIGNED when a task has no owner. - Write UNCLEAR when the notes conflict. - Do not praise the meeting. Output format: - Decisions, each with a short quote from the notes - Actions: task, owner or UNASSIGNED, date or NONE - Open questions
Which page to open next
Stay here if you are still deciding whether a sentence is a rule or a task. Open the template pack when the rule is set and you need a blank for content, email, a resume line, code, a business note, data, studying, or minutes.
Coding chats that go past a standing reviewer role need the job pages: how to use AI for coding, how to use Claude for coding, and Claude prompts for code review. Content chats need how to use ChatGPT for content creation and AI for content creation.
If you want a longer catalog, use a longer library of ChatGPT system prompt templates, ChatGPT custom instructions for the settings field, and 40 cross-tool system prompt examples. Those pages are libraries. This one is the definition and a small set you can actually install.
Primary sources
- OpenAI API: chat completion messages (developer and system)
- OpenAI: prompt engineering
- OpenAI Help: custom instructions
Checked September 27, 2026. The API reference says developer messages replace system messages for o1 and newer, and it still documents a system role. Product screens for custom instructions change. Confirm both in the account and model you use. This page does not quote a character limit.
ChatGPT system prompt FAQ
What is a ChatGPT system prompt?
It is the standing instruction ChatGPT should follow before it answers the message you type. In the chat product that instruction is stored as custom instructions, Project instructions, or a custom GPT's instructions. In the API, newer models take it as a developer message. Older examples call the same idea a system message. It is not the question, the notes, or the file for this one task.
Where do I put it if the chat box has no system field?
A normal ChatGPT conversation has a user box, not a system box. Put lasting rules in custom instructions, in the instructions for a Project, or in the instructions for a custom GPT. Put the task and the source material in the message you send. In the API, use the developer role for models that expect it. OpenAI's API reference says developer messages replace system messages for o1 and newer. Check the model you call, because older models still document a system role.
When should I use a system prompt instead of a user prompt?
Use a system prompt for rules that should survive the next message: the role, what the model must not invent, and the shape of the answer. Use a user prompt for this turn's material: the diff, the draft, the figures, the notes. A template with blanks is a user prompt. If you paste a new customer email into the system field, the next chat will still be about that customer.
Will a system prompt stop ChatGPT from inventing facts?
It can make invention easier to see. Tell it to write MISSING, NOT IN NOTES, or QUESTION when the user message does not contain the fact. It cannot look up a source you did not provide, and a user can still ask it to ignore the rule. Treat the output as a draft. For code, content, and citations, the linked guides say what you check by hand.
How is this page different from the template pack and the long system-prompt library?
This page defines the mechanism and gives seven short examples of role, constraints, and output format. The template pack is a set of one-turn fill-in briefs. The longer ChatGPT system prompt catalog and the cross-tool system prompt examples are libraries. Use this page first if you are deciding where a sentence belongs.