You give ChatGPT an instruction, it follows it properly for three messages, then it drifts back to its usual style. That's not a flaw in the tool: you wrote a one-off instruction, not a system prompt. A system prompt is the frame you set before the conversation starts: a role, some context, rules, an output format. The model keeps it in view at every turn, which makes it far more stable than a sentence you toss in halfway through.
Here you'll see exactly what a system prompt contains, the five blocks that make it solid, a complete example you can adapt in ten minutes, and three places to install it depending on the tool you use. You'll also see why rules sometimes start slipping in long conversations, and the simple habits that fix most cases. No technical skills needed: everything happens in plain English, in a standard interface.
What a system prompt actually is
A system prompt is a set of instructions placed above the conversation: it defines who the AI should be, what it should know and what it must respect, and it stays active for every message you send.
The most accurate comparison is a job description. The system prompt is the job description: role, scope, internal rules, deliverable formats. Your message today is the one-off request: "draft the quote for Mrs. Lefèvre." A new hire would find it strange if you repeated the whole job description with every request. An AI, on the other hand, needs it — and that's exactly what the system prompt does: it re-reads it at every turn without you having to retype it.
The name changes from tool to tool, but the mechanism is identical everywhere:
- in ChatGPT, they're called custom instructions (plus project instructions);
- in Claude, project instructions;
- in Gemini, Gems;
- on the developer side, a parameter named system sent with every request.
The concrete difference from a normal message: your message stays one item of history among dozens of others, whereas the system prompt is treated as a priority framing instruction. That's why a rule written in the right place holds up ten times better than the same rule sent in your third message.
The five blocks of a system prompt that goes the distance
A reliable system prompt always contains the same five blocks: role, context, rules, format and safeguards. If one is missing, that's almost always where the AI goes off the rails.
| Block | The question it answers | Short example |
|---|---|---|
| Role | Who's speaking? | You are a copy editor for an accounting firm. |
| Context | For whom, in what setting? | Our readers run small businesses and have no accounting background. |
| Rules | What to do, what to avoid? | Sentences of 20 words max, no jargon. |
| Format | What does a good answer look like? | One heading, three paragraphs, a three-bullet list. |
| Safeguards | What to do when in doubt? | If a detail is missing, ask instead of inventing. |
Here's a complete example, deliberately ordinary and very specific, for a small business:
Role: you are a writing assistant for Atelier Vermeil, a craft joinery workshop in Nantes with two employees.
Context: our customers are private individuals having a staircase or a custom bookcase installed. Project budgets run from 4,000 to 12,000 euros. They usually compare three quotes before choosing.
Rules: write in English, addressing the customer politely and professionally. Sentences of 20 words max. Never give a specific lead time (I approve those myself). Never use the word "cheap." Mention solid wood whenever the topic allows.
Format: one heading, three 60-word paragraphs, a three-bullet list, a one-line call to action.
Check: if information is missing (dimensions, wood species, access constraints), ask me instead of making it up. At the end of the message, flag any rule you weren't able to follow.
This prompt is under 150 words, and yet it removes most of the corrections you'd otherwise make by hand. Three details do the work: the rules are countable and verifiable ("20 words max" is measurable, "be concise" isn't), the bans are named explicitly, and the last block gives the AI permission to ask a question rather than fill a gap. To go further on how to phrase the instructions themselves, the complete method for writing a good prompt breaks down each building block.
When the rules start slipping
In a conversation that stretches on (thirty, forty exchanges), you may see the model let go of one or two rules. The reason is mechanical: the model works with a limited context window and naturally gives more weight to the latest messages than to the instructions at the start. Four habits are enough:
- Keep the system prompt short and ordered by priority. Past 400 or 500 words, rules get diluted and start contradicting each other.
- Number your rules. You'll be able to say "you skipped rule 3" instead of explaining everything again.
- Phrase things positively when you can. "Use short sentences" works better than "don't write long sentences."
- Open a new conversation for a new task instead of stacking fifteen topics in the same thread. That's the basic principle of context engineering: a clean, short context beats a long, cluttered one.
Method 1: build and test your system prompt in Skilzy's AI Lab
The fastest way to know whether your system prompt holds up is to test it on several requests in a row in one place, instead of guessing.
That's what the AI Lab built into Skilzy is for: you write your system prompt once, run several typical requests against it back to back, and adjust the rules that don't hold. Usage credits are included, so you don't need an extra subscription to run these tests. It's especially useful for system prompts, because not every rule holds the same way: your 20-words-per-sentence limit may be followed to the letter while your banned-vocabulary rule slips by the third message.
On the learning side, the Prompt Basics program walks through the build step by step: you start from an empty system prompt, add the five blocks one at a time, and see live what each addition changes in the answer. The exercises use real cases (writing assistant, customer follow-up assistant, email triage assistant) and you walk away with reusable system prompts, not lecture notes.
Two state-recognized, fundable certifications (RS7439 for AI content marketing, RS6792 for AI and sales) are available if you want a more formal framework.
The limitation worth stating plainly: if your need boils down to locking in three style rules in ChatGPT for your emails, the tool's built-in feature is plenty. That's method 2.
Method 2: ChatGPT's custom instructions
ChatGPT includes a free, permanent system prompt: custom instructions, which apply automatically to all your new conversations.
The path: open settings, then the Personalization section, then Custom instructions. You'll find two fields. The first is for describing yourself (job, industry, audience, constraints). The second is for describing how you want the answers (length, tone, format, bans). The official OpenAI documentation on custom instructions covers how they work and the cases where they don't apply.
In practice, put your role and context blocks in the first field, and your rules, format and safeguards blocks in the second. The fields are short (around 1,500 characters each), which forces a healthy discipline: keep the rules that genuinely cost you time when they're not followed.
Two limits to keep in mind. First, these instructions apply everywhere, including when you ask for a cake recipe: an overly specialized system prompt gets in the way outside its domain. Second, they don't work retroactively, so your older conversations carry on without them. For framing that only concerns one specific project, use a ChatGPT project instead: you write instructions that apply only to that project's conversations, which keeps the rest clean. If you're brand new to the tool, the guide how to use ChatGPT as a beginner walks through the interface in order.
Method 3: Claude projects, Gemini Gems and the system parameter
Claude, Gemini and the APIs each offer their own version of the system prompt, with one clear advantage: framing per use case rather than one global setting.
In Claude, you create a project and fill in its instructions. Everything you write there applies to every conversation in the project, and you can attach reference documents (style guide, pricing, approved sample copy). It's the most comfortable format when you're working on several accounts in parallel: one project per client, one system prompt per project.
In Gemini, the equivalent is called a Gem. You name it, give it its instructions, and call it up whenever you need it. Same principle, but with the logic of named assistants rather than workspaces.
On the technical side, if you go through an API or an automation tool like n8n, the system prompt is a dedicated field called system, separate from the user's messages. The Anthropic documentation on system prompts explains why that separation matters: whatever goes in the system field is treated as a frame, not as a negotiable request. It's also where you can slip in two or three examples of ideal answers, a technique covered in our article on few-shot prompting.
One tip that applies to all three: write your system prompt in your own text file, then copy it into the tool. Interfaces change, your rules stay, and you can reuse them elsewhere in thirty seconds.
Key takeaways
A system prompt is five blocks (role, context, rules, format, safeguards), under 400 words, with measurable, numbered rules, and a stable place to store it. Start small: take the task you repeat most often this week, write the five blocks, install them in your tool's instructions, and fix one rule every time an answer disappoints you. After three or four iterations, you'll have a frame that saves you real time, every day, without explaining anything twice.
