A system prompt is a set of standing instructions that tells an AI model how to behave for a whole conversation: its role, tone, format and limits. Instead of repeating “keep it short” in every message, you say it once in the system prompt. Below are 12 examples you can copy, plus tips for writing prompts that small models running on a phone will actually follow.
The short version: be specific, keep it brief, say what to do rather than what to avoid, and include an example of the output you want.
What does a system prompt do? #
Every message you send goes to the model along with the system prompt. The model treats it as the ground rules. A good system prompt can:
- Give the AI a role (“You are a patient Spanish tutor”)
- Set the tone (“friendly and direct, no filler”)
- Fix the format (“answer in bullet points,” “under 100 words”)
- Add rules (“ask a question before giving advice”)
- Supply background (“I’m vegetarian and cook for two”)
Businesses use system prompts to turn one model into a support bot, a writing assistant or a coding helper. You can do the same thing for your own tasks.
In Personal LLM, each chat can have its own system prompt, so your recipe assistant doesn’t have to sound like your code reviewer. Everything runs on your phone, so your prompts and chats stay there.
12 system prompt examples to copy #
1. Plain, no-filler assistant #
Answer directly. Skip introductions, disclaimers and summaries unless I ask. Use short paragraphs. If a question is ambiguous, ask one clarifying question first.
2. Help-desk style explainer #
You are a patient tech support helper. Ask what device and app I’m using if I haven’t said. Give one step at a time and wait for me to confirm before the next. Use plain language, no jargon.
3. Recipe assistant #
You help me cook. I cook for two, I’m vegetarian, and I have about 30 minutes on weeknights. Suggest recipes using what I list. Give ingredients first, then numbered steps. Mention substitutions for anything I might not have.
4. Editor #
You are a careful editor. When I paste text, fix grammar, spelling and clarity. Keep my voice and meaning. Return the edited version first, then a short list of the main changes.
5. Tone checker #
For each message I paste, say in two sentences how it will likely come across and which words cause that. Then offer one rewrite in the tone I ask for. Keep my meaning and roughly my length.
There’s more on this use in our AI tone checker guide.
6. Language tutor #
You are a friendly Spanish tutor. I’m at a beginner level. Reply in simple Spanish, then give the English translation in parentheses. Correct my mistakes gently at the end of each reply, one or two at a time.
7. Study coach (no answer giving) #
You are a tutor. Don’t give me final answers. Ask guiding questions, give hints, and check my reasoning. If I’m stuck after two hints, show the next step only.
8. Code reviewer #
You review code. Point out bugs first, then readability issues. Show corrected code in a code block. Be concise. If you’re not sure something is a bug, say so.
9. Fact-careful assistant #
Be accurate. If you don’t know something or it depends on events after your training data, say so instead of guessing. Never invent sources, quotes or numbers. Keep answers brief.
Small models make things up more often, so this prompt pairs well with our guide to AI hallucinations.
10. Summarizer #
When I paste text, reply with: a one-sentence summary, then up to five bullet points of key facts, then any deadlines or action items. Don’t add information that isn’t in the text.
11. Brainstorm partner #
Help me brainstorm. Give 10 varied ideas, from safe to unusual. Keep each to one line. After the list, ask which ones I want to develop.
12. Calm listener #
You are a calm, kind listener. Let me talk. Ask one gentle question at a time. Don’t rush to fix things or give lists of advice unless I ask. If something sounds serious or unsafe, encourage me to reach out to a real person or a crisis line.
How to write system prompts that small models follow #
Models that run on a phone, like 3B to 9B models, follow instructions well, but they get confused by long, complicated prompts more easily than big cloud models do.
- Keep it short. Three to eight lines is plenty. Long prompts also take up context and slow every reply.
- Say what to do, not what to avoid. “Answer in three bullet points” works better than “Don’t write long answers.”
- Be concrete. “Under 80 words” beats “be concise.” “I’m a beginner” beats “keep it simple.”
- Show an example. One sample of the output format is often worth a paragraph of description.
- Put the most important rule first. Small models weight early instructions more.
- Test and adjust. Try three or four messages, see where it drifts, and tighten the prompt.
- Remind it mid-chat. In long conversations, a model can drift from its instructions. A quick “remember, bullet points” resets it.
System prompts vs presets #
A system prompt shapes what the model says. Settings like temperature shape how it says it. They work together:
| Goal | System prompt | Setting |
|---|---|---|
| Accurate facts | Fact-careful assistant | Low temperature |
| Creative writing | Story or brainstorm prompt | High temperature |
| Math or logic | “Show each step” | Thinking mode on |
| Quick edits | Editor prompt | Balanced |
Personal LLM has one-tap presets for this: Creative (temperature 1.2), Balanced (0.7), Precise (0.3) and Thinking (reasoning on). You can also set temperature, top-p, top-k, context size and max tokens yourself.
For a fun use of system prompts, see how to turn one into a game master for an offline AI text adventure.
Frequently asked questions #
What is a system prompt in AI? #
It’s a set of instructions given to an AI model before the conversation, telling it how to behave for every reply: its role, tone, format and rules. Users usually don’t see it in the chat, but it shapes every answer.
What’s the difference between a system prompt and a regular prompt? #
A regular prompt is one message, like a question. A system prompt applies to the whole conversation. You write it once, and the model follows it for every message in that chat.
How long should a system prompt be? #
For small on-device models, three to eight lines is usually best. Longer prompts can confuse smaller models, and they take up space in the context window that the conversation needs.
Why does the AI ignore my system prompt? #
The prompt may be too long or vague, or the conversation has gone on long enough that the model has drifted. Shorten it, make the key rule concrete and put it first, and remind the model mid-chat if needed.
Can I use a different system prompt for each chat? #
In apps that support it, yes. Personal LLM lets you set a system prompt per chat, so you can keep a tutor chat, a recipe chat and an editor chat side by side.