Last updated: July 12, 2026
TL;DR
A system prompt, also called instructions, is the standing rule set that defines who your AI assistant is, how it talks, and what it can and cannot do. A good one covers:
- Identity, tone, and conversation flow, so every reply sounds like your business.
- Hard rules and limits, so the assistant never guesses at policy or pricing.
- Scenario handling and handoff, so edge cases route to a human instead of improvising.
- Key prompt components include: identity, tone, conversation flow, response rules, scenario handling, knowledge base reference, and defined limitations.
The eleven elements below are the difference between a generic chatbot and an assistant that sounds like your team.
What is a system prompt?
A system prompt, also called instructions, is the set of standing rules you give an AI assistant before any customer ever types. It defines who the assistant is, how it sounds, and what it can and cannot do, and it stays active in every conversation. Write it well and every reply sounds like your business; write it vaguely and you get a generic chatbot.
System prompt vs user prompt
A system prompt and a user prompt play different roles. The system prompt is set by you, the builder: it defines the assistant’s identity, rules, and limits, and it stays active across the whole conversation. A user prompt is what a customer types in a single turn. When the two conflict, a well-designed assistant follows the system prompt: a customer can ask for a refund, but if your rules say refunds need human approval, the assistant routes it to your team instead of improvising.
System prompt vs user prompt: when they conflict, the system prompt wins.
Why a good system prompt matters
An assistant with clear instructions resolves more conversations on the first try, stays on brand under pressure, and knows when to bring in your team. An assistant without them improvises: it answers in different tones, guesses at policy, and treats every edge case as a surprise.
Customers feel the difference immediately. Clear instructions show up as fast understanding and simple steps. Vague ones show up as repeating yourself, robotic replies, and answers that sound like every other bot.
The eleven elements below are the anatomy of a strong system prompt, each with an example and the practical reason it earns its place.
Want a copy-paste starting point? Grab our system prompt template with worked examples and adapt it to your business.
The 11 elements of a great system prompt.
1. Identity & Purpose
Clearly state who the assistant is and what its core mission/goal is.
Example:
"You are Gigi, a customer service assistant for Los Santos Credit Union. Your primary purpose is to help customers resolve issues with their products, answer questions about services, and ensure a satisfying support experience."
Without a stated role, the model improvises one, and its answers drift with it. A one-line identity is the cheapest hallucination protection you can buy.
Reference:
OpenAI’s prompting guidance makes the same point: give the model a clear role and purpose, and responses get more accurate and consistent.
2. Tone, voice & persona
Defines the character, tone, and “voice” the assistant should use, personality traits, text patterns, and pace.
Example:
- Friendly, patient, professional
- Uses natural contractions and a conversational tone
- Varies sentence complexity to sound human
Tone is what customers remember. The same correct answer lands completely differently when it sounds like your brand instead of a default model voice.
3. Conversation flow
Breaks down the ideal structure for interactions step-by-step: introductions, diagnosing, troubleshooting, resolving, and closing.
Example flow:
- Warm greeting
- Acknowledgment of issues/frustration
- Steps or "Happy path"
- Stepwise problem-solving
A defined flow keeps the assistant from skipping steps, and skipped steps are where first-contact resolution dies: unconfirmed details, missed follow-ups, conversations that end without an answer.
4. Response Guidelines
Short, actionable rules for how the assistant should respond during the interaction.
Examples:
- Keep responses under 30 words
- Avoid multiple questions at once
- Match the customer’s technical level
Concrete response rules are also your tuning dials: when replies run long or drift off-topic, you adjust one line instead of rewriting the whole prompt.
5. Scenario handling
Instructions for handling specific, common situations (e.g. password resets, frustrated customers, billing).
Example:
For frustrated customers, “Acknowledge feelings. Take ownership. Focus on solutions.”
Real conversations are mostly edge cases: the frustrated customer, the double refund request, the message in another language. Scripting the common ones keeps the assistant consistent exactly when it matters.
6. Knowledge
While the instructions tell your assistant how to act, what tone to use, and how to manage conversations, they are not meant to hold all your detailed information. Instead, a well-designed AI assistant should have explicit access to a separate knowledge base. Think of it as your assistant's brain.
The knowledge base is not part of your instruction guide, but your guide should always reference it clearly.
In short:
- Instructions = How to act (the “manual”).
- Knowledge base = What to say (the “brain”).
Write the rules once in the instructions; update the facts anytime in the knowledge base.
The knowledge is where all the facts live:
- Product information
- Troubleshooting guides
- Company policies
- FAQs and resource links
- and more
Example:
TechSolutions offers... flagship products include TaskMaster Pro $389 (productivity), SecureShield $120 (security)...
This split is also what keeps answers current: prices, hours, and policies change in the knowledge base once, without touching the prompt.
For the file formats Invent supports, see our knowledge base guide.
7. Available actions
You can instruct your assistant to perform actions in a specific order based on the scenario. For example, to add a new client: first search the database (Zoho, Airtable, Notion), then create the user only if they don’t already exist. Clear step-by-step instructions ensure accurate results.
8. Limitations or constraints
Explicitly defines what the assistant cannot do.
Example:
- Cannot process refunds
- Cannot make changes to account ownership
- Cannot schedule an event for the same client on the same week
Limits are where compliance lives. An assistant that knows what it cannot do escalates instead of guessing, and that protects you with regulators as much as with customers.
9. Confirmation & refinement protocols
Guides for verifying info and refining responses for clarity.
Examples:
- Explicit confirmation (“So your email is... Is that correct?”)
- Use analogies for technical topics
One confirmation line costs the customer a second and prevents the expensive class of mistakes: wrong bookings, wrong accounts, wrong charges.
10. Quality Assurance
Provide instructions to ensure accuracy with numbers, emails, ID photos, dates, and links. The model is smart, but adding specific instructions will further improve the accuracy of the results you expect.
Example:
- "Heads up: Many users may try to upload meme images instead of real photos. Make sure to instruct the assistant to only accept genuine headshots, no memes or cartoons."
This ensures it analyzes only suitable images provided by your clients or users, guaranteeing accuracy and relevance."
The model is smart, but smart does not mean careful with your specifics. Explicit checks on numbers, dates, and uploads catch exactly the errors customers screenshot.
11. Brand-specific adjustments & warm closures
Customization for your target audience, cultural context, or temporary campaigns.
Example:
Always ending with a positive, brand-aligned closing as “That’s wonderful to hear, your satisfaction means the world to us! We truly appreciate you taking the time to share your feedback. If there’s anything else we can help with, just let us know. Have a fantastic day!”
The last message is the one customers remember, and a branded closure is the difference between ending a ticket and ending a conversation.
FAQ
How detailed should my instructions be?
Match the detail to the complexity of the use case. A booking assistant might need half a page; a support assistant with refund rules, account tiers, and compliance requirements needs much more. The test: every line should change behavior. If it does not, cut it.
Can I update the instructions?
Yes, and you should. Review it after real conversations, tighten the wording where the assistant drifts, and update it when your policies change.
Why does sharing information about your business with your assistant matter?
The more context the assistant has, the fewer wrong answers and clarifying questions your customers sit through. Specifics about your products, policies, and customers are what turn generic replies into answers that sound like they came from your team.
What's the best way to test?
Test in the playground with real customer questions before going live, then review actual transcripts and tighten the wording wherever the assistant drifts. Repeat after every meaningful change.
Do I need a system prompt?
If you want consistent, on-brand replies, yes. Without one, the model falls back to generic defaults: helpful, but it sounds like everyone else’s bot and improvises on policy. Even three or four clear rules covering identity, tone, limits, and handoff noticeably change behavior.
Can a user prompt override a system prompt?
By design, no. The system prompt outranks user messages, which is what protects your rules when customers push back ("just process the refund"). A well-built assistant treats conflicting requests as escalation triggers, not instructions.
Which platforms give you full control over the system prompt?
Developer APIs like OpenAI’s and Anthropic’s expose the system message directly in code. Among no-code platforms, look for an editable instructions field paired with a testing playground. On Invent, the Instructions field gives you full control over the system prompt, and you can test changes in the playground before going live.
Conclusion
Detailed instructions give your assistant a consistent identity, rules it will not improvise around, and a clear script for the hard moments. That shows up as fewer errors, fewer escalations, and conversations that end with the customer trusting you a little more than before.
Ready to craft your instructions for your assistant?
Start with clear purpose, tone, and flow, iterate based on your testing through the playground, and keep improving it according to your needs and conversations with real users.
