System Prompts Explained: How to Shape AI Behaviour

What Are System Prompts?

Every time you use ChatGPT, Claude, or any AI assistant, there is a layer of instructions you do not see. This layer — the system prompt — tells the model who it is, how it should behave, and what rules it must follow.

Think of it like a job description. Before an employee starts their first day, they receive a brief: their role, responsibilities, tone, and boundaries. A system prompt does the same thing for an AI model.

When you use ChatGPT’s “Custom Instructions” or Claude’s “Project Instructions”, you are writing a system prompt. When developers build AI applications, they write system prompts that define how the AI behaves for every user.

Anatomy of a System Prompt

An effective system prompt typically contains five components:

1. Identity: Who is the AI? What is its name, role, and expertise?
2. Behaviour rules: How should it communicate? Formal or casual? Concise or detailed?
3. Knowledge scope: What does it know about? What should it refuse to discuss?
4. Output format: How should responses be structured? Markdown, plain text, JSON?
5. Guardrails: What must it never do? What safety boundaries exist?

Using Custom Instructions

You do not need to be a developer to use system prompts. Both ChatGPT and Claude offer built-in features that let you set persistent instructions:

• ChatGPT: Settings → Personalisation → Custom Instructions
• Claude: Create a Project and add instructions in the Project Settings
• Gemini: Extensions and Gems allow custom behavioural configuration

A well-written custom instruction saves you from repeating context in every conversation. Instead of starting each prompt with “I am a marketing manager at a SaaS company...”, you set that context once and it applies to every interaction.

Tips for Writing Effective System Prompts

1. Be specific about tone. “Be professional” is vague. “Write like a friendly colleague — warm but concise, no corporate jargon” is specific.
2. Use examples. Show the model what good output looks like. One example is worth a hundred words of description.
3. Define boundaries clearly. State what the model should not do. Explicit prohibitions prevent common failure modes.
4. Keep it structured. Use headings, numbered lists, and clear sections. The model parses structured text better than dense paragraphs.
5. Test and iterate. Write a draft, test it with ten different user messages, observe where it fails, and refine. Good system prompts are rarely written in one pass.

Common Mistakes

• Too long. System prompts that exceed 2,000 words often cause the model to forget early instructions. Keep it focused.
• Contradictory rules. “Be concise” followed by “Always provide comprehensive detail” creates confusion. Prioritise your instructions.
• No output format. If you do not specify format, the model will guess differently each time. Always define the expected structure.
• Forgetting edge cases. What should the model do when it does not know the answer? When the user is rude? When the request is out of scope? Define these scenarios.

Taking It Further

System prompts become truly powerful when combined with other techniques. Use RCTF as a framework for structuring each user interaction, and chain-of-thought prompting for tasks that require reasoning.

For developers building AI applications, system prompts are the foundation of the entire user experience. Every chatbot, copilot, and AI agent starts with a carefully crafted system prompt that defines its behaviour. Mastering this skill is one of the highest-leverage investments you can make in AI.