Lesson 14. System prompts and persistent instructions
Custom Instructions in ChatGPT, the System field in Claude, the system parameter in API — all implement one idea: giving AI persistent instructions before the conversation starts. Write once — and they work in every response.
Topic breakdown
Unlike a regular prompt, a system prompt is active throughout the entire conversation. It tells the model who it is, how to write, what not to do, and what format to use by default.
For business this means: creating a specialized assistant for a specific task, maintaining a unified brand tone across all responses, and standardizing team work.
In ChatGPT this is configured through Custom Instructions. In Claude through the System field. When working with API — through the system role. Each approach has its own capabilities and limitations.
A good system prompt consists of three parts: who (role), how (behavior rules), what not (restrictions). These three elements keep all responses on track.
What you'll learn
- understand the difference between system and regular prompts
- configure Custom Instructions in ChatGPT and System in Claude
- translate brand tone and rules into persistent instructions
- write a complete system prompt for a support bot or writing assistant
Lesson plan
What is a system prompt?
ChatGPT Custom Instructions, Claude's System field, the system role in API — all are mechanisms for setting persistent instructions before a conversation begins.
Role, rules, and restrictions
A good system prompt: who (role), how (behavior), what not (restrictions). Three elements keep responses on track.
Standardization for teams
A shared system prompt is the simplest way to ensure the entire team gets consistent quality responses. Brand tone and format are maintained automatically.
Testing and updating
After writing a system prompt, test it on 5-10 different situations. If responses aren't right — refine rules or add restrictions.
Weak vs strong prompt
You are a professional copywriter. Write in English.
Role: you are a content manager for Prompter.uz. Brand: technology and AI education. Behavior: always write in simple English. Response 150-200 words. Tone: friendly and inspiring. Include 1 practical tip in each post. Prohibited: technical jargon; exaggerated promises. Format: headline + body text + 1 CTA.
The second version sets role, tone, format, and restrictions. As a result, every response meets the brand standard without additional clarification.
Ready prompt template
Copy and adaptRole: you are [specific role and specialization]. Company: [name]. Product: [brief]. Behavior: always write in [language]. Response length: [short/medium/full]. Tone: [professional/friendly/business]. Required: [what must be in every response]. Prohibited: [don't write 1]; [don't write 2]. Format: [response structure].
Why it works
The role block sets the model's perspective: copywriter, support specialist, legal consultant, or another specialization.
Behavior rules — tone, language, length, and required elements — stay the same across all responses.
Restrictions protect against unwanted topics, incorrect claims, and words that don't match the brand.
The format block ensures the client receives a structured response of the same type every time.
Practice
- Write a simple system prompt for your profession: role, tone, and one restriction.
- Insert it into ChatGPT Custom Instructions or the System field in Claude.
- Ask 3 different questions and observe the system prompt's influence.
- Add one new rule and see what changes in the responses.
Mini-project
Mini-project: system prompt for your assistant
Think about how you use AI at work and write a persistent system prompt for that task. Goal: get consistently quality responses without repeated explanations.
Tasks
- Identify which repeating task you use AI for.
- Describe the role, behavior, and restrictions.
- Insert into ChatGPT Custom Instructions or Claude System.
- Test on 5 real tasks.
Deliverables
- 1 complete system prompt
- 5 test results
- list of edits for the next version
Checklist
Common mistakes
- making the system prompt too long — the more rules, the higher the chance some will be ignored
- giving contradictory instructions: for example 'write briefly' and 'explain in detail' simultaneously
- not adding a list of restrictions — the model may add unwanted content
- writing the system prompt once and never updating it
Lesson FAQ
How does a system prompt differ from a regular prompt?
A regular prompt works once. A system prompt is active throughout the entire conversation and influences every response.
How long should a system prompt be?
100-300 words is sufficient. Overly long system prompts may lead to some rules being ignored.