Skip to main content

Chapter 6: The CO-STAR Framework & System Prompts

6.1 Introduction​

Ad-hoc prompting is hard to scale. Frameworks make prompt quality repeatable.

CO-STAR is a practical structure for high-control prompt design, especially in team settings. Combined with system prompts, it enables consistent behavior across sessions and users.

6.2 CO-STAR Framework Breakdown​

CO-STAR stands for:

  • C: Context
  • O: Objective
  • S: Style
  • T: Tone
  • A: Audience
  • R: Response Format

Each component reduces ambiguity and narrows the model's output space.

6.3 CO-STAR in Practice​

Template​

Context: [background and relevant facts]
Objective: [specific outcome]
Style: [writing style or structure preference]
Tone: [formal, friendly, neutral, persuasive, etc.]
Audience: [who will read/use this]
Response: [required format, length, sections]

Example​

Context: A SaaS company is launching a new analytics feature for SMB customers.
Objective: Draft a feature announcement email.
Style: Clear and concise business communication.
Tone: Professional and optimistic.
Audience: Existing SMB customers.
Response: Subject line + body under 140 words + 3 bullet benefits.

6.4 System Prompts vs User Prompts​

System prompts define global behavior rules. User prompts define task-level requests.

System Prompt Should Include​

  • Role boundaries
  • Safety and policy constraints
  • Response behavior defaults
  • Handling of uncertainty
  • Formatting standards

User Prompt Should Include​

  • Task objective
  • Relevant context
  • Specific constraints
  • Output contract

6.5 Designing Strong System Prompts​

A strong system prompt is:

  • Stable: reusable across many tasks
  • Explicit: no vague instructions
  • Non-contradictory: rules are compatible
  • Minimal: only enduring behavior policies

Example System Prompt Skeleton​

You are an enterprise prompt assistant.
Always follow these rules:
1) Use only provided context when factual certainty is required.
2) If required information is missing, ask clarifying questions.
3) Never invent citations or external facts.
4) Return outputs in user-requested format exactly.
5) Keep responses concise unless user requests detail.

6.6 Layered Prompt Architecture​

Use three layers in production:

  1. System layer: global behavior and guardrails
  2. Developer/application layer: workflow rules and tool policies
  3. User layer: task-specific request

This separation improves maintainability and debugging.

6.7 Common Implementation Mistakes​

  • Duplicating constraints across layers with contradictions
  • Putting temporary task details in system prompt
  • Overly long system prompts that reduce effective context budget
  • Missing fallback behavior for uncertain inputs

Fix by versioning prompts and testing each layer independently.

6.8 Chapter 6 Practical Exercise​

  1. Write one reusable system prompt for a "Prompt QA Assistant."
  2. Create two user prompts for different tasks:
  • Task A: evaluate prompt quality
  • Task B: rewrite weak prompts
  1. Apply CO-STAR to both user prompts.
  2. Test on 5 inputs and record:
  • Consistency
  • Format compliance
  • Hallucination rate
  • Instruction adherence

6.9 Key Takeaways​

  • CO-STAR provides a reliable structure for prompt clarity.
  • System prompts control stable behavior across tasks.
  • Layered architecture improves governance and scale.
  • Prompt versioning and layer-level testing are essential in production.

6.10 Next Chapter​

In Chapter 7, we expand beyond text and learn how to design prompts for multimodal systems involving images, video, and audio.