Introduction

Instructions provide clear direction to the model so that generated responses align with brand, policy, and reader expectations.

• Definition: Structured rules describing tone, format, scope, and prohibited content.
• Why: Without guidelines, the model may improvise, leading to inconsistent or non-compliant answers.
• Where: Help centers, internal knowledge bases, onboarding assistants, marketing copy review.

Syntax

[
  {
    "role": "system",
    "content": "You are an editor for ItTechGenie. Follow the checklist: tone=confident, format=Markdown, include key takeaways." 
  },
  {
    "role": "user",
    "content": "Topic: Benefits of Infrastructure as Code. Avoid marketing buzzwords." 
  }
]

Example

[
  {"role": "system", "content": "Respond in numbered steps with short sentences. Conclude with a 3-item checklist."},
  {"role": "user", "content": "Explain how to roll back a failed Azure deployment."}
]

1. Pause incoming traffic using the deployment slot swap.
2. Redeploy the previous stable build from the release pipeline.
3. Run smoke tests on the restored version before reopening traffic.

Checklist:
- [x] Incident ticket updated
- [x] Stakeholders notified
- [ ] Post-incident review scheduled

[
  {"role": "system", "content": "Never invent numbers. Cite the source section when quoting policies."},
  {"role": "user", "content": "What onboarding steps are required for a contractor according to section 5.2?"}
]

Explanation

• Instruction hierarchy: Place non-negotiable guidelines in the System message and optional hints in the User message.
• Clarity beats length: Bullet lists and checklists help the model prioritize requirements.
• Conflict resolution: If instructions conflict, the earliest (System) direction usually wins—review prompts to avoid ambiguity.
• Iteration: Gather feedback from readers and refine instructions to cover edge cases.

Real-World Use Case

A customer-success team builds a knowledge base assistant. System instructions enforce inclusive language, require short paragraphs, and demand links to official docs. Analysts can now trust that every answer aligns with brand tone and compliance rules.

Key Notes / Tips

• Maintain a shared library of reusable instruction snippets for different personas.
• Spell out “do not” rules to prevent hallucinated references or promises.
• Limit the number of goals per prompt—crowded instructions dilute model attention.
• Use evaluation datasets to check whether guidelines are consistently followed.
• Document revisions; even small wording changes can shift response quality.

Practice Exercise

1. Draft a System instruction for a release-notes generator that enforces a change log template.
2. Create a User instruction that prevents mention of internal project names while summarizing incidents.
3. Challenge: Combine System and User instructions to produce bilingual (English + Spanish) answers while staying concise.

Summary

Strong instructions anchor the model’s output. Define tone, structure, and guardrails clearly, then iterate until the assistant behaves like a domain expert for your audience.