How to Prompt ChatGPT: The Complete Guide

ChatGPT processes instructions in a clear priority order: System → User → Tools → Assistant. Understanding this hierarchy is key to reliable prompt engineering — conflicting instructions are resolved by priority level.

// System prompt (highest priority)
"You are a technical writer. Always use British English. Never exceed 200 words."

// User prompt
"Write a product description for our API monitoring tool."

// Tools (function calling)
{ name: "format_output", parameters: { format: "markdown", sections: ["headline", "body", "cta"] } }

// Assistant (prefilled context)
"Here is the description following your style guide..."

💡 Place hard rules in the system prompt where they can't be overridden by user input — critical for deployed chatbots.

Function calling is ChatGPT's most powerful feature for structured output. Define a JSON schema and GPT returns data matching it exactly. No parsing errors, no retry logic, no "here is the JSON" preamble.

// Define the function
{
  name: "extract_entities",
  description: "Extract structured entities from text",
  parameters: {
    type: "object",
    properties: {
      people: { type: "array", items: { type: "string" } },
      organisations: { type: "array", items: { type: "string" } },
      dates: { type: "array", items: { type: "string" } },
      sentiment: { enum: ["positive", "negative", "neutral"] }
    },
    required: ["people", "organisations", "sentiment"]
  }
}

💡 Function calling is the most reliable way to get structured output from any OpenAI model — prefer it over prompt-based JSON requests.

Different GPT models respond differently to temperature. GPT-4o uses temperature 0.0-1.0 conventionally. o3/o3-mini use internal chain-of-thought and are designed for temperature 1.0 by default — lowering temperature can actually hurt reasoning quality.

// GPT-4o: Creative writing
{ model: "gpt-4o", temperature: 0.8 }

// GPT-4o: Data extraction
{ model: "gpt-4o", temperature: 0.0 }

// o3-mini: Complex reasoning
{ model: "o3-mini", temperature: 1.0 }  // Default, don't lower!

// o3-mini: Math/logic
{ model: "o3-mini", reasoning_effort: "high" }  // Use effort, not temp

💡 For o3 models, control output quality with reasoning_effort (low/medium/high) instead of temperature. High effort = more thinking tokens = better accuracy.

GPT-4o can process images alongside text. Provide images via URL or base64 encoding with specific instructions about what to analyse. Be explicit about what you want extracted — "describe this image" produces generic output.

{
  role: "user",
  content: [
    { type: "text", text: "This is a screenshot of our checkout page. Identify:\n1. UI/UX issues\n2. Accessibility violations\n3. Mobile responsiveness concerns\n\nFormat as a prioritised list with severity ratings." },
    { type: "image_url", image_url: { url: "https://..." } }
  ]
}

💡 Be as specific with vision prompts as with text prompts. "Analyse this screenshot" produces vague output. "Identify accessibility violations in this checkout form" produces actionable results.

For production ChatGPT deployments, structure system prompts in four blocks: Identity, Rules, Output Format, and Boundaries. Keep under 500 tokens for GPT-4o to maintain strong instruction following.

## Identity
You are Aria, a customer support specialist for TechCorp.
Experience: Enterprise SaaS support, 5 years.
Tone: Professional, warm, concise.

## Rules
1. Always check the knowledge base before answering.
2. Never share internal pricing or roadmap.
3. Escalate billing disputes to human agents.
4. Respond in the customer's language.

## Output Format
- Greeting (1 sentence)
- Answer (2-4 sentences)
- Next step or follow-up question

## Boundaries
Refuse: legal advice, competitor comparisons, personal opinions.
Redirect: "I'd recommend contacting our [team] for that."

💡 Test your system prompt against adversarial inputs — users will try to override it. Place the most important rules first.

Few-shot extraction minimizes context window usage vs zero-shot verbose.

3 well-crafted few-shot examples (150 tokens) outperform a 600-token verbose instruction block, saving 75% on input costs per request.

Without concise few-shot examples, developers write lengthy prose instructions that consume 4x more tokens for equivalent or inferior output quality.

JSON Schema enforcement eliminates parse errors.

OpenAI structured outputs with JSON Schema achieve 99.9% schema adherence vs <70% with unconstrained generation — a 30x reduction in parse failures.

Without schema enforcement, every 1M requests generate 300K+ malformed responses requiring retries, error handling, and downstream data corruption.

Chain-of-thought prompting improves complex reasoning accuracy.

Adding 'Let's think step by step' improves accuracy on GSM8K math benchmarks from 17.7% to 78.7% — a 4.4x improvement on multi-step reasoning tasks.

Without chain-of-thought, models attempt to produce answers in a single leap, failing on problems requiring intermediate steps.