The 4-part framework + copy-paste templates
Practical techniques • 7 min read
TL;DR
Great prompts have four things: context (who/what), task (specific action), format (how to output), and examples (when possible). Iterate 2-3 times to refine.
The 4-Part Prompt Framework
1. Context
Tell the AI who it is and what it's dealing with.
You are a senior technical writer for a SaaS company.
The audience is engineers evaluating our API.
2. Task
State exactly what you want.
Write a 300-word intro for our webhooks documentation.
3. Format
Specify structure, length, tone.
Use H2/H3 headings. Include one code example.
Tone: friendly but technical. Avoid marketing language.
4. Examples (Optional but Powerful)
Show what good looks like.
Example intro style:
"Webhooks let you..." (calm, direct, no hype)
Advanced Techniques
Chain-of-Thought
Add "Think step by step" or "First, list your considerations, then answer." Improves reasoning on complex tasks by 20-40%.
Few-Shot Learning
Show 2-3 input/output pairs before the real question.
`Q: What's the capital of France?
A: Paris
Q: What's the capital of Japan?
A: Tokyo
Q: What's the capital of Ghana?
A:`
Role Prompting
Assigning an expert role improves quality: "As a nutritionist..." or "As a database administrator..."
Delimiters
Use triple backticks, XML tags, or dashes to separate input from instructions:
`Summarize the following text in one sentence:
<text>
[your content here]
</text>`
Constraints First
Put "must" and "must not" at the top, not buried at the end. Models read prompts sequentially and give more weight to early instructions.
Common Prompt Mistakes
- Being too vague: "Make it better" → "Make it more concise, cut 30% of words, use active voice"
- Asking for opinion when you need fact: "Is Python good?" → "Compare Python vs Rust for CLI tools"
- No output format: Always specify structure if you'll process the output
- Multi-question overload: Break complex asks into steps
- Ignoring the temperature: Use 0 for factual, 0.7-1.0 for creative
Prompt Templates You Can Steal
For Writing
Role: You are a [role].
Task: Write [content type] about [topic].
Audience: [description].
Length: [word count].
Tone: [tone].
Must include: [3 key points].
Must avoid: [what to skip].
For Coding
`Write a [language] function that:
- Input: [signature]
- Output: [signature]
- Constraints: [performance/style]
- Handle edge cases: [list] Include type hints and docstring. Skip comments in code.`
For Analysis
`Analyze this data: [data]
Steps:
- Identify key patterns
- Explain each pattern's likely cause
- Suggest one actionable insight Format: Markdown with H3 headers per section.`
How Prompts Differ Across Models
- Claude: Responds well to XML tags for structure and long context
- GPT-4: Handles more casual prompts, benefits from "step by step"
- Gemini: Long context strong; explicit format instructions help
Related: ChatGPT vs Claude · Claude API Guide · AI Glossary
🎧 Get Daily AI News in 5 Minutes
Stay current on AI with a 5-minute daily podcast:
Original: https://ainews.q-sci.org/how-to-write-chatgpt-prompts.html
Top comments (0)