Prompt Engineering

The same AI model gives wildly different outputs depending on how you ask. "Write a blog post about AI" gives a generic 500-word essay. "Act as a senior tech journalist writing for an Indian startup audience. Write a 600-word piece on how AI is changing jobs in India in 2025. Use data points, a surprising opener, and end with what readers should do next." gives something publishable.

Prompt engineering is the skill of communicating clearly with AI — and it has a bigger impact on output quality than which model you use.

R — Role: Set who the AI should be. "Act as a senior product manager at a startup."

C — Context: Give the relevant background. "We are launching a B2B SaaS product for Indian SMEs in the accounting space."

T — Task: Be specific about what you need. "Write 3 subject lines for a cold email campaign targeting CFOs."

F — Format: Specify the output format. "Format as a numbered list. Each subject line under 8 words. Include an emoji."

The more specific each element, the better the output.

Chain of thought: Add "Think step by step" before complex tasks. This forces the model to reason through the problem instead of jumping to an answer.

Few-shot prompting: Give 2-3 examples of what you want before the actual request. The model will pattern-match to your examples.

Self-consistency: Generate the same output 3 times with a higher temperature, then pick the best or average them.

System prompts: Use the system message to set persistent instructions that apply to the entire conversation.

Negative prompting: "Do not use jargon, do not exceed 300 words, do not use bullet points."

For writing: "Act as [expert]. Write [format] for [audience] about [topic]. Tone: [tone]. Length: [words]. Include: [specific elements]. Avoid: [what to skip]."

For analysis: "Analyse [subject]. Identify: 1) Top 3 strengths 2) Top 3 weaknesses 3) Your recommendation. Be specific and use data where possible."

For coding: "You are a senior [language] developer. Write [code task]. Requirements: [list]. Add comments explaining non-obvious logic. Handle edge cases for [specific scenarios]."

For debugging: "Here is my code: [paste code]. Here is the error: [paste error]. Explain what is causing it and provide the fixed code with explanation."

Too vague: "Write something about AI" — too open. Be specific about topic, audience, format, and length.

No context: The AI doesn't know your industry, audience, or constraints unless you tell it.

One-shot thinking: AI conversation is iterative. Generate, critique, refine. Never accept the first output as final.

Ignoring system prompts: Setting a system prompt with your background and preferences saves you from repeating context every time.

Not iterating: If the output is wrong, don't regenerate — tell it specifically what to change.