Why Prompt Engineering Matters
The same AI model can produce a vague, unhelpful response or a precise, actionable answer — the only difference is how you ask. Prompt engineering is the practice of structuring your input to reliably get the output you need. It's not magic; it's communication design applied to language models.
Good prompts can eliminate 80% of the back-and-forth you'd otherwise spend refining AI outputs. For developers building on LLM APIs, better prompts mean fewer tokens, lower costs, and more predictable outputs.
The Anatomy of a Good Prompt
Every effective prompt has up to five components. You don't always need all five, but knowing each one helps:
- Role: Who should the AI be? ("You are a senior financial analyst...")
- Context: What background information does it need? ("I'm building a B2B SaaS for HR teams...")
- Task: What exactly should it do? ("Write a 3-paragraph executive summary...")
- Format: How should the output be structured? ("Respond in JSON with keys: title, summary, action_items")
- Constraints: What limits apply? ("Under 200 words, no technical jargon, written for a C-suite audience")
Technique 1: Chain-of-Thought Prompting
For any task requiring reasoning, logic, or multi-step calculation, appending "Let's think step by step" (or asking the model to reason before answering) dramatically improves accuracy.
Without chain-of-thought:
Prompt: "If a store sells 3 items at $4.50 each and gives a 15% discount, what's the total?"
Response: "$11.47" (often wrong)With chain-of-thought:
Prompt: "If a store sells 3 items at $4.50 each and gives a 15% discount, what's the total? Think step by step."
Response: "3 items × $4.50 = $13.50. Discount: $13.50 × 0.15 = $2.025. Total: $13.50 - $2.025 = $11.48."Chain-of-thought works because it forces the model to generate intermediate reasoning tokens before the answer, reducing the chance of the model "jumping" to a wrong conclusion.
Technique 2: Few-Shot Prompting
Show 2–3 examples of the input/output pattern you want before giving the actual request. This is especially powerful for consistent formatting, tone matching, and domain-specific outputs.
Classify the sentiment of product reviews as Positive, Negative, or Neutral.
Review: "Arrived on time, exactly as described."
Sentiment: Positive
Review: "The zipper broke after two uses. Very disappointed."
Sentiment: Negative
Review: "It's okay. Does what it says."
Sentiment: Neutral
Review: "Absolutely love it, already ordered a second one!"
Sentiment:The model now has a crystal-clear pattern. Without examples, it might return "The sentiment is positive" — with examples, it returns exactly "Positive".
Technique 3: Role Assignment
Assigning a specific expert persona produces more authoritative, domain-appropriate responses.
Generic: "Explain how to improve my website's SEO."
With role: "You are an SEO consultant with 10 years of experience specializing in e-commerce sites. Explain the three highest-impact changes I can make to improve my product pages' search rankings."
The role doesn't just change tone — it shifts the model toward vocabulary, frameworks, and recommendations appropriate for that domain.
Technique 4: Output Format Control
Be explicit about the format you want. LLMs are highly responsive to format instructions.
- JSON: "Respond with a JSON object. Fields: name (string), tags (array of strings), priority (1-5 integer)"
- Markdown table: "Format the comparison as a markdown table with columns: Feature, Tool A, Tool B"
- Numbered list: "Give me exactly 5 recommendations, numbered, each under 30 words"
- XML/structured tags: "Wrap your response in <analysis> and <recommendation> tags"
When building applications, always request JSON output and use response_format: { type: "json_object" } (OpenAI) to enforce it at the API level.
Technique 5: Constraint Setting
Constraints focus the model and prevent it from padding responses with filler. Common constraints:
- Word/character limits: "In exactly 50 words..."
- Audience level: "Explain as if to a 12-year-old" or "Assume the reader is a PhD in machine learning"
- Tone: "Use a formal, professional tone" / "Keep it conversational and friendly"
- Scope: "Focus only on open-source solutions" / "Only consider options under $100/month"
- What to exclude: "Do not include caveats or disclaimers"
Before/After Examples
Scenario 1: Writing a bio
Bad: "Write a bio for me."
Good: "Write a 3-sentence professional bio for a product manager with 5 years of experience at SaaS startups. Tone: confident but approachable. Do not use the word 'passionate'."
Scenario 2: Code review
Bad: "Review this code."
Good: "You are a senior Python developer. Review the following function for: (1) correctness, (2) edge cases not handled, (3) performance issues. Format as a numbered list. Be specific — point to line numbers."
Scenario 3: Email drafting
Bad: "Write an email to a client about a delay."
Good: "Draft a professional email to a B2B client informing them their project delivery will be delayed by 2 weeks due to unexpected technical issues. Acknowledge the inconvenience, explain briefly without over-detailing, offer a specific revised date, and include a small gesture of goodwill. Keep it under 150 words."
Scenario 4: Data extraction
Bad: "Extract the key info from this job posting."
Good: "Extract structured data from this job posting. Return JSON with: title, company, location (remote/hybrid/onsite), salary_range (null if not listed), required_years_experience, top_5_required_skills (array)."
Scenario 5: Summarization
Bad: "Summarize this article."
Good: "Summarize this article in 3 bullet points, each under 20 words. Focus on: the core finding, the method used, and the practical implication. Avoid jargon."
Common Mistakes to Avoid
- Being too vague: "Help me with marketing" gives the model no direction. Always specify the task.
- Asking for too much at once: Break complex requests into sequential prompts or use chain-of-thought
- Assuming context: The model doesn't know your industry, audience, or goals unless you tell it
- Ignoring negative instructions: "Don't include..." works, but "Focus only on..." is usually more reliable
- Not specifying format: If you need structured data, always specify the exact format
Prompt Chaining for Complex Tasks
For tasks too complex for a single prompt, break them into a chain. Each output becomes the input for the next step:
- Prompt 1: "Given this brief, generate 10 potential blog post angles."
- Prompt 2: "From these 10 angles, pick the 3 most likely to perform well for SEO and explain why."
- Prompt 3: "Write a full outline for angle #2."
- Prompt 4: "Write the full article from this outline."
Quick Reference Cheat Sheet
- Add role: "You are an expert [role]..."
- Force reasoning: "Think step by step before answering."
- Get JSON: "Respond only with a JSON object. No prose."
- Consistent format: Provide 2–3 examples before your request
- Constrain length: "In exactly N words / bullet points / sentences"
- Audience: "Explain to [specific audience]"
- Verify: Add "If you are unsure, say so." to reduce hallucinations