Zero-shot, few-shot, chain-of-thought, and other fundamental techniques explained clearly.
These are the prompting techniques that have held up across models and use cases. Not theory — practical methods with real uses.
What it is: Just ask. No examples, no elaborate setup. You describe the task and let the model work.
When to use it: Most of the time. For well-defined, common tasks, zero-shot prompting is sufficient. Start here and escalate only if the output isn't right.
Example: > "Summarize this customer complaint in two sentences, focusing on the core issue."
Zero-shot works well because modern models have seen vast amounts of examples during training. You don't need to show them what a summary looks like — they already know.
What it is: You provide two or three examples of what you want before making your actual request. The model pattern-matches to your examples.
When to use it: When you have a specific format, style, or structure that's hard to describe in words. If you've struggled to describe what you want, just show it.
Example: > Input: "Product ships in 3-5 days" → Output: "Delivery: 3–5 business days" > Input: "Free returns within 30 days" → Output: "Returns: Free, 30-day window" > Input: "Price includes all taxes and fees" → Output: [your actual task]
Few-shot is particularly powerful for classification tasks, consistent formatting, and matching an existing writing style.
What it is: You ask the model to reason through a problem step by step before giving an answer.
When to use it: Complex reasoning, multi-step math, logic problems, analysis where the reasoning matters as much as the conclusion.
Example: > "A project has three phases. Phase 1 takes 6 weeks, Phase 2 takes half as long as Phase 1, and Phase 3 takes twice as long as Phase 2. If we start on January 1, when do we finish? Think through this step by step."
Without "think step by step," models often produce wrong answers for problems like this by jumping to a conclusion. Making reasoning explicit significantly improves accuracy on complex problems.
The phrase "think step by step" or "let's work through this" reliably triggers more careful reasoning in most major models.
What it is: You assign the AI a specific persona, role, or area of expertise.
When to use it: When you want responses calibrated to a particular expertise level, perspective, or professional context.
Example: > "You are a senior UX designer with 10 years of experience in enterprise software. Review this onboarding flow and identify the three biggest friction points from a user experience perspective."
Role prompting changes more than tone — it shifts what the model considers relevant and important. An "experienced editor" focuses on different things than a "supportive writing coach," even given the same text.
What it is: You explicitly tell the model what structure, format, or length you want.
When to use it: Anytime you're going to use the output in a specific way. Saves cleanup time and produces more consistent results.
Examples: > "Return your answer as a JSON object with keys: name, description, price." > "Give me five options, formatted as a numbered list. One sentence per option." > "Write this as a table with columns: Action, Owner, Due Date." > "Under 100 words. No bullet points."
Models follow format instructions reliably. If you're not specifying format, you're leaving output structure to chance.
What it is: Using the first output as a starting point and improving it through follow-up instructions.
When to use it: Always — for anything you care about. The first draft is a draft, not the final product.
Example flow: 1. "Draft an introductory email to a potential client explaining our product." 2. "Good, but shorten it by 40% and make the opening less generic." 3. "Now make the second paragraph more concrete — include a specific example of how we helped a similar client." 4. "Adjust the tone — it's still too formal. Make it sound like a smart person wrote it, not a corporation."
Iteration is the single most underused technique. Most people accept the first response. Professionals iterate until it's right.
Have a follow-up question about this topic?
Ask AI