Good prompts aren't about being polite or verbose — they're about being structured and precise. These 6 techniques are what separate a $0.05 prompt that gets mediocre results from a $0.03 prompt that gets excellent results.
Force the model to reason step by step instead of jumping to an answer. Best for: math, logic, debugging, analysis.
Think step by step: 1. Identify the core problem 2. Break it into sub-parts 3. Address each part with evidence 4. Synthesize into a final answer Problem: [your question here]
~35 tokens overhead. Increases output quality significantly for reasoning tasks.
Show the model 2-3 examples of the input/output pattern you want. Best for: classification, formatting, extraction.
Input: "The service was terrible"
Output: {"sentiment": "negative", "confidence": 0.95}
Input: "Pretty good, would recommend"
Output: {"sentiment": "positive", "confidence": 0.80}
Input: "[your text here]"
Output:
~50 tokens per example. 2-3 examples is usually optimal — more doesn't help much.
Structure your prompt with XML-like tags. Models (especially Claude) parse these as semantic boundaries. Best for: complex prompts with multiple sections.
<context> [paste relevant documentation here] </context> <task> Based on the context above, write a migration plan. </task> <constraints> - Maximum 3 steps - No breaking changes </constraints>
~20 tokens overhead for tags. Dramatically improves output quality for multi-section prompts.
Give the model a specific role with years of experience. Best for: domain-specific tasks, code review, writing.
You are a senior PostgreSQL DBA with 15 years of experience in high-traffic systems. You prioritize query performance and data integrity. Review this query for performance issues: [query]
~25 tokens. The persona primes the model to use domain-specific terminology and reasoning patterns.
Explicitly state what the model should NOT do. Best for: preventing hallucination, controlling output format.
Constraints: - Do not make up information. If unsure, say "I don't know." - Stay within scope. Do not add features not requested. - Maximum 200 words. - Respond in valid JSON only.
~30 tokens. Essential for production prompts where reliability matters more than creativity.
Give the model scoring criteria before asking it to evaluate. Best for: code review, content quality, grading.
Evaluate using this rubric: - Correctness (0-5): Is the code logically correct? - Security (0-5): Are there vulnerabilities? - Readability (0-5): Is it easy to understand? Provide scores first, then detailed feedback.
~35 tokens. Forces structured, consistent evaluation output.
The real power is in combining: use a Persona + XML Tags + Constraints for complex system prompts, or Few-Shot + CoT for tricky classification tasks.
Use WeighMyPrompt's Prompt Builder to combine these techniques visually — pick a task, select techniques, and see the prompt build in real time with a live token counter.
A typical production prompt using 3 techniques costs 80-120 tokens of overhead. At GPT-4o pricing ($5/M input), that's $0.0004-$0.0006 per request. The quality improvement easily justifies this cost.
Calculate your exact cost with WeighMyPrompt — free, private, and accurate.