Requirements Input
Step 1 of 2
Fill in the fields below — only the ones you have. Empty fields are skipped automatically.
Feature / Task Name *
Language / Stack *
Developer Role
Business Requirements
// paste the BA description or key business rules
API Contract / Endpoints
// REST endpoints, request/response format
Database Schema
// table names, key fields, relations
Events / Messages (Kafka / Queue)
// event names, payload structure
Constraints & Rules
// DO NOT / performance limits / forbidden approaches
Expected Output
// what should the developer produce?
Prompt Techniques to Apply
// select based on task complexity
Role (Persona)
XML Structure
DO / DO NOT
Chain-of-Thought
Few-Shot Example
Pseudocode Logic
Output Directives
Why XML Prompts Work
Research-backed
Semantic errors account for >50% of LLM code mistakes. Structure eliminates ambiguity.
<role>
Persona anchors the model
A model assigned an expert role generates more accurate, security-conscious code. Role Prompting shows +3–8% accuracy at minimal token cost.
Technique cost: +5–15% tokens · Gain: +3–8% accuracy
<constraints>
DO / DO NOT eliminates guessing
Explicit forbidden approaches save the model from "rejecting" bad solutions internally. This is Negative Space technique — what not to do is as valuable as what to do.
Technique cost: +10% tokens · Gain: fewer hallucinations
<output_format>
Explicit output cuts noise by 30–50%
Without output directives, models add explanations, markdown wrappers, and examples. "Code only, no explanation" removes 30–50% of unnecessary output tokens.
Directive: "code only" · Effect: -30–50% output tokens
<task> + pseudocode
Codified logic → 55–87% fewer tokens
Replacing natural language with pseudocode (IF/ELSE/RETURN) reduces input tokens by 55–87% with no accuracy loss. The model parses intent, not words.
CodeAgents · 56% success rate on VirtualHome benchmark
Combine wisely
Start with Signature + Few-Shot
Stacking all 5 techniques increases tokens without proportional quality gain. Best default: Function Signature + Few-Shot. Add CoT only for complex algorithms.
CodePromptEval · 7072 prompts tested on GPT-4o, Llama3, Mistral
Generated Output
Step 2 — Copy & Send to Developer
Ready to use
XML Prompt
Example Code Output
QA Checklist
Before handing this prompt to a developer — verify the checklist below.