6 Melhores Práticas para Geração de Código com LLM

1. Use YAML structured output instead of JSON format - YAML has higher error tolerance, JSON is prone to format errors that cannot be parsed - YAML content consumes fewer tokens 2. Bullet points analysis - When asking LLM to analyze problems, requesting output in bullet points yields better results - Bullets promote deep understanding and logical semantic segmentation - Example: self-reflection questions with bullets for general description, goals and rules, input structure, output structure 3. LLMs excel at generating modular code - Instruct: 'Split the generated code into multiple small sub-function modules and give them meaningful names' - Results in fewer errors and higher success in iterative fixes 4. Flexible decision-making and double validation - After initial output, ask the model to generate the same output again and correct if necessary - Example: feed the model its own AI tests as input, regenerate tests, and correct errors - More effective than asking 'Is this test correct?' yes/no questions 5. Delay decision-making, avoid direct questions, give exploration space - Start with simplest task: self-reflection and reasoning about public test cases - Then generate additional AI tests and possible solutions - Only after model answers, proceed to actual code generation and iterative fixing - Evaluate and rank multiple possible solutions, prioritize top-ranked ones for initial coding 6. Flow-oriented supervision - Design a scientific flow, accumulate data step by step, add new data at each stage - Example for CodeContests: two-phase flow with steps for reflection, public test analysis, solution generation - Validate and select data at each step, with iterative correction if errors occur