The guide provides 6 strategies for getting better results. It also gives examples, links and resources for further learning and experimentation with GPT models.
—
For those in a hurry, the six GPT strategies are:
- Write clear instructions: Provide details, examples, delimiters, steps and desired length for the model output.
- Provide reference text: Give the model relevant information or citations to avoid fabrications or errors.
- Split complex tasks into simpler subtasks: Decompose a task into a workflow of simpler tasks with specific instructions for each stage.
- Give GPT time to “think”: Ask the model to reason from first principles, work out its own solution, or use inner monologue before answering a question.
- Use external tools: Use embeddings, code execution, or APIs to augment the model’s capabilities or accuracy.
- Test changes systematically: Use evaluation procedures, gold-standard answers, or model-based evals to measure and optimize the model’s performance.