Applicability of Structured Prompts to Different Models
Different models vary in their capability dimensions. From the perspective of maximizing model performance, it is necessary to develop targeted prompts:
- Basic task scenarios: Simple prompts (such as one or two sentences) show little difference in performance across different models.
- Complex task scenarios: The effectiveness of structured prompts is strongly related to model capabilities: Optimization Strategies for GPT-3.5
If the effect is not good when using GPT-3.5, the following adjustments can be made:
- Simplify structural complexity: Reduce multi-level structures to two-level structures (such as using 1., 2., 3. as first-level headings and - as second-level sub-items).
- Adjust attribute words: Refer to AutoGPT prompts and use more intuitive attribute words such as Goals and Constraints instead of complex terms.
- Continuous iterative optimization: Example: The GPT-3.5 version of the LangGPT assistant improves the stability of model responses through structural simplification and attribute word adjustment. markdown
1. Role: Data Analyst # 2. Goals - Analyze the trends of sales data provided by users
- Identify outliers and provide suggestions
3. Constraints - Data integrity must be verified first (prompt if missing values > 30%)
- Conclusions must be visualized with charts (line charts + bar charts)
4. Workflow 1. User uploads CSV data file
- Output data overview → anomaly analysis → visualized charts
- Optimization logic: Replace # levels with numerical numbers, and clarify task boundaries with Goals and Constraints to reduce GPT-3.5’s understanding cost.