Tips for LLM Coders
Here are a few practical tips for effectively working with Large Language Models (LLMs) to improve your code productivity and results.
Prompt Engineering
Write and Save Your Prompts
- Always write and save your prompts in a text file first
- Lets you review and iterate until it looks good
- Avoids submitting typos to the LLM
- Allows you to rollback to previous versions
- Enables easy sharing and reuse
- If you need to modify a prompt:
- Make a copy (in the same file) of the original prompt and modify the copy
- This lets you see the progression and evolution of your prompts
- Use clear naming or versioning (e.g., “v1”, “v2”, or timestamps)
Use Structured Formats
- Write your prompts using either JSON or XML for complex requests
- Makes delimiters clear and easy to parse
- XML is often preferred for complex structured data and nested content
- JSON works well for simple key-value pairs and lists
- Always tell the LLM which format to use in its response
- Example: “Please respond in XML format with clear section tags”
Iterative Prompt Development
- Use an LLM to generate and improve your prompts
- Instead of trying to write a detailed prompt from scratch, write a simpler prompt that calls out high-level requirements
- Ask your LLM to generate a detailed, step-by-step process
- Revise the prompt iteratively until it’s perfect
- Test with edge cases and different scenarios
Version Control and Workflow
Use Git for Checkpoints
- Create frequent git commits as you generate code or content
- As you generate code (or other text) and it looks acceptable, commit it to git
- Makes rollback easy if something goes wrong
- Use descriptive commit messages like “feat: add user authentication logic”
- Consider using conventional commit formats for consistency
- Create branches for experimental changes
Quality Assurance
Multi-LLM Code Review
- Review your code with a different LLM than the one that generated it
- Each LLM has different strengths and sees things differently
- Use a different LLM to review your code for fresh perspective
- Ask it for bullet points for each finding and use that as a checklist
- Focus review on:
- Correctness and logic
- Performance and efficiency
- Idiomatic code patterns
- Security vulnerabilities
- Documentation and readability
Best Practices Summary
- Start simple, then iterate and refine
- Document your process and save your work
- Use multiple LLMs for different perspectives
- Leverage version control for safe experimentation
- Structure your inputs and outputs clearly
- Always review and validate generated content