Hasty Briefsbeta

My tips for using LLM agents to create software

a day ago
  • #software development
  • #LLM tips
  • #AI coding agents
  • Use Anthropic’s Claude Sonnet model for complex coding tasks.
  • Experiment with various agents and models; be adaptable as the field evolves quickly.
  • Use pay-as-you-go pricing for heavy users; free or subscription-based chatbots for light users.
  • Provide the AI model with relevant context to improve results, but avoid irrelevant content.
  • Store context in a fixed place and update standing prompts to guide the agent.
  • Reinforce instructions or add special instructions in comments within source code files.
  • Monitor context size to avoid 'out of context' errors and session failures.
  • Use tools to extract only relevant information from large files to save tokens.
  • Start new sessions instead of compaction to save tokens and maintain focus.
  • Generate context upfront for challenging areas to assist the agent.
  • Communicate clearly with agents; explanations are only needed if they reveal new goals or guidelines.
  • Have a design in mind and document it to guide the agent and avoid random refactoring.
  • Use 'think deeply' or 'analyze deeply' to enter deep reasoning mode for planning.
  • Collaborate with the agent to create and update planning and tracking documents.
  • Enable inter-agent communication to resolve integration problems between different codebases.
  • Implement extensive logging to help the agent diagnose issues and understand program flow.
  • Use defensive prompts to ensure the agent completes tasks without errors.
  • Prevent agents from disabling tests; ensure they diagnose and fix failures.
  • Use git defensively: frequent commits, new branches for big changes, and prevent unauthorized git actions.
  • Break big tasks into small tasks and use TODO lists to track progress.
  • Interrupt the agent with corrections if it starts doing something undesirable.
  • Review and refine agent-created tests to ensure they catch bugs.
  • Build custom tools to help the agent perform complex tasks more efficiently.
  • Avoid asking the agent for opinions on design as it can be inconsistent.
  • Resist the urge to implement all suggested improvements; focus on solving existing problems.