My tips for using LLM agents to create software

Summary

This post details my experiences creating software with LLM coding agents, emphasizing that what you do with AI agents is ‘creation’, not just 'coding,' and sharing what worked for me. This is not 'The One True Path To AI Success.'tl;dr:I’m not a professional developer, just a hobbyist with aspirationsI wanted to accomplish a coding project beyond my skill level and have been experimenting with agentic coding tools for several months (spoiler: mostly success)You should use Anthropic’s Claude Sonnet model for complex coding tasks.Experiment with various agents and models; be adaptable as the field evolves quickly. I prefer Claude Code and Roo Code at this time.If you are a heavy user, you should use pay-as-you go pricing; TANSTAAFL.If you are a light user, use your favorite free or “comes-with-my-monthly-subscription” chatbot and whatever model it comes with. I expand on what “light” means later but think “not very much at all, bash one-liners, single-file python scripts, etc.”Part 1 of this post has background and describes how I chose agent, model, and subscription type.Mad AI SkillzI’ve learned a lot of tips and tricks (maybe even “skills”?) in getting more productive output from agentic AI and I wanted to share. This was actually the main point of me writing a blog post. I hope this helps you. I hope you share your tips with me and others.Context“Context” refers to the agent's short-term memory, including system prompts, standing instructions, and provided information, which the model uses for reasoning.Coding agents are really good at getting additional context if needed - they can add source code files, they can invoke local tools to extract particular information, and they’ll send that up to the model if needed.How the agent gets additional context is really important, as it significantly affects your experience. If you use a chat-based agent, you’ll have to paste or upload files manually. Some IDE-integrated agents require to to explicitly add files to contex...