How I write software with LLMs
8 hours ago
- #Workflow Optimization
- #LLM
- #Software Development
- The author has shifted from traditional programming to using LLMs for creating projects, finding it more aligned with their passion for making things.
- LLMs have enabled the author to write software with a low defect rate, surpassing manual coding in reliability while maintaining system understanding.
- Engineering skills have evolved; system architecture and usability choices are now more critical than writing correct code.
- Different people achieve varying results with LLMs, suggesting that interaction methods significantly impact outcomes.
- The author's workflow involves an architect (strong model), a developer (efficient model), and reviewers (multiple models) to ensure quality.
- Projects built using this method include Stavrobot (secure LLM assistant), Middle (voice note recorder), Sleight of hand (artistic clock), and Pine Town (multiplayer canvas).
- The harness used must support multiple models from different companies and allow custom agents to autonomously call each other.
- A detailed example of adding email support to Stavrobot illustrates the workflow, involving planning, implementation, review, and iterative improvements.
- Key benefits include high reliability, deep system familiarity without reading every line of code, and the ability to quickly implement and refine features.
- Challenges include maintaining architectural oversight and avoiding compounding errors when unfamiliar with underlying technologies.