My LLM Coding Workflow Going Into 2026 - Addy Osmani
Key Arguments
1. Planning Precedes Code — Begin with detailed spec.md containing requirements, architecture decisions, data models, and testing strategy. "Waterfall in 15 minutes."
2. Iterative Task Decomposition — Break work into small chunks. Large monolithic requests produce inconsistent results.
3. Extensive Context Provision — Context packing: relevant code, docs, constraints, pitfalls. Tools like gitingest, repo2txt. Claude Skills as reusable instruction packages.
4. Model Selection — "Model musical chairs" — switch between Claude, Gemini, others when one hits limitations.
5. Human Oversight as Non-Negotiable — Treat AI code like junior developer contributions. "You absolutely have to test what it writes."
6. Version Control as Safety Net — Frequent, granular commits. Small commits make pinpointing problems easier.
7. Behavioral Customization — CLAUDE.md files, project-specific rules, coding style examples.
8. Automation as Force Multiplier — CI/CD pipelines, automated testing create feedback loops.
Notable Quotes
"At Anthropic, engineers adopted Claude Code so heavily that ~90% of the code for Claude Code is written by Claude Code itself"
"AI coding assistants are incredible force multipliers, but the human engineer remains the director of the show"
"The LLM is an assistant, not an autonomously reliable coder...I am the senior dev"
Main Conclusions
- Classical software engineering practices become more critical, not obsolete, when incorporating AI
- AI handles mechanical tasks while developers manage architecture, design, and oversight
- Context and communication are paramount — AI performance correlates with information provided
- Accountability remains human
- Structured workflows outperform ad-hoc use