How I use LLMs as a staff engineer

The post describes a practical, limited way of using LLMs in staff-engineer work. The author relies on them for boilerplate, unfamiliar tactical changes, throwaway research code, learning new topics, occasional bug hunting, and checking long-form writing for typos and logic errors. He avoids using them for work he can do better himself, such as production logic in familiar areas, full technical drafts, and large-codebase research.

Reading notes

#

• Uses Copilot completions for code, mostly for boilerplate such as arguments and types
• Accepts more LLM help in unfamiliar languages or systems, but still asks a subject-matter expert to review the change
• Uses LLMs heavily for one-off research code that only needs to work once and does not need maintenance
• Treats the model as an on-demand tutor for learning new domains, asking follow-up and self-check questions
• Feeds learning notes back to the LLM for review and correction
• Uses Copilot chat as a last resort when stuck on a bug, usually only once and without much iteration
• Uses LLMs to review drafts for typos and logic mistakes, but does not let them write the documents
• Avoids using LLMs for full PRs in familiar areas, ADRs, and research in large codebases