本文由 AI 分析生成
建立時間： 2026-03-28 來源： https://www.seangoedecke.com/how-i-use-llms/
Summary
Sean Goedecke (GitHub staff engineer, former Copilot team) provides a concrete breakdown of how he uses LLMs: mostly as smart autocomplete and for throwaway research code, not for production business logic or technical writing. Notable for its honest “I don’t use LLMs for X” list.
Sean Goedecke（GitHub 資深工程師，前 Copilot 團隊）具體說明了如何使用 LLM：主要作為智能自動補全和一次性研究代碼，而非用於生產業務邏輯或技術寫作。以其誠實的「我不使用 LLM 做 X」清單而著稱。
Key Points
- Production code: Copilot for boilerplate/autocomplete only; rarely for business logic; more reliance in unfamiliar languages (operates at “smart intern baseline”)
- Throwaway research code: 2-4x faster with LLMs; perfect fit because maintainability doesn’t matter
- Learning new domains: LLM as on-demand tutor; ask follow-up questions, feed back notes for review; low hallucination risk in well-understood domains
- Bug fixes: last resort only; tries once and moves on if the LLM fails; currently better at bug-hunting than AI models
- Writing: never lets LLM write documents; uses it only for proofreading, ignores stylistic suggestions
- Does NOT use LLMs for: writing whole PRs in familiar areas, ADRs/technical writing, large codebase research
Insights
The “throwaway code” vs “production code” distinction is the most practically useful framework in this article — LLMs excel at the former (no maintenance concerns, any working code is good enough) and are risky for the latter (maintainability, idiomaticity, long-term correctness matter). The “smart intern baseline in unfamiliar languages” framing sets realistic expectations: useful for making tactical changes, but still requires domain expert review. The explicit “I ignore stylistic suggestions” note reflects well-calibrated trust: LLMs are better at typo detection than stylistic judgment.
Connections
Raw Excerpt
LLMs excel at writing code that works that doesn’t have to be maintained. Non-production code that’s only run once (e.g. for research) is a perfect fit for this. I would say that my use of LLMs here meant I got this done 2x-4x faster than if I’d been unassisted.