本文由 AI 分析生成
建立時間： 2026-03-28 來源： https://harper.blog/2025/02/16/my-llm-codegen-workflow-atm/
Summary
Harper Reed’s LLM codegen workflow (February 2025): three discrete loops — brainstorm/spec with conversational LLM (ChatGPT 4o), plan/document the spec, then execute in small steps using coding LLM. The workflow emphasizes iteration, small steps, and managing the context boundary between planning and implementation.
Harper Reed 的 LLM 代碼生成工作流程（2025 年 2 月）：三個離散的循環——使用對話式 LLM（ChatGPT 4o）進行頭腦風暴/規格設計，規劃/記錄規格，然後使用代碼生成 LLM 小步驟執行。工作流程強調迭代、小步驟，以及管理規劃和實現之間的上下文邊界。
Key Points
- Three loops: (1) Idea honing with conversational LLM (one question at a time → comprehensive spec); (2) plan documentation; (3) execution in small steps with coding LLM
- Greenfield workflow: brainstorm via chat → compile to developer-ready spec → execute → iterate
- Legacy/modern code: read existing code first, understand patterns, then introduce AI coding against established patterns
- Spec prompt: “ask me one question at a time” → forces thorough iterative spec development vs. one-shot prompting
- Compile step: “compile our findings into a comprehensive, developer-ready specification” → clean handoff to coding LLM
- Small steps matter: avoid large context requests; break implementation into small, executable chunks
Insights
The three-loop structure is the most important insight: planning and implementation require different LLM behaviors (broad/creative vs. precise/focused), and using them in separate sessions with a clean spec as the handoff artifact is better than one long “design and code this for me” session. The one-question-at-a-time spec elicitation prompt is genuinely effective — it produces a better spec than asking for everything at once because it forces sequential consideration of each dimension. The “this probably won’t work in 2 weeks” caveat reflects how fast the space is moving.
Connections
Raw Excerpt
Brainstorm spec, then plan a plan, then execute using LLM codegen. Discrete loops. The idea is that you keep the planning and execution in separate contexts so the model stays focused on each task.