本文由 AI 分析生成
建立時間： 2026-03-24 來源： https://steipete.me/posts/2025/optimal-ai-development-workflow
Summary
Peter Steinberger’s production-tested Claude Code workflow after months of iteration: Ghostty terminal as primary environment, Claude Code as the main driver, VS Code only for code lookup, and GPT-5 for plan review. The core thesis is that less tooling means more productivity — he removed even his last MCP server because Claude reaching for Playwright polluted the context unnecessarily.
Peter Steinberger 歷經數月反覆調整後的實戰工作流程：以 Ghostty 為主要終端環境、Claude Code 為主力、VS Code 僅用於查閱程式碼、GPT-5 負責計畫審查。核心主張是「越少工具越有生產力」——他甚至移除了最後一個 MCP，因為 Claude 在不必要的情況下啟動 Playwright 會汙染 context。
Key Points
- Ghostty over VS Code terminal: VS Code terminal freezes when pasting large text; Ghostty is stable
- 4 Claude instances simultaneously: 3840×1620 monitor lets him run 4 agents + Chrome without window management
- Agent count by task type: 1-2 for regular work, ~4 for cleanup/tests/UI; depends on blast radius
- Plan mode for big tasks: small prompts for small tasks, write bigger plans to file then have GPT-5 review before executing
- Testing in same context: automated tests aren’t great, but asking the model to write tests in the same session reliably surfaces issues
- CLI-first services: Vercel, psql, gh, Axiom — one line in CLAUDE.md is enough for agents to use them
- Removed last MCP: Claude would spin up Playwright unprompted when reading code was faster and cheaper on context
Insights
- The “blast radius” framing for deciding agent count is practical: the more a task can affect unrelated code, the fewer parallel agents you run — it’s a risk management heuristic, not a capability limit
- Writing plans to a file and sending to GPT-5 for review before executing them is an interesting meta-strategy: use a more literal, precise model to sanity-check plans from a more creative model
- The MCP removal is counterintuitive but consistent with context economy: every tool a model can use is a tool it will sometimes misuse, burning tokens and steering the session off course
- “Claude often makes a mess but it’s equally great at refactoring and cleaning up” — alternating between generation and cleanup passes to manage debt is a discipline, not a workaround
- Steering running agents matters: background agents are harder to correct mid-drift; keeping humans in the loop during complex tasks is still the productive choice
Connections
- Clippings-lessons-from-building-claude-code-how-we-use-skills — Anthropic’s internal perspective on skill/tool design complements this practitioner view
- Clippings-你不知道的-claude-code — deeper engineering internals behind the Claude Code workflow
- Clippings-6-prompt-techniques — prompt strategies that pair well with this agentic workflow
Raw Excerpt
“Even removed my last MCP, since Claude sometimes would go off spinning up Playwright unasked when it could simply read the code — which is faster and pollutes the context less.”