Back to Details

planning-workflow

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Planning Workflow — The Foundation of Agentic Development

规划工作流 — Agentic开发的基础

Core Philosophy: "Planning tokens are a lot fewer and cheaper than implementation tokens."
The models are far smarter when reasoning about a detailed plan that fits within their context window. This is the key insight behind spending 80%+ of time on planning.
核心理念："规划所用的token数量更少、成本更低，远低于实现阶段的token消耗。"
当模型在其上下文窗口内针对详细规划进行推理时，表现会聪明得多。这就是我们将80%以上时间投入规划的核心原因。

Why Planning Matters

规划的重要性

Before burning tokens with a big agent swarm:
  • Measure twice, cut once — becomes "Check your plan N times, implement once"
  • A very big, complex markdown plan is still shorter than a few substantive code files
  • Front-loading human input in planning enables removing yourself from implementation
  • The code will be written ridiculously quickly when you start enough agents with a solid plan
在动用大量Agent集群消耗token之前：
  • 量两次，切一次 —— 在这里变成“多次检查你的规划，一次完成实现”
  • 一份非常庞大、复杂的Markdown规划，仍然比几个有实质内容的代码文件更短
  • 在规划阶段前置人工输入，能让你在实现阶段无需参与
  • 当你基于完善的规划启动足够多的Agent时，代码编写速度会快得惊人

The Planning Process (Overview)

规划流程（概述）

┌──────────────────────────────────────────────────────────────┐
│  1. INITIAL PLAN (GPT Pro / Opus 4.5 in web app)             │
│     └─► Explain goals, intent, workflows, tech stack         │
├──────────────────────────────────────────────────────────────┤
│  2. ITERATIVE REFINEMENT (GPT Pro Extended Reasoning)        │
│     └─► 4-5 rounds of revision until steady-state            │
├──────────────────────────────────────────────────────────────┤
│  3. MULTI-MODEL BLENDING (Optional but recommended)          │
│     └─► Gemini3 Deep Think, Grok4 Heavy, Opus 4.5           │
│     └─► GPT Pro as final arbiter                             │
├──────────────────────────────────────────────────────────────┤
│  4. CONVERT TO BEADS (Claude Code + Opus 4.5)                │
│     └─► Self-contained tasks with dependency structure       │
├──────────────────────────────────────────────────────────────┤
│  5. POLISH BEADS (6+ rounds until steady-state)              │
│     └─► Cross-model review, never oversimplify               │
└──────────────────────────────────────────────────────────────┘
┌──────────────────────────────────────────────────────────────┐
│  1. INITIAL PLAN (GPT Pro / Opus 4.5 in web app)             │
│     └─► Explain goals, intent, workflows, tech stack         │
├──────────────────────────────────────────────────────────────┤
│  2. ITERATIVE REFINEMENT (GPT Pro Extended Reasoning)        │
│     └─► 4-5 rounds of revision until steady-state            │
├──────────────────────────────────────────────────────────────┤
│  3. MULTI-MODEL BLENDING (Optional but recommended)          │
│     └─► Gemini3 Deep Think, Grok4 Heavy, Opus 4.5           │
│     └─► GPT Pro as final arbiter                             │
├──────────────────────────────────────────────────────────────┤
│  4. CONVERT TO BEADS (Claude Code + Opus 4.5)                │
│     └─► Self-contained tasks with dependency structure       │
├──────────────────────────────────────────────────────────────┤
│  5. POLISH BEADS (6+ rounds until steady-state)              │
│     └─► Cross-model review, never oversimplify               │
└──────────────────────────────────────────────────────────────┘

Phase 1: Creating the Initial Plan

第一阶段：创建初始规划

Where to Write It

在哪里编写

Use GPT Pro with Extended Reasoning in the web app. No other model can touch Pro on the web when dealing with input that fits its context window.
Alternative: Claude Opus 4.5 in the webapp is also good for initial plans.
在网页应用中使用开启Extended Reasoning的GPT Pro。当处理的输入内容在其上下文窗口范围内时，没有其他模型能比得上网页版的GPT Pro。
**替代方案：**网页应用中的Claude Opus 4.5也适用于创建初始规划。

What to Include

应包含的内容

  1. Goals and Intent — What you're really trying to accomplish
  2. Workflows — How the final software should work from the user's perspective
  3. Tech Stack — Be specific (e.g., "TypeScript, Next.js 16, React 19, Tailwind, Supabase")
  4. Architecture Decisions — High-level structure and patterns
  5. The "Why" — The more the model understands your end goal, the better it performs
You don't even need to write the initial markdown plan yourself. You can write that with GPT Pro, just explaining what it is you want to make.
  1. 目标与意图 —— 你真正想要实现的是什么
  2. 工作流 —— 最终软件从用户视角看应如何运作
  3. 技术栈 —— 要具体（例如："TypeScript, Next.js 16, React 19, Tailwind, Supabase"）
  4. 架构决策 —— 高层结构与模式
  5. “为什么” —— 模型对你的最终目标理解得越透彻，表现就越好
你甚至不需要自己编写初始Markdown规划。你可以通过GPT Pro来生成，只需向它说明你想要制作的内容即可。

Phase 2: Iterative Refinement

第二阶段：迭代优化

THE EXACT PROMPT — Plan Review (GPT Pro Extended Reasoning)

精准提示词 —— 规划评审（开启Extended Reasoning的GPT Pro）

Paste your entire markdown plan into GPT Pro with Extended Reasoning enabled and use this EXACT prompt:
Carefully review this entire plan for me and come up with your best revisions in terms of better architecture, new features, changed features, etc. to make it better, more robust/reliable, more performant, more compelling/useful, etc. For each proposed change, give me your detailed analysis and rationale/justification for why it would make the project better along with the git-diff style change versus the original plan shown below:

<PASTE YOUR EXISTING COMPLETE PLAN HERE>
将你的完整Markdown规划粘贴到开启Extended Reasoning的GPT Pro中，并使用以下精准提示词：
Carefully review this entire plan for me and come up with your best revisions in terms of better architecture, new features, changed features, etc. to make it better, more robust/reliable, more performant, more compelling/useful, etc. For each proposed change, give me your detailed analysis and rationale/justification for why it would make the project better along with the git-diff style change versus the original plan shown below:

<PASTE YOUR EXISTING COMPLETE PLAN HERE>

THE EXACT PROMPT — Integration (Claude Code)

精准提示词 —— 整合优化（Claude Code）

After GPT Pro finishes (may take 20-30 minutes for complex plans), paste the output into Claude Code with this EXACT prompt:
OK, now integrate these revisions to the markdown plan in-place; use ultrathink and be meticulous. At the end, you can tell me which changes you wholeheartedly agree with, which you somewhat agree with, and which you disagree with:

```[Pasted text from GPT Pro]```
GPT Pro完成后（复杂规划可能需要20-30分钟），将输出内容粘贴到Claude Code中，并使用以下精准提示词：
OK, now integrate these revisions to the markdown plan in-place; use ultrathink and be meticulous. At the end, you can tell me which changes you wholeheartedly agree with, which you somewhat agree with, and which you disagree with:

```[Pasted text from GPT Pro]```

Repeat Until Steady-State

重复直到进入稳定状态

  • Start fresh ChatGPT conversations for each round
  • After 4-5 rounds, suggestions become very incremental
  • You'll see massive improvements from v2 to v3, continuing to the end
  • This phase can take 2-3 hours for complex features — this is normal
  • 每一轮都开启新的ChatGPT对话
  • 经过4-5轮后，建议的修改会变得非常细微
  • 从版本2到版本3你会看到巨大的改进，这种提升会持续到最后
  • 对于复杂功能，这个阶段可能需要2-3小时 —— 这是正常的

Phase 3: Multi-Model Blending (Advanced)

第三阶段：多模型融合（进阶）

Why Blend Models

为什么要融合模型

Different models have different strengths. Blending gets "best of all worlds."
不同的模型有不同的优势。融合模型可以实现“集各家之长”。

The Process

流程

  1. Get competing plans from Gemini3 (Deep Think), Grok4 Heavy, and Opus 4.5
  2. Use GPT Pro as final arbiter
  1. 从Gemini3（Deep Think）、Grok4 Heavy和Opus 4.5获取竞争性规划
  2. 以GPT Pro作为最终的决策方

THE EXACT PROMPT — Multi-Model Blend

精准提示词 —— 多模型融合

I asked 3 competing LLMs to do the exact same thing and they came up with pretty different plans which you can read below. I want you to REALLY carefully analyze their plans with an open mind and be intellectually honest about what they did that's better than your plan. Then I want you to come up with the best possible revisions to your plan (you should simply update your existing document for your original plan with the revisions) that artfully and skillfully blends the "best of all worlds" to create a true, ultimate, superior hybrid version of the plan that best achieves our stated goals and will work the best in real-world practice to solve the problems we are facing and our overarching goals while ensuring the extreme success of the enterprise as best as possible; you should provide me with a complete series of git-diff style changes to your original plan to turn it into the new, enhanced, much longer and detailed plan that integrates the best of all the plans with every good idea included (you don't need to mention which ideas came from which models in the final revised enhanced plan):

[Paste competing plans here]
I asked 3 competing LLMs to do the exact same thing and they came up with pretty different plans which you can read below. I want you to REALLY carefully analyze their plans with an open mind and be intellectually honest about what they did that's better than your plan. Then I want you to come up with the best possible revisions to your plan (you should simply update your existing document for your original plan with the revisions) that artfully and skillfully blends the "best of all worlds" to create a true, ultimate, superior hybrid version of the plan that best achieves our stated goals and will work the best in real-world practice to solve the problems we are facing and our overarching goals while ensuring the extreme success of the enterprise as best as possible; you should provide me with a complete series of git-diff style changes to your original plan to turn it into the new, enhanced, much longer and detailed plan that integrates the best of all the plans with every good idea included (you don't need to mention which ideas came from which models in the final revised enhanced plan):

[Paste competing plans here]

Real-World Examples

实际案例

Example Plan Documents

规划文档示例

ProjectPlan Link
CASS Memory SystemPLAN_FOR_CASS_MEMORY_SYSTEM.md
CASS GitHub Pages ExportPLAN_TO_CREATE_GH_PAGES_WEB_EXPORT_APP.md
项目规划链接
CASS Memory SystemPLAN_FOR_CASS_MEMORY_SYSTEM.md
CASS GitHub Pages ExportPLAN_TO_CREATE_GH_PAGES_WEB_EXPORT_APP.md

Example AGENTS.md Files

AGENTS.md文件示例

Project TypeLink
NextJS webapp + TypeScript CLIbrenner_bot/AGENTS.md
Bash script projectrepo_updater/AGENTS.md
项目类型链接
NextJS webapp + TypeScript CLIbrenner_bot/AGENTS.md
Bash script projectrepo_updater/AGENTS.md

What Makes a Great Plan

优秀规划的特质

Good vs. Great

良好 vs 优秀

Good PlanGreat Plan
Describes what to buildExplains WHY you're building it
Lists featuresDetails user workflows and interactions
Mentions tech stackJustifies tech choices with tradeoffs
Has tasksHas tasks with dependencies and rationale
~500 lines~3,500+ lines after refinement
良好规划优秀规划
描述要构建的内容解释构建的原因
列出功能详细说明用户工作流与交互
提及技术栈结合权衡取舍论证技术选择的合理性
包含任务包含带有依赖关系与论证的任务
~500行经过优化后达到~3500+行

Essential Elements

核心要素

  1. Self-contained — Never need to refer back to external docs
  2. Granular — Break complex features into specific subtasks
  3. Dependency-aware — What blocks what?
  4. Justified — Include reasoning, not just instructions
  5. User-focused — How does each piece serve the end user?
  1. 自包含 —— 无需参考外部文档
  2. 粒度精细 —— 将复杂功能拆分为具体的子任务
  3. 感知依赖 —— 哪些任务会阻塞其他任务？
  4. 有论证依据 —— 不仅包含指令，还要包含推理过程
  5. 以用户为中心 —— 每一部分如何服务于终端用户？

Common Mistakes

常见误区

  1. Starting implementation too early — 3 hours of planning saves 30 hours of rework
  2. Single-round review — You continue to get improvements even at round 6+
  3. Not using GPT Pro — Extended Reasoning is uniquely good for this
  4. Skeleton-first coding — One big comprehensive plan beats incremental coding
  5. Losing context — Convert plans to beads so agents don't need the original
  1. 过早开始实现 —— 3小时的规划能节省30小时的返工时间
  2. 仅进行单轮评审 —— 即使到第6轮及以后，你仍能获得改进建议
  3. 未使用GPT Pro —— Extended Reasoning在此场景下的表现独一无二
  4. 先搭建骨架再编码 —— 一份全面的详细规划优于增量式编码
  5. 丢失上下文 —— 将规划转换为beads，让Agent无需依赖原始规划

FAQ

常见问题

Q: Shouldn't I code a skeleton first? A: You get a better result faster by creating one big comprehensive, detailed, granular plan. That's the only way to get models to understand the entire system at once. Once you start turning it into code, it gets too big to understand.
Q: What about problems I didn't anticipate? A: Finding the flaws and fixing them is the whole point of all the iterations and blending in feedback from all the frontier models. If you follow the procedure using those specific models and prompts, after enough rounds, you will have an extremely good plan that will "just work." After implementing v1, you create another plan for v2. Nothing says you can only do one plan.
Q: How do I divide tasks for agents? A: Each agent uses bv to find the next optimal bead and marks it in-progress. Distributed, robust, fungible agents.
Q: Do agents need specialization? A: No. Every agent is fungible and a generalist. They all use the same base model and read the same AGENTS.md. Simply telling one it's a "frontend agent" doesn't make it better at frontend.
Q: Which tech stack should I use? A: This is part of the "pre-planning" phase. Usually I already know based on project type:
  • Web app: TypeScript, Next.js 16, React 19, Tailwind, Supabase (performance-critical parts in Rust compiled to WASM)
  • CLI tool: Golang or Rust if very performance critical
  • If unsure, do a deep research round with GPT Pro or Gemini3 to study libraries and get suggestions.
Q: Should design decisions be in markdown or beads? A: The beads themselves can and should contain this markdown. You can have long descriptions/comments inside the beads—they don't need to be short bullet point type entries.
Q: 我应该先编码搭建骨架吗？ A: 创建一份庞大、全面、详细、粒度精细的规划，能让你更快获得更好的结果。这是让模型一次性理解整个系统的唯一方法。一旦你开始将其转换为代码，内容就会变得过于庞大，模型无法完全理解。
Q: 对于我没有预料到的问题该怎么办？ A: 发现缺陷并修复它们，正是所有迭代过程以及融合前沿模型反馈的核心目的。如果你按照流程使用这些特定模型和提示词，经过足够多轮次后，你会得到一个极其优秀的规划，它将“直接生效”。在实现版本1后，你可以为版本2创建新的规划。并没有规定你只能做一次规划。
Q: 我该如何为Agent分配任务？ A: 每个Agent使用bv找到下一个最优的bead，并将其标记为进行中。实现分布式、健壮、可替代的Agent机制。
Q: Agent需要专业化吗？ A: 不需要。每个Agent都是可替代的通用型Agent。它们都使用相同的基础模型，并读取相同的AGENTS.md文件。仅仅告诉某个Agent它是“前端Agent”并不会让它更擅长前端工作。
Q: 我应该使用哪种技术栈？ A: 这属于“预规划”阶段的内容。通常我会根据项目类型提前确定：
  • **Web应用：**TypeScript, Next.js 16, React 19, Tailwind, Supabase（性能关键部分使用编译为WASM的Rust）
  • **CLI工具：**如果对性能要求极高，使用Golang或Rust
  • 如果不确定，使用GPT Pro或Gemini3进行深度研究，调研各类库并获取建议。
Q: 设计决策应该放在Markdown还是beads中？ A: beads本身可以且应该包含这些Markdown内容。你可以在beads中加入长篇描述/注释——它们不需要是简短的要点式条目。

Best Practices Guides

最佳实践指南

Keep best practices guides in your project folder and reference them in AGENTS.md:
Have Claude Code search the web and update them to latest versions.
在你的项目文件夹中保存最佳实践指南，并在AGENTS.md中引用它们：
让Claude Code搜索网络并将其更新到最新版本。

Complete Prompt Reference

完整提示词参考

GPT Pro — Plan Review

GPT Pro —— 规划评审

Carefully review this entire plan for me and come up with your best revisions in terms of better architecture, new features, changed features, etc. to make it better, more robust/reliable, more performant, more compelling/useful, etc. For each proposed change, give me your detailed analysis and rationale/justification for why it would make the project better along with the git-diff style change versus the original plan shown below:

<PASTE YOUR EXISTING COMPLETE PLAN HERE>
Carefully review this entire plan for me and come up with your best revisions in terms of better architecture, new features, changed features, etc. to make it better, more robust/reliable, more performant, more compelling/useful, etc. For each proposed change, give me your detailed analysis and rationale/justification for why it would make the project better along with the git-diff style change versus the original plan shown below:

<PASTE YOUR EXISTING COMPLETE PLAN HERE>

Claude Code — Integrate Revisions

Claude Code —— 整合优化

OK, now integrate these revisions to the markdown plan in-place; use ultrathink and be meticulous. At the end, you can tell me which changes you wholeheartedly agree with, which you somewhat agree with, and which you disagree with:

```[Pasted text from GPT Pro]```
OK, now integrate these revisions to the markdown plan in-place; use ultrathink and be meticulous. At the end, you can tell me which changes you wholeheartedly agree with, which you somewhat agree with, and which you disagree with:

```[Pasted text from GPT Pro]```

GPT Pro — Multi-Model Blend

GPT Pro —— 多模型融合

I asked 3 competing LLMs to do the exact same thing and they came up with pretty different plans which you can read below. I want you to REALLY carefully analyze their plans with an open mind and be intellectually honest about what they did that's better than your plan. Then I want you to come up with the best possible revisions to your plan (you should simply update your existing document for your original plan with the revisions) that artfully and skillfully blends the "best of all worlds" to create a true, ultimate, superior hybrid version of the plan that best achieves our stated goals and will work the best in real-world practice to solve the problems we are facing and our overarching goals while ensuring the extreme success of the enterprise as best as possible; you should provide me with a complete series of git-diff style changes to your original plan to turn it into the new, enhanced, much longer and detailed plan that integrates the best of all the plans with every good idea included (you don't need to mention which ideas came from which models in the final revised enhanced plan):

[Paste competing plans here]
I asked 3 competing LLMs to do the exact same thing and they came up with pretty different plans which you can read below. I want you to REALLY carefully analyze their plans with an open mind and be intellectually honest about what they did that's better than your plan. Then I want you to come up with the best possible revisions to your plan (you should simply update your existing document for your original plan with the revisions) that artfully and skillfully blends the "best of all worlds" to create a true, ultimate, superior hybrid version of the plan that best achieves our stated goals and will work the best in real-world practice to solve the problems we are facing and our overarching goals while ensuring the extreme success of the enterprise as best as possible; you should provide me with a complete series of git-diff style changes to your original plan to turn it into the new, enhanced, much longer and detailed plan that integrates the best of all the plans with every good idea included (you don't need to mention which ideas came from which models in the final revised enhanced plan):

[Paste competing plans here]