Signal to BuildSIGNAL TO BUILD中文Owner sign in ↗
← Knowledge Notes
AI Engineering ／ Knowledge note · English

Wang Kai's Claude Code Terminal Workflow Methodology (Full Version)

The complete engineering accumulation of Wang Kai's primary workflow for 2026. Terminal window as Agent / multi-model hybrid orchestration / Skill system / Feishu CLI collaboration bus / three-layer configuration hierarchy / Chrome MCP controlling real browsers / 10-minute patrol for long tasks / counter-consensus view "Claude subscription is the cheapest"

Source collection：Wang Kai · Published here：2026-09-26

Claude CodeAgent SkillsEngineering workflows

Wang Kai's primary workflow starting from 2025-12. Refined from real-world practice of the 1119 post + a 92KB overview document.Core Philosophy:The directory is the Agent's body, CLAUDE.md is its persona,.mcp.json is its tools, and feedback.md is its growth record.

In one sentence:Exponentially increase token consumption volume, consumption quality, and consumption efficiency——porting the division of labor/review/iteration logic of traditional internet engineering into AI collaboration.

1. Ultimate Vision

"In the future, everyone will have 10 computers, and every computer will have an Agent installed (Claude Code, OpenClaw, etc.), forming the standard organization of the AI era (1 person + 10+ Agents)"

2. Core Mental Models (memorize 8 sentences)

# Mental Model One Sentence
1 A terminal window is an Agent Not a sub-agent / Task — literally "open a new terminal and run claude"
2 One window, one function Content specialist / technical specialist / CTO review / CEO dispatch; the finer the split, the more token consumption doubles and the better the quality
3 A Claude subscription is the cheapest kimi API 1 task 1.5h = 15 RMB > amortized 200 USD monthly fee
4 Multi-model hybrid orchestration Claude Code as primary + Codex for CTO review + Gemini/Kimi as fallback
5 Skills are the Agent's wiki Internalize work into reusable assets; neither each window nor yourself 3 minutes later can see the full picture
6 10-minute patrol replaces complex Loops Agents will stop when they get clever; a simple 10-minute check makes it continue if it's not running
7 Give all data/information to the AI The more links and the more MCPs in context, the more useful the AI is
8 PUA Claude Iterate analysis → feedback → improvement over and over, and the prompt can eventually be polished to 6000+ words

3. Terminal Window as Agent

3.1 One window = one agent = one function

"First treat a window as a complete affair/project ➡️ split it into departments ➡️ split it into individual people ➡️ finally split it into a specific subdivided function. With each split, token consumption should double, and the quality of the resulting output gets better and better."

Common function splits:

  • Design / Content / SEO / Video Operations / Engineering / Review
  • Outbound link team / Content team / Product optimization team
  • CEO / CTO / Promotion

3.2 Engineering division of labor and review loop

graph LR
    A[内容专员 Skill A] --> D[Codex CTO 审查]
    B[技术专员 Skill B] --> D
    C[推广专员 Skill C] --> D
    D --> E[Claude Code 修复]
    E --> F[Codex 验收]
    F --> G[全局监控 + 强化学习]
    G --> H[复盘 Agent 持续优化 Skill]
    H --> A

1. Claude Code produces the product architecture:Do not base it on the current state of the existing product; first clarify the product goals, then have it produce the architecture + feature/logic details, then reverse-compare the current product's shortcomings and have it update (to avoid being led around by constant trivial issues).

2. Load Codex into Claude Code for CTO review:codex-plugin-cc Help build the full technical landscape (performance, etc.), review code → hand off to CC to resolve → Codex accepts the result.

3. Bug resolution closed loop:CC controls the browser to test the bug → resolves it → controls the browser again itself to verify whether the bug is truly resolved; loop this way to avoid infinite loops.

3.3 Multi-model hybrid orchestration in practice

新下载多个终端 → 每个终端开多窗口 → 每个窗口不同任务 → 每个终端分配不同大模型
(Claude Code、Codex、Gemini、Kimi 各跑各的)

Cost observation(Counter-consensus view):

  • kimi API same task 1.5h ≈ 15 RMB > amortized 200 USD monthly fee
  • Codex is cheaper than kimi API
  • Conclusion: If usage is high, subscriptions from first-tier vendors like Claude / Codex are the cheapest

3.4 paperclip / OpenClaw evaluation

Claude Code multi-window paperclipai/paperclip OpenClaw
Visualization ✅ View dynamics in terminal Framework encapsulation ❌ Not visualizable
Intervention cost Low High (frequent crashes) Heavily dependent on Skill
Recommendation ✅ ⚠️ Immature Use for cloud runs; slow locally

"In the early stage I highly recommend everyone to use visualization (that is, run OpenClaw on the computer at hand), and still most recommend Claude; wait until you are very familiar with AI's logic, error patterns, and deviation patterns before considering cloud servers and so on."

4. Three Fatal Pitfalls of Skill / Memory / Rules

4.1 Pitfall 1: Rules must specify "scope of application"

"This is so much like teaching a child: you tell them the rule 'don't get angry', and the premise of that rule is 'when you are in the wrong, don't lose your temper arbitrarily', but the child is very likely to say when being bullied that adults told me 'don't get angry, so I can't get angry'. For every rule and memory, it's best to clearly specify the 'scope of application'."

Bad example:Don't write "add tags later, okay" Good example:"When you post on Xiaohongshu, you must add tags" (platform-limited)

4.2 Pitfall 2: Memories get cross-wired

While talking about something in one window, it replies with a completely unrelated other project → check ~/.claude/CLAUDE.md whether too many project names are written

4.3 Pitfall 3: The more Skills, the more conflicts

Need to continuously organize the structure → Karpathy LLM Wiki thinking:Organize all files, build folders according to projects, have hierarchical relationships between files, be able to cross-reference each other, etc., similar to Wikipedia.

Wang Kai one-click upgrade prompt: tokrepo LLM Wiki Memory Upgrade Prompt

4.4 For precision → write Skills in code

"Various motion effects are often used in video production... If I tell Claude Code to follow them in text, it always takes all kinds of routes and cannot guarantee the output effect. Later I had it write code, put the knowledge points into code in the video skill, and the implementation path became relatively unique."

Principle:Text has high breadth, while code/mathematics has the least information decay. Scenarios requiring precision are very suitable for conversion into skills to ensure a unique effect.

5. Three-Layer Configuration Hierarchy (tested on v2.1.x)

Layer MCP configuration Skill path CLAUDE.md path Loading method
Enterprise (managed) (managed) /Library/Application Support/ClaudeCode/CLAUDE.md Highest priority
User (global) ~/.claude.json's mcpServers field ~/.claude/skills/<name>/SKILL.md ~/.claude/CLAUDE.md Inherited by all agents
Project ./.mcp.json ./.claude/skills/<name>/SKILL.md ./CLAUDE.md Only takes effect in this directory
Local ./.claude/settings.local.json — ./CLAUDE.local.md Not committed to git

5.1 Key rules

  • MCP write command:claude mcp add <name> -s user|project|local <command> [args...] Do not hand-edit JSON
  • Skill priority:Enterprise > Personal > Project > Plugin —— upper layers with the same name override lower layers
  • CLAUDE.md is concatenated, not overridden —— all content written in the global CLAUDE.md will appear in every agent → must be extremely minimal

5.2 4 questions to decide which layer to put it in

1. 这个能力绑定特定资源吗?(如某个浏览器端口)  → Project
2. 不同身份/agent 凭证不一样吗?(如各自 API key)→ Project
3. 不同身份/agent 行为不一样吗?(如不同语气)   → Project
4. 纯逻辑、所有 agent 都用同一份?               → User(全局共享)

5.3 Counterexamples (wrong in v1.0)

  • ❌ ~/.claude/mcp.json —— This path does not exist, global MCP is in the ~/.claude.json's mcpServers field
  • ❌ Registering chrome MCP at user level → cannot switch identities when doing matrix operations
  • ❌ Writing "I work on SillyTavern Online" in global CLAUDE.md → pollutes all projects
  • ❌ Skills with the same name under multiple identity directories → get overridden

6. Configuration Path Quick Reference (v2.1.x)

~/.claude.json                              ← 全局 MCP(mcpServers 字段)
~/.claude/CLAUDE.md                         ← 全局 CLAUDE.md(必须极简)
~/.claude/skills/<name>/SKILL.md            ← 全局 skill
~/.claude/settings.json                     ← 全局 hooks/permissions
./.mcp.json                                 ← 项目 MCP
./.claude/skills/<name>/SKILL.md            ← 项目 skill
./CLAUDE.md                                 ← 项目 CLAUDE.md
./.claude/settings.local.json               ← 本地(不入 git)
./CLAUDE.local.md                           ← 本地 CLAUDE.md

Key commands

# 加 MCP
claude mcp add <name> -s user|project|local <command> [args...]

# 看 MCP
claude mcp list
claude mcp get <name>

# Claude Code 内
/mcp     # 看 MCP server + 工具
/skills  # 看 skill 列表

7. Wang Kai's Own Tool Stack

7.1 chrome-mcp-proxy (William Wang's own)

"Killer Chrome MCP" — the official Google version defaults to headless browser; the author modified it to support user data + session reuse + no focus stealing.

  • Path:~/chrome-mcp-proxy/chrome-mcp-proxy.sh
  • Port pair:bash chrome-mcp-proxy.sh <PROXY_PORT> <CHROME_PORT>
  • Health check:curl http://127.0.0.1:9401/proxy/status
  • One-click install:tokrepo Chrome MCP Background Proxy
  • Verified environment: macOS + Chrome 147 + Claude Code v2.1.126+

7.2 cc-status-board (status bar)

Inspired by GSD (get-shit-done), solves 3 problems:

  1. Model information used by the current conversation
  2. Current conversation context length → reminds to compress in time
  3. Matches your installed AI assets (MCP, Skill, etc.) based on context —— matches the right AI assets in the right context

7.3 weixin-claude-router (WeChat → iTerm2 → Claude Code)

Scenario: About 10 terminal tabs are always open running CC; when away from the computer, remotely control Claude Code on iTerm2 via WeChat/Feishu.

7.4 tokrepo.com (AI asset hosting)

Positioning: "GitHub for everyone in the AI era" — hosts prompts, Skills, MCPs, memory files, rules, etc.

Entry:tokrepo.com/en

Published:

  • Chrome MCP Background Proxy
  • LLM Wiki Memory Upgrade Prompt
  • Remotion AI Skill (programmatic video production)
  • Weight loss tool (Feishu IM + Feishu Bitable + Claude Code)
  • CC Status Board

7.5 Feishu CLI (inter-Agent collaboration bus)

"Suppose I get 10 computers, each equipped with a Claude 200 USD subscription, how to efficiently organize them to collaborate with each other? Feishu CLI can achieve this efficiently: pull all computers/Claude Code into one Feishu organization, use Feishu CLI to write documents, messages, even calendars, etc. to carry out collaboration, reporting, etc., then set up a 'brain' machine to organize them all."

  • Two major sources of work information: Claude (conversations) + Feishu (documents)
  • Multi-Agent collaboration: same Feishu organization + Feishu CLI; humans only watch 1-3 "brain" computers
  • Feishu documents can pull dozens of articles via CLI as context (previously needed to have CC control the browser to read them)

7.6 Remotion AI Skill (programmatic video production)

Cost: "The cost of making videos with Claude Code is much lower than Jimeng etc.", can be combined with Jimeng premium membership + installing Jimeng CLI in CC, running automatically in multiple windows, with each YouTube account producing at least 10+ clips per day.

Effect breakthrough: instead of abstracting into rules, directly have CC imitate Dropbox explainer videos.

8. Computer Use vs Claude Code Terminal

Claude client Computer Use Claude Code terminal
Control scope Whole computer Terminal + MCP
Visualization Yes Yes (dynamic)
Onboarding threshold Low Medium
Use case Handling WeChat/local documents/Feishu Primary development
Note Cowork and Code run on different network environments Recommendation

WeChat resists AI control → buy a mouse simulator, let the AI control WeChat through the mouse simulator.

9. Debugging and Best Practices

9.1 Long-term task scheduling

"I have more and more tasks where I let Claude Code keep executing nonstop... but it always stops when it gets clever. At first I wanted to set a detailed scheduled loop task via /loop to keep it running all the time. Then I had a turning point in my thinking——simply set a 10-minute check scheduled loop task; if the Agent is not continuing to run, just make it continue. "

→ Conclusion: Abandon /loop, use a simple 10-minute patrol + make it continue if it's not running

9.2 Give all data/information to the AI

"To say it again: when you use Claude Code etc., the more data and more information you connect, the more useful your AI will be and the more it will help you."

  • Control the browser to get real information
  • Pull all documents via Feishu CLI
  • Build a self-hosted financial MCP to get accurate data (mandatory for investment scenarios)
  • Give complete securities trading order data, let CC do offensive/defensive drills

9.3 Common pitfalls (by frequency)

  1. Overlong context:Inevitably encountered; cc-status-board will remind you. Handling: have the original window write a summary → send it to a new window
  2. Chrome MCP disconnects:Open another window to take over
  3. Opening too many Agents at the same time → computer memory runs out:Control concurrency (Wang Kai long-term runs 6 projects / minimum 10 agents)
  4. AI makes repeated mistakes:Check whether the rules specify a "scope of application"
  5. More Skills lead to memory cross-contamination:Structure and organize using LLM Wiki thinking
  6. Anti-spam restrictions on new platform accounts:Follow the rules to warm up accounts, do not bypass them
  7. WeChat anti-AI control:Buy a mouse simulator
  8. Agent gets clever and stops long tasks on its own:10-minute patrol checks
  9. AgentMail SDK property names are unintuitive:.inbox_id / .preview
  10. Cowork and Code run on different network environments:Test Computer Use via Code
  11. Reddit/Medium protections cannot be bypassed:Choose platforms with weaker protections
  12. Don't be intimidated by entry barriers:Account bans / thinking it's too expensive / not knowing the terminal; in practice, a Claude subscription is the cheapest

9.4 chrome MCP tools are not called by the agent (v1.1 key point)

Symptom:After the agent sees the task,it takes a detour to write Node.js/Python scripts to call raw CDP, instead of calling mcp__chrome__* tools.

Cause:Claude Code v2.x's tool discovery mechanism does not eager-load MCPs registered only at the user level.

Forced registration method(Also register chrome in each agent project directory):

cd ~/agents/identity-A/publisher
claude mcp add chrome -s project bash $HOME/chrome-mcp-proxy/chrome-mcp-proxy.sh 9401 9222

And in the project CLAUDE.md state clearly: "Only use mcp__chrome__* tools; it is forbidden to use Bash to write scripts to detour and call CDP".

If it still takes a detour → send /mcp to confirm chrome is listed before assigning tasks.

10. Evolution path of token consumption (Wang Kai's personal experience)

Stage Tools Monthly expenditure Key changes
Cannot code Claude + ChatGPT < 100 USD Ordinary conversation
Using AI Coding Cursor, Qoder, Trae, v0, Claude, ChatGPT all open simultaneously ~ 200 USD Token volume ×10
VS Code + Claude Code By project (one folder per project) Increase Still constrained by project-based thinking
Terminal + Claude Code Break free from project-based thinking 125 → 250 → 200 USD (Claude) + Codex Token volume ×2 again
Multiple terminal windows + multiple models Claude Code, Codex, Gemini, Kimi each run independently Same as above Token volume ×3 again, "the number of windows open daily is comparable to web pages opened in a browser"
Multi-browser concurrency + Skill-ification + recurring tasks Chrome MCP × N + Skill + Feishu CLI + /10min patrol checks Claude subscription fully utilized + API billing at the 200 USD tier, with 50 USD used up in 3 hours Split work into dozens of Skills: "design/content/SEO/video/technical/review"

Measurement dimensions:Token consumption volume / consumption quality / consumption efficiency.

11. Work mindset shift (important)

"Subconscious rules" that Wang Kai repeatedly emphasizes:

  • ✅ Avoid using the browser yourself as much as possible —— First reaction is to let CC control the browser
  • ✅ Avoid doing it yourself as much as possible —— First reaction is "Why should a human do this instead of AI?"
  • ✅ Don't write the first word of a document —— Collaborate directly at the Demo / code level
  • ✅ When in doubt, ask GPT/Claude first —— Don't search first
  • ✅ When encountering an open-source project: look at the name → URL → CC install ➡️ explanation ➡️ testing
  • ✅ When seeing any long article/project link → throw it into a new CC window and ask questions(NotebookLM mindset)
  • ✅ Collaborate by writing code directly —— Commit directly within permission scope, let the code review Agent review
  • ✅ PUA Claude —— Iterate analysis → feedback → improvement repeatedly, the prompt eventually exceeds 6000 words

12. Engineering-level tips for Claude Code

  • ultrathink instruction:Require the model to provide reasoning logic when outputting, significantly improving output quality
  • CLAUDE.md:Long-term memory, no need to repeat foundational context when opening new windows
  • Build context:Refer to the Claude Code in Action course
  • Must-read materials:

13. Key reference link list

Wang Kai's own resources

Core external resources

14. Wang Kai says

"Once again I appeal: as long as you start using Claude Code, your token consumption will definitely increase exponentially, and your AI capabilities will definitely be at the forefront of the entire era."

"An agent is its working directory; the directory is the agent's body, CLAUDE.md is its persona,.mcp.json is its tools, and feedback.md is its growth record."

"Any scheme that tries to skip account warming and batch-send links will be instantly killed by anti-cheat systems; following the rules and accumulating trust is the only sustainable path for matrix operations."

References