Claude Code Workflows (Cross-Author Thematic Synthesis)

When to open this page: You want to decide how to use Claude Code, learn how people from different communities and identities use it, or choose between the two philosophies of "single-Agent engineering" and "multi-Agent matrix". This wiki is not a tutorial; it extracts the Wang Kai Claude Code Terminal Workflow Methodology (Full Version) / Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk) / newtype · AI Coding Practice and Tool Stack Evolution / dontbesilent AI Self-Media Course / 13-Lesson Systematic Methodology / Lenny's Newsletter / 2026 Trio Discussion on Organizational Operations in the AI Era (Anthropic + OpenAI) / Xiaohongshu / AI Tools + Skill Tutorial Collection / Wang Kai AI Coding Practice: From Cursor to Claude Code Terminal Workflow perspective fusion, consensus, and explicit disagreements across 7 author wikis on the topic of Claude Code.

One-sentence core: Claude Code is not an IDE / not Copilot / not a code-writing tool——it is an "Agent that lives in the terminal", and all perspective disagreements stem from two mental models: "treating it as an engineer colleague" vs "treating it as an Agent decomposition unit".

0. What is Claude Code? — 7 Perspectives

Each author defines Claude Code very differently; we list these clearly before discussing consensus:

Perspective	Source	Positioning	Core metaphor
Anthropic internal engineer perspective	Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk)(Cal Rueb)	Collaborative engineer colleague —— You make product decisions, it handles implementation details, you review and commit	"It lives in the terminal, because the terminal is the environment engineers are most familiar with"
Anthropic PM Head perspective	Lenny's Newsletter / 2026 Trio Discussion on Organizational Operations in the AI Era (Anthropic + OpenAI)(Cat Wu)	Multi-form product matrix —— The terminal version carries first releases of new features / the desktop version is for non-technical users / web+mobile lets you kick off tasks on the go / Cowork (Claude with hands) handles non-coding tasks (Slack 0 / writing decks)	"Claude Code spans forms; the smarter the model gets, the more crutches we can remove"
Independent developer matrix perspective	Wang Kai Claude Code Terminal Workflow Methodology (Full Version) / Wang Kai AI Coding Practice: From Cursor to Claude Code Terminal Workflow(Wang Kai)	A terminal window is an Agent —— Literally "open a new terminal and run claude", one window per function	"The directory is the Agent's body, CLAUDE.md is its persona,.mcp.json is its tools, feedback.md is its growth record"
AI tool stack evolution perspective	newtype · AI Coding Practice and Tool Stack Evolution / Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk)(Huang Yihe)	Main battleground for AI Coding after 2025-H2 —— "showed me the possibility of a relatively general-purpose Agent"	"Cursor cannot replace Claude Code's stack of Sub-agents / Skills / Slash commands"
Non-programmer zero-code perspective	dontbesilent AI Self-Media Course / 13-Lesson Systematic Methodology(Lao D)	True Agent (in contrast to "LLM question-and-answer") —— Most domestic translations of "intelligent agent" do not refer to real Agents; Claude Code is the real thing	"A continuous multi-step executor that can read/write files / run commands / call APIs"
Creative application + Skill-based perspective	Xiaohongshu / AI Tools + Skill Tutorial Collection(Xiaohongshu vibecoding)	Execution container for the Skill system —— IELTS 7.5 / Vibe Marketing batch videos / pixel sprite agent virtual office / Obsidian drawing	"The ability of Skills to call local services is something N8N does not have"
Liberal arts co-founder perspective	zara zhangzala / Vibe Coding Mental Model System(zara zhangzala)	Co-founder rather than employee —— Brainstorm together with Claude Code,bring problem not solution; the Tab Out entry design was proposed by Claude, not thought up by zara	"Claude Code understands technology better than I do; I give it the idea, it gives me the solution"

Key observation:Anthropic's own PM (Cat Wu) and engineer (Cal) describe Claude Code with a bias toward "single-Agent engineering";Independent developer (Wang Kai) pushes it to a "multi-Agent matrix";Non-programmers (Lao D / Xiaohongshu) treat it as an "executor for Skills / prompt engineering";zara treats it as a co-founder (creative partner). This is the root of all disagreements that follow.

1. Cross-author consensus (7 points)

1.1 Claude Code ≠ IDE ≠ Copilot

Tool	Mental model	Trust scale
Cursor / Copilot	"Completer" inside an IDE	You are driving
Devin / Manus	"Robot" that runs tasks automatically	It is driving
Claude Code	Agent in the terminal / engineer colleague / Skill execution container	Collaboration

Cal Rueb said: "It 'lives' in the terminal, because the terminal is the environment engineers are most familiar with; don't treat it like a GUI." Huang Yihe (newtype · AI Coding Practice and Tool Stack Evolution §6.1) switched his main battleground after 2025-07 with the same judgment: "Claude Code showed me the possibility of a relatively general-purpose Agent."

1.2 Agentic Search is stronger than RAG (counterintuitive but consensus across 4 authors)

Retrieval mode	Tool	Advantages	Disadvantages
Agentic Search	Claude Code	Dynamically decides which file to look at / which line to read, high precision	High token cost
Real-time search	Cline	Imitates humans, regex + high-speed file content	High token cost
Embedding + RAG	Cursor	High cost-effectiveness	"Function definition at line 200 / call at line 300 / dependency at line 400 → RAG can only read fragments, breaking the logic"
Low-latency RAG	WindSurf	Enterprise deployment	Slow initial indexing

Cal Rueb's explanation(Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk) §2.2): "RAG scans a book's table of contents into vectors, then uses similarity to find pages. But code is not a book; code is a graph——files import each other, functions call each other.An LLM holding grep / glob is like an engineer holding find / ag, and that is the correct way to search code."

Practical implications(consistent across authors):

Do not add RAG MCP to Claude Code——its built-in search capability is stronger than RAG
Large codebases (>100k lines) still work, because it only looks at a few relevant files each time
Wang Kai's extension:Build a self-hosted financial MCP to get accurate data(mandatory for investment scenarios), use Feishu CLI to pull all documents (replacing embed indexing)

1.3 CLAUDE.md is the core configuration (but there is disagreement on the number of layers, see §2.1)

All authors agree that CLAUDE.md is the core of Claude Code's behavior definition. Its content should include:

Project structure(directory tree, key modules)
Build/test commands(pnpm dev, pytest)
Code style + naming conventions
Architectural iron rules("API calls must go through src/api/")
Pitfalls to avoid("Do not touch legacy/auth.py")
MCP / Skill usage preferences

Cal's recommended anti-pattern: "You must X" style commands — use "We prefer X because Y" (instructions with reasons are followed better by LLMs).

Lao Huang (newtype · AI Coding Practice and Tool Stack Evolution §6.4) exclusive insight: "After finishing each issue,update CLAUDE.md once. If the issue is particularly large, when hitting milestones in the process, in addition to updating CLAUDE.md, I also manually execute /compact."——treat CLAUDE.md as working memory and maintain it continuously.

1.4 Consensus on the three tiers of Think Hard

Keyword	Thinking Token	When to use
`think`	~4k	Simple decisions (no need to use for fixing typos)
`think hard` / `think harder`	~10k	Debugging / refactoring
`ultrathink` / `Ultra think`	~32k(HIGHEST)	Architecture design / complex bugs

Lao Huang (newtype · AI Coding Practice and Tool Stack Evolution §6.2) practical principle: "When Claude Code gives a wrong answer,first add 'Ultra think' and retry; 90% of the time it's insufficient budget, not a bad model."

Cal Rueb's counterintuitive addendum: "95% of conversations don't need think; just let Claude Code do it. Only 5% of complex tasks need 'think hard' or 'ultrathink'.Don't ultrathink every time; it wastes tokens and slows down responses. If you find yourself repeatedly using ultrathink, that means your task granularity is too large; split the task into smaller ones, and using think hard for each small task is enough."

1.5 `/clear` and `/compact` are severely underrated context management tools

Cal Rueb's data: he himself averages 10–20 /clear commands per day. One /clear per small task.

/clear = clear the entire conversation context and start over → use when switching tasks / going off-topic / repeatedly retrying the same error
/compact = do not clear, but have Claude Code summarize the current context into a compressed abstract → use when approaching the token limit (default 200k, Sonnet 4.6 can reach 1M) / when a task is very long and needs a savepoint

Lao Huang's rhythm: Proactively /compact + update CLAUDE.md midway through large issues → preserve the train of thought while reducing consumption. Cal's warning: Do not start a new task immediately after /compact——first confirm that the compacted summary contains key information.

1.6 The Skill system is Claude Code's "second layer"

The 4 Major Stacks of Claude Code (Lao Huang newtype · AI Coding Practice and Tool Stack Evolution §6.1 Post 2148):

Stack	Function
Skills	Capability loading (encapsulating SOPs from your mind)
MCP	Tool calling (external server integration)
Slash commands	Workflow reuse (/research / /draft / custom commands)
Sub-agents	Multi-threaded division of labor

Cross-author consensus on Skill philosophy:

Wang Kai (Wang Kai's Complete Methodology for Claude Code Terminal Workflows §4):"Skill is the Agent's wiki" — internalizing work into reusable assets
Xiaohongshu (Xiaohongshu / AI Tools + Skill Tutorial Collection §3):"The advantage of Skill over N8N = it can directly call local stitching tools and local services"
Lao D (dontbesilent AI Self-Media Course / 13 Lessons of Systematic Methodology §L13):"Skill = a cross-platform reusable prompt engineering product" — Git version control + frontmatter trigger conditions
newtype(Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk) §1): Skill is a core component of Claude Code's collaboration model

See Claude Skill Engineering (Cross-Author Thematic Synthesis).

1.7 Multi-model orchestration / tool division of labor is the norm

Lao Huang's v0 + Claude + Cursor three-piece suite(newtype · AI Coding Practice and Tool Stack Evolution §2):

"Handing everything to one tool will most likely not yield good results. These AI tools are not yet strong enough for that,each has its own strengths. "

Tool	Role	Stage
v0.dev	UI / design drafts	Visual layer
Claude (Web/Desktop)	Framework / architecture	Abstraction layer
Cursor / Claude Code	Details / debug	Implementation layer

Wang Kai's multi-model orchestration(Wang Kai's Complete Methodology for Claude Code Terminal Workflows §3.3):

Claude Code (primary) + Codex (loaded into CC as CTO reviewer) + Gemini (large-context scanning) + Kimi (fallback)
Counterintuitive conclusion:A Claude subscription is the cheapest option(Kimi API: 1 task for 1.5h = 15 RMB / the $200 monthly fee amortizes well)

Lao D's domestic China adaptation(dontbesilent AI Self-Media Course / 13 Lessons of Systematic Methodology §L01 + §L08):

No VPN access in mainland China → DeepSeek + Monica wrapper + Openrouter three-piece suite
Access Claude Code via Openrouter Key (no paid membership required)

2. Explicit disagreements (4 items — not smoothed over)

2.1 Disagreement 1: Does CLAUDE.md have 3 layers or 4 layers?

Perspective	Number of layers	Path
Cal Rueb(Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk) §3.1)	3 layers	`~/.claude/CLAUDE.md`(global) + `<project>/CLAUDE.md`(team-shared, committed to git) + `<project>/CLAUDE.local.md`(personal, not committed to git)
Wang Kai (Wang Kai's Complete Methodology for Claude Code Terminal Workflows §5)	4 layers	Enterprise(`/Library/Application Support/ClaudeCode/CLAUDE.md`)+ User(`~/.claude/CLAUDE.md`)+ Project(`./CLAUDE.md`)+ Local(`./.claude/settings.local.json`) priority decreases in order

Why the difference: Cal speaks from the perspective of an Anthropic internal engineer (2 dimensions: personal + team are sufficient); Wang Kai speaks from the perspective of matrixed multi-identity operations (4 dimensions: enterprise-level / global / project-level / local — must account for "different credentials / different behaviors / resource binding").

Complementary conclusion:

Individuals/small teams → use Cal's 3-layer mental model
Multi-identity matrix / matrixed operations → use Wang Kai's 4 layers + 4-question decision framework (table below)

Wang Kai's 4 questions to decide which layer to place content 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)

2.2 Disagreement 2: Are Sub-Agents the primary tool or a supporting tool?

Perspective	Position	Data/reasoning
Wang Kai ("terminal window = Agent")	Primary reliance	"One window per function; the finer the split, tokens ×2, the better the quality". With 10+ parallel projects,he keeps 10+ Claude Code instances open long-term(physical multiple terminals)
Cal Rueb("Task tool sub-agent")	Use cautiously	"4 parallel Sub-Agents result in total time ~1.3x that of a single Agent(not 4x); startup + return overhead is very large". It is only cost-effective when there are 4 independent modules

Key insight: they are not talking about the same thing when they say "Sub-Agent":

Wang Kai's "Agent" = independent real terminal windows(physically isolated, viewable and intervenable)
Cal's "Sub-Agent" = child Claude Code instances launched by the Task tool(logically isolated, scheduled by the main Agent)

Complementarity:

Cross-task parallelism / multi-identity operations → use Wang Kai's multi-terminal window model
Intra-task parallelism (scanning different directories / refactoring multiple files) → use Cal's Task sub-agent

Cal's Sub-Agent boundaries(hold true at all times):

Do not let Sub-Agents perform commit / push
Do not let Sub-Agents start dev servers (orphaned processes)
The main Agent must review code written by Sub-Agents
Sub-Agents do not inherit the main Agent's CLAUDE.md — starting a new session re-reads it

2.3 Disagreement 3: Should Permissions be broad or strict?

Perspective	Position	Operation
Cal Rueb(Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk) §4.2)	Never use BypassPermissions in production codebases	Even when in a hurry, at minimum use AcceptEdits + Bash review; maintain a dangerous command list `rm -rf / git push --force / DROP TABLE / sudo`
Lao D (dontbesilent AI Self-Media Course / 13 Lessons of Systematic Methodology §L08)	Directly select "Don't ask again" (equivalent to BypassPermissions)	"Use Don't ask again mode to let Claude Code automatically approve permission requests" — teaching beginners to get started quickly
Wang Kai	AcceptEdits + Hooks interception (implicit)	Uses cc-status-board to monitor context; did not explicitly state whether to Bypass or not, but the "PUA Claude" workflow implies automatic permissions for long tasks

Real disagreement: Cal holds the strict Anthropic internal position for production codebases; Lao D holds the quick-onboarding position for non-programmer sandbox environments; Wang Kai holds the engineering position for matrixed operations.

Application judgment:

Production codebases / company projects → follow Cal's position, never Bypass
Local sandboxes / learning and exploration / non-technical tasks → Lao D's position is acceptable
Multi-Agent matrix → follow Wang Kai's position, use Hooks to block dangerous commands (detailed in §3 below)

2.4 Disagreement 4: Views on cost

Perspective	Position
Wang Kai(counterintuitive)	"A $200 Claude subscription is the cheapest" — Kimi API: 1 task for 1.5h = 15 RMB > amortized $200 monthly fee,for heavy usage, subscriptions from first-tier vendors like Claude / Codex are the cheapest
Lao D(domestic China adaptation)	"Use Openrouter without a membership, starting at ¥40" — for users in mainland China without VPN access, $5 can run many tasks
Lao Huang(tool stack observation)	"Running Cline with local large models is not recommended" — local models' coding capability cannot match Claude 3.5 Sonnet; even with Cursor/CC subscription costs, it is cheaper than hiring developers
Cal Rueb	Does not discuss cost (Anthropic internal)

Judgment criteria: usage volume (how many tokens are consumed per day?) + task granularity (heavy coding vs light self-media work?) + whether you are in mainland China without VPN access.

3. Configuration / Hooks / Permission Implementation Quick Reference (integrating Cal + Wang Kai)

3.1 Configuration path quick reference (tested by Wang Kai on 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 rules:

MCP write commands:claude mcp add <name> -s user|project|local <command> [args...] Do not hand-edit JSON
Skill priority: Enterprise > Personal > Project > Plugin — same-named upper layers override lower layers
CLAUDE.md is concatenated, not overwritten — all content written in the global CLAUDE.md appears in every agent,so it must be extremely minimal(Wang Kai's personal experience: writing "I work on SillyTavern Online" in global CLAUDE.md → polluted all projects)

3.2 Cal's recommended Hook usage

Hook event	Recommended usage
PreToolUse	Intercept dangerous commands (if bash contains `rm -rf` pop up a notification or block)
PostToolUse	Auto lint (run `prettier --write` or `pnpm lint`)
UserPromptSubmit	Inject context (append `git status`)
Stop	Auto commit + play a sound (`afplay /System/Library/Sounds/Glass.aiff`) — Cal: "so I can focus on other things and look back when I hear the sound"
Notification	Send macOS notification when waiting for user input

Cal's own hooks (simplified version):

{
  "hooks": {
    "PreToolUse": [{"matcher": "Bash", "hooks": [{"type": "command", "command": "~/.claude/hooks/block-dangerous.sh"}]}],
    "PostToolUse": [{"matcher": "Edit|Write", "hooks": [{"type": "command", "command": "~/.claude/hooks/auto-lint.sh"}]}],
    "Stop": [{"hooks": [{"type": "command", "command": "afplay /System/Library/Sounds/Glass.aiff"}]}]
  }
}

3.3 Wang Kai's anti-pattern quick reference (matrixed operations)

❌ Registering chrome MCP at the user level → cannot switch identities during matrixed operations
❌ Writing "I work on X" in global CLAUDE.md → pollutes all projects
❌ Same-named skills under multiple identity directories → get overwritten
❌ ~/.claude/mcp.json(This path does not exist, global MCP is in the ~/.claude.json field of mcpServers)

4. Multi-Agent Orchestration: Wang Kai's Exclusive "Engineered Division of Labor"

Wang Kai Wang Kai's Complete Methodology for Claude Code Terminal Workflows §3.2 presents an engineered division of labor that no other author has systematically covered:

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

3 major modes:

Claude Code produces product architecture → not based on the current state of the existing product; first clarify product goals, then have it produce the architecture + reverse-compare gaps against the current state
Load Codex into Claude Code for CTO review(codex-plugin-cc) → Codex helps build the full technical landscape → hand off to CC for resolution → Codex performs acceptance
Bug resolution closed loop: CC controls the browser for testing → resolves → controls the browser again itself to verify → loop

Wang Kai's token consumption evolution path:

Stage	Tools	Monthly spend	Key change
Cannot code	Claude + ChatGPT	< $100	Ordinary conversation
Using AI Coding	Cursor, Qoder, Trae, v0 all open simultaneously	~ $200	Token volume ×10
VS Code + Claude Code	Per-project	upgrade	Still constrained by project-based thinking
Terminal + Claude Code	Break free from project-based thinking	$125 → $250 → $200	Token volume ×2 again
Terminal multi-window + multi-model	CC, Codex, Gemini, Kimi each running	Same as above	Token volume ×3 again
Multi-browser concurrency + Skill + looped tasks	Chrome MCP ×N + Feishu CLI + 10min patrol	$200; API billing uses $50 in 3h	"Split work into dozens of Skills"

5. Long-task Scheduling: Wang Kai's Exclusive "10min Patrol"

Wang Kai's Complete Methodology for Claude Code Terminal Workflows §9.1 Counterintuitive judgment:

"I have more and more tasks that I let Claude Code execute nonstop…… but it always gets clever and stops on its own. At first I wanted to set up a detailed scheduled loop task via /loop to keep it running continuously. Then I had a change of thought——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 + if it's not running, make it continue.

(Other authors did not discuss this point separately; Cal implicitly advocates using a Stop hook to play a beep to draw the user back to look, rather than automatically resuming runs. This is a different choice for different task types.)

6. Cat Wu (Anthropic PM Head) Exclusive Perspective: Claude Code as a Product Matrix

Lenny's Newsletter / 2026 Three-Person Discussion on Organizational Operations in the AI Era (Anthropic + OpenAI) The first installment of the three-person discussion, this is a rare perspective in external materials on "how Anthropic's internal PMs view Claude Code":

6.1 Four Product Forms and Applicable Scenarios

Product	Applicable Scenarios	Key Features
Claude Code Terminal Edition	One-off coding tasks	New features land here first
Claude Code Desktop Edition	Frontend work + task overview + non-technical users	Real-time preview of web apps
Cloud Code Web / Mobile	"kicking off tasks while out" (starting tasks even when away)	Start tasks even when not at your computer
Cowork(Claude with Hands)	Non-code tasks(Slack 0 / Inbox 0 / writing docs / slide decks)	Connects to Google Calendar / Slack / Gmail / Drive, synthesizes information → 20-page deck

6.2 As Models Get Smarter, Products Do Subtraction

The early to-do list(task list) feature in Claude Code was a crutch "to help models do large refactors" — new models don't need it, so it was removed.

Why: as each generation of models gets smarter,products must proactively remove crutch features, otherwise the UI becomes increasingly bloated.

6.3 Anthropic PM Team Structure (30–40 people)

Team	Function
Research PM	Led by Diane, customer feedback + model releases
Cloud Developer Platform	API
Claude Code team	Core code product
Enterprise	Enterprise adoption
Growth	Overall product suite growth

Average ship cadence:Ships a feature every week, and in extreme cases, once per day.

6.5 zara zhangzala: Liberal Arts Background + Feishu CLI Hands-On Practice (Exclusive)

For details, see zara zhangzala / Skill & Independent Product Portfolio (Strategy Retrospective) + zara zhangzala / Vibe Coding Mental Model System. zara's perspective fills the gap in "real workflows of liberal arts graduates using Claude Code".

6.5.1 Feishu CLI + Claude Code Bus Pattern (zara's Real-World Practice)

zara uses Claude Code + Feishu CLI (self-built skill) to implement the "meeting todo → done" workflow:

飞书会议 → 妙记自动转写
            ↓
Claude Code skill 抓妙记 → 提取 todo
            ↓
分给不同 agent 执行(写文档 / 发消息 / 建议下一步)
            ↓
回写飞书文档(自动同步)

→ Cross-checked against Wang Kai's Claude Code Terminal Workflow Methodology (Full Version)'s "Feishu CLI collaboration bus"; zara's is a domestic China self-built skill in practice.

6.5.2 OpenAI Realtime + Feishu CLI Dictation-Style Document Writing (Exclusive)

A browser extension built by zara demonstrates the combination of the Real-Time model and Claude Code skills:

浏览器右下角小人(zara 插件)
   ↓ 实时语音(OpenAI Real-Time API)
说话:"我想写个文档介绍飞书 CLI 各种能力"
   ↓
Real-Time 模型同步整理成开场段
   ↓
追问:"展开群聊总结功能"
   ↓
继续填充段落
   ↓
追问:"用妙记 search 这场活动"
   ↓
飞书 CLI skill 调度 → 妙记 search → 内容回插
   ↓
完整飞书文档自动生成

Counter-consensus view: zara demonstrates that vibe coding is not just about writing code,writing documents / interacting with Miaoji / doing advanced summaries are all extensions of vibe coding. For details, see zara zhangzala / Vibe Coding Mental Model System §3 opinionated products.

6.5.3 Four Mental Principles for Chatting with Claude Code (Exclusive)

Principle	Differences from Cal / Wang Kai / Lao Huang
Use the best model (don't fear the cost)	Cal / Wang Kai talk about ROI / tokens; zara says cheaper models end up being more expensive(failures waste more tokens)
Treat it as a co-founder, not an employee	Wang Kai says "the directory is the Agent's body" (tool-oriented); zara says "it understands technology better than I do, I give it my ideas" (partner-oriented)
Try everything and push with high standards	Cal talks about "ultrathink"; zara says "AI won't get annoyed by you, push it all the way" (behavioral level)
Cut features before shipping	Exclusive finding — models are good at adding, not good at cutting. You must deliberately cut before shipping

6.5.4 Onboarding Path for Liberal Arts Users (Exclusive)

Stage	Action
0. Psychological ice-breaking	Directly vibe code a project, don't study CS, build first learn later (for details, see zara zhangzala / Vibe Coding Mental Model System §2)
1. First product	Website (faces.site) — entirely natural language / 0 code
2. Second product	Browser extension (Tab Out) — Chrome API involved
3. Third product	Claude Skill (Frontend Slides / follow-builders) — engineering encapsulation
4. End-to-end ship	Excalicord (8 systems connected) — Coding ≠ Shipping, 70/30 split

For details, see zara zhangzala / Skill & Independent Product Portfolio (Strategy Retrospective) §1-4.

6.5.5 Differentiation from the Other Six Schools

Dimension	Cal / Wang Kai / Lao Huang	xhs / Lao D / Cat Wu	zara
Mental model	Engineer colleague / Sub-agent / Tool	Skill executor	Co-founder / brainstorm partner
Focus	Implementation details / Permission / Hooks	Skill engineering / prompts	idea + ship, leave middle details to Claude
Liberal arts friendliness	Medium	Medium–High	Highest

7. How to Choose — Implementation Paths by Identity

Who you are	Primary method	Starting point
Anthropic-style engineer / Single-Agent engineering	Cal Rueb's perspective (Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk))	Configure hooks + CLAUDE.md + three permission tiers, use Sub-Agents cautiously via the Task tool
Independent developer / multiple products in parallel	Wang Kai's perspective (Wang Kai's Claude Code Terminal Workflow Methodology (Full Version))	Terminal window as Agent, Codex does CTO review, 10-minute patrols, Feishu CLI collaboration bus
Non-technical PM / content creator / based in China	Lao D's perspective (dontbesilent AI Self-Media Course / 13-Lesson Systematic Methodology §L08)	Openrouter Key + AI command line + Don't ask again, chat directly in Chinese
AI tool stack enthusiast / wanting to understand tool evolution	Lao Huang's perspective (newtype · AI Coding Practice and Tool Stack Evolution)	v0 + Claude + Cursor three-piece set → switch to Claude Code as the main battlefield + Sub-agents / Skills
Creative applications + Skill practice	Xiaohongshu vibecoding (Xiaohongshu / AI Tools + Skill Tutorial Collection)	Pixel character VS Code plugin visualization + Vibe Marketing Skill batch videos + IELTS Skill three-piece set
Product lead / wanting to understand PM work in the AI era	Cat Wu's perspective (Lenny's Newsletter / 2026 Three-Person Discussion on Organizational Operations in the AI Era (Anthropic + OpenAI))	Understand Cloud Code's multiple forms + Cowork boundaries + weekly ship cadence + removing crutches
Liberal arts user / afraid of code / wanting to build small products	zara's perspective(zara zhangzala / Vibe Coding Mental Model System)	Treat it as a co-founder, not an employee(see §6.5.4 for details) + bring problem not solution + Use the best model (cheaper ones end up more expensive) + cut features before shipping(models are not good at cutting) + build first learn later(for details, see zara zhangzala / Vibe Coding Mental Model System §2 §5)

Universal starting path(for any identity):

Cal §3 three-layer CLAUDE.md → clearly write out project structure / commands / style
Cal §6 add ultrathink keyword (retry first when it answers incorrectly)
Lao Huang "update CLAUDE.md after handling each issue" + node /compact
Cal §8 add PostToolUse auto lint + Stop beep
Wang Kai "the directory is the Agent's body" → one project per directory with one independent CLAUDE.md
When multi-task needs arise → choose between Wang Kai's "multiple terminal windows" / Cal's "Task sub-agent" (see §2.2)

8. Engineering-Level Tips / Must-Read Materials (Integrated)

Collected by Wang Kai

ultrathink instruction — requires the model to provide reasoning logic when outputting
CLAUDE.md — long-term memory, no need to repeat context when opening a new window
MinusX "What makes Claude Code so damn good"
Boris Cherny (CC founder) usage insights
Claude Code in Action course

Wang Kai's own tool stack

tokrepo.com — AI asset hosting / Skill / MCP / prompts
chrome-mcp-proxy — supports user data + session reuse + does not steal focus
cc-status-board — status bar, 3 major functions (model / context length / match installed MCP/Skills based on context)
weixin-claude-router — WeChat remote control of iTerm2 Claude Code

Lao D's domestic adaptation stack

rap.dev — AI command line + one-click template distribution (fully automated Xiaohongshu publishing)
Openrouter starting at ¥40 + no Claude Code subscription required

9. Reference List

Main original author wikis:

Wang Kai's Claude Code Terminal Workflow Methodology (Full Version) — independent developer matrix / terminal window as Agent / 4-layer configuration / Codex CTO review / 10-minute patrols
Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk) — Cal Rueb Anthropic internal talk / Agentic Search / 3 permission tiers / 6 Hook events / cautious use of Sub-Agents / Cal's personal configuration template
newtype · AI Coding Practice and Tool Stack Evolution — AI Coding tool stack evolution 2024-08→2026-05 / Cursor 4 retrieval modes comparison / Ultra think 32K / Plan Mode / Vibe Coding cleaning prophecy
dontbesilent AI Self-Media Course / 13-Lesson Systematic Methodology — domestic no-scientific-Internet-access Openrouter + AI command line / Claude Code is the real Agent / real-world cases
Lenny's Newsletter / 2026 Three-Person Discussion on Organizational Operations in the AI Era (Anthropic + OpenAI) — Cat Wu Anthropic PM Head perspective / Claude Code 4 forms / removing crutches / 30–40 person PM team
Xiaohongshu / AI Tools + Skill Tutorial Collection —— Pixel Character Agent / Vibe Marketing Skill Batch Videos (FFmpeg) / IELTS 7.5 Skill Three-Piece Set
Wang Kai AI Coding Hands-On: From Cursor to Claude Code Terminal Workflow —— AI Coding Evolution Path / Transforming Collaboration Processes / Disruptive Nature of Data Analysis

2026-05-19 Supplement: Anthropic Engineering Team First-Party Perspective

In 2026-05, we captured 25 articles from anthropic.com/engineering and introduced Anthropic’s own engineering judgments on Claude Code. This is the "original manufacturer perspective," cross-checked against the previous 7 author perspectives (Wang Kai hands-on / Cal lecture / Lao Huang / Lao D / Cat Wu interview / xhs / Wang Kai AI Coding) — where they align, where details are supplemented, and where official "hard rules" exist.

A. Sandboxing + Auto Mode Triangle (2025 Q4)

Claude Code defaults to "ask for every write operation." Pain point = approval fatigue(approval fatigue) → 93% of users click yes without even looking.

The 4 modes officially provided by Anthropic:

Mode	Security	Maintenance Cost	Autonomy	Applicable Scenarios
Manual approval (default)	Medium	Low	Low	Ordinary projects
`--dangerously-skip-permissions`	Extremely poor	Zero	High	Not officially recommended, temporary experiments
sandbox (old approach)	High	High (must be configured for every new capability)	Medium	Strict isolation + no network / host access required
Auto mode(New in 2025)	High	Low	High	Recommended for most daily use —— sandbox default + privilege escalation for sensitive operations

Anthropic actual measurement:sandboxing safely reduces permission prompts by 84%.

This means: it conflicts with Wang Kai’s hands-on habit of using "--dangerously-skip" for daily work.Anthropic official stance: auto mode is a safer equivalent solution, and the permission model should not be bypassed.

B. Two Postmortems Culture (2025-08 + 2026-04)

Anthropic publicly disclosed two large-scale Claude Code quality regressions in 2025–2026:

Time	Event	Truth
2025-08~09	A large number of users reported "Claude got dumber"	Three independent infrastructure bugs, not the model / not load
2026-04	Three change regressions in Claude Code (reasoning effort default changed to medium)	Wrong trade-off, reverted on 4/7

Official public commitment:"We never reduce model quality due to demand, time of day, or server load." Translation: Never reduce model quality due to load / time of day / server pressure.

Implications for the workflow:

If users feel quality has degraded → it may be a real bug, don’t put up with it, reports will be taken seriously + public postmortem
This is the engineering culture foundation that corroborates Wang Kai’s "Claude subscription is the cheapest contrarian take" / Lao Huang’s "Claude is the most stable"

See details in Anthropic Engineering Team First-Party Perspective: Agent Engineering / Skills / MCP / Claude Code (2025–2026 Complete Collection) §6.2.

C. Building Effective Agents 5 Patterns Are Canonical

Anthropic divides LLM applications into:Workflow(predefined paths) vs Agent(dynamic autonomy), and provides 5 building blocks:

1. Prompt chaining       ── 拆步骤,前→后串
2. Routing               ── 分类→不同 prompt/模型处理
3. Parallelization       ── 多 LLM call 同时跑,合并(sectioning / voting)
4. Orchestrator-workers  ── 编排器拆任务,动态生成子 task
5. Evaluator-optimizer   ── 一个 LLM 输出,一个 LLM 评,反馈循环

Core discipline:Successful implementations are simple composable patterns, not complex frameworks. This cross-checks consistently with Wang Kai’s "keep the tool stack as minimal as possible."

D. Using Claude Code to Modify Itself (Meta-Engineering)

From writing-tools-for-agents:Let Claude Code itself participate in modifying your tool descriptions.

Approach:

Show Claude Code "the tool + the transcript of the agent failing to use the tool"
Have it analyze "where in the tool description the description is unclear, causing the agent to use it incorrectly"
Iterate on the tool description until the agent’s accuracy meets the standard

Implications for the workflow:Claude Code is not just a code-writing tool, but also a tool for modifying prompts / modifying schemas / modifying its own onboarding. This fully aligns with Wang Kai / Cal Rueb’s "integrated tool stack."

E. Key Cross-Checks with the Previous 7 Author Wikis

Topic	Three Practical Camps	Anthropic Official	Status
`--dangerously-skip-permissions`	Wang Kai uses it daily	Not recommended, auto mode is available as a replacement	New divergence
CLAUDE.md layering	Wang Kai 4 layers / Cal 3 layers	Official documentation 3 layers	Cal aligns with official, Wang Kai has one extra layer
Sub-Agent	Lao Huang cautious / Wang Kai uses frequently	Listed as a standard capability in official best-practices	Aligned
Cost of connecting sandbox	Lao Huang does not use it	84% permission reduction + auto mode follows automatically	Anthropic pushes it more aggressively
Postmortem transparency	(No corresponding entry)	Public + detailed + remediation	New dimension of engineering culture

For more details see Anthropic Engineering Team First-Party Perspective: Agent Engineering / Skills / MCP / Claude Code (2025–2026 Complete Collection) §6 + AI Agent Evaluation Methodology (Anthropic Engineering Team demystifying-evals Deep Dive).

History

2026-05-19: Added §2026-05-19 section (Anthropic Engineering Team First-Party Perspective). Integrated 25 first-hand articles from anthropic.com/engineering, added 5 sub-sections (Sandboxing + Auto Mode / Two Postmortems / 5 Building Blocks Canonical / Using Claude Code to Modify Itself Meta / Cross-Checks with the Previous 7 Authors).
2026-05-18: Topic-dimension wiki first release. Synthesized and rewritten from 7 author wikis (Wang Kai CC overview / Cal CC lecture / Lao Huang AI Coding / Lao D self-media course / Lenny's Cat Wu interview / Xiaohongshu vibecoding tools / Wang Kai AI Coding). Explicitly marked 4 divergences: CLAUDE.md number of layers / Sub-Agent primary-secondary / Permission strictness / Cost perspective