Original Title: Google engineer automated 80% of his work with Claude Code. here's the exact system he built.
Original Author: @noisyb0y1
Compiler: Peggy, BlockBeats
Editor's Note: As "AI writing code" gradually becomes the industry consensus, what truly changes productivity is not the model itself, but how you set rules for the model, organize processes, and embed it into a sustainable operating system.
Starting from a simple CLAUDE.md file, to multi-agent collaboration, and then to an automated development loop, this method transforms the development process from "dialogue between man and AI" to "management of an AI engineering team." In this process, errors are constrained in advance, processes are structured, and code generation, testing, and review are gradually detached from manual execution and taken over by the system.
Moreover, it is worth noting that the article also reveals an overlooked detail: in long contexts and complex systems, model behavior is not entirely controllable. Whether it's hidden token consumption or diluted instructions, both can invisibly affect output quality. This makes "how to manage AI," rather than just "how to use AI," a new core competency.
At this point, developers are no longer centered around coding but work around rule design, process scheduling, and result verification. Those who completed this step first have begun to shift from "doing things themselves" to "letting the system do things for them."
The following is the original text:
A Google engineer with 11 years of experience has automated 80% of his work using Claude Code and a simple .NET application.
Now, he only needs to work 2–3 hours a day instead of the previous 8 hours, and spends most of his remaining time "relaxing," while the system operates on its own, bringing him a monthly passive income of $28,000.
What he has mastered is a set of methods you have yet to understand.
Part 1—Writing CLAUDE.md According to Karpathy's Principles
Andrej Karpathy—one of the most influential AI researchers globally—has systematically summarized the most common mistakes made by large language models when writing code: overdesign, neglecting existing patterns, and introducing unnecessary dependencies.
Andrej Karpathy used Claude Code and a simple .NET tool to create an automated pipeline that strings together processes such as task acquisition, assessing feasibility, auto-writing code, submitting PRs, and modifying based on feedback, thereby handing off about 80% of the development work to the system, leaving him responsible only for the final review and testing, reducing his work hours from 8 to 2–3, essentially achieving a transformation from "coder" to "manager of an automated development system."
Someone summarized these observations into a unified CLAUDE.md file.
As a result, this project gained 15,000 stars on GitHub within a week, meaning that, in some sense, 15,000 people have changed their work methods because of it.
The core idea is quite simple: if errors are predictable, they can be preemptively avoided with clear instructions. By simply placing a markdown file in the code repository, you can provide a complete set of structured behavioral rules for Claude Code, thus unifying decision-making and execution throughout the project.
This file mainly contains four core principles:
· Think first, then code → Avoid erroneous assumptions and overlooked trade-offs
· Simplicity first → Avoid overdesign and bloated abstraction
· Surgical changes → Avoid altering code that no one requested to change
· Goal-driven execution → Test first, then verify according to clear success criteria
It doesn’t rely on any framework nor requires complex tools—just one file can change the behavior of Claude at the project level.
The real difference lies in:
· Without using CLAUDE.md: Claude violates the norms in about 40% of cases
· Using Karpathy's CLAUDE.md: Violation rate drops to about 3%
· Setup time: Only 5 minutes
Command for automatically generating your own CLAUDE.md file:
claude -p "Read the entire project and create a CLAUDE.md based on:
Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution.
Adapt to the real architecture you see." --allowedTools Bash,Write,Read
What it replaces is a Claude that overdesigns and introduces unnecessary dependencies when facing simple tasks, and modifies files that shouldn't be altered at will.
Part 2 Everything Claude Code: A Complete Engineering Team in One Repository
Everything Claude Code (over 153,000 stars on GitHub)
This is not just a collection of prompts but resembles a complete AI operating system for building products.
30+ specialized agents: planner.md → functional planning architect.md → system design decisions tdd-guide.md → test-driven development code-reviewer.md → code quality and security review security-reviewer.md → vulnerability analysis loop-operator.md → automated loop execution
180+ skills: TDD, security, research, content generation—all pre-configured
Built-in AgentShield: 1282 security tests directly integrated into configurations
Supports running on multiple tools such as Claude, Codex, Cursor, OpenCode, Gemini, etc.—a single system, available everywhere.
Installation method:
/plugin marketplace add affaan-m/everything-claude-code
Or install manually—just copy the components you need into the project's .claude/ directory. Do not load everything at once—loading 27 agents and 64 skills simultaneously may exhaust the context quota before you input the first prompt. Only keep the necessary parts.
The real difference lies in:
· Before: You were conversing with AI
· After: You are managing an automatically running AI engineering team
What it replaces is the need for you to spend weeks building your own agent system, configuring different tools for planning/review/security, and paying $200–500 per month for various AI services.
Part 3 A Hidden "Scandal": Claude Code v2.1.100 is Quietly Consuming Your Tokens
Someone built an HTTP proxy to intercept and analyze complete API requests from four different versions of Claude Code.
They discovered:
v2.1.98: 169,514 bytes request → 49,726 tokens charged
v2.1.100: 168,536 bytes request → 69,922 tokens charged
difference: -978 bytes but +20,196 tokens
v2.1.100 sends fewer data bytes but charges an additional 20,000 tokens. This "inflation" occurs entirely on the server end—you cannot see it or verify it through the /context interface.
This image primarily reveals: Certain versions of Claude Code (especially v2.1.100) exhibit "hidden token consumption," leading to an abnormal acceleration of usage limits. Specifically, versions like Claude Code v2.1.100 introduce about 20,000 "invisible tokens" quietly on the server side, resulting in higher billing even with less request data, causing users' context to be invisibly occupied, limits to be exhausted more quickly, and even weakening model output quality; this issue was validated by comparing API requests from different versions through the HTTP proxy, essentially being a form of "hidden consumption" that users find difficult to detect and cannot verify themselves.
The importance of this issue goes beyond billing; this additional 20,000 tokens will be stuffed into Claude's actual context window.
This means:
→ Your CLAUDE.md instructions will be diluted by these 20,000 "hidden contents"
→ In long conversations, output quality will decline faster
→ When Claude ignores your rules, it will be hard to find out why
→ The usage limit for Claude Max will be consumed about 40% faster than normal
Fixing it takes only 30 seconds: npx claude-code@2.1.98
This is a temporary solution before an official fix from Anthropic, but in actual use, you can almost immediately feel a change in conversation effectiveness.
What it replaces is the need for you to guess why Claude suddenly stops following your instructions.
Case Study: What a Complete Automated System Looks Like
An engineer with 11 years of experience built a system consisting of three parts:
Results After One Week:
· Before: Coding for 8 hours a day
· After: Only 2–3 hours a day for code review and testing
· Code quality: Essentially unchanged—because he reviews each one
· Teams status: Always online—mouse moves automatically every minute
· Remaining time: Freely allocated throughout the day
This is not "magic," but rather the result of CLAUDE.md + suitable agents + a loop mechanism every 15 minutes working together.
Complete Checklist:
What You Gain After Reading:
· Before: Claude violates existing norms in 40% of cases
· After: Using Karpathy's CLAUDE.md, the violation rate drops to 3%
· Before: You need several weeks to set up agents
· After: 27 agents are ready to use out of the box
· Before: Claude Max runs out of limits within 2–3 hours
· After: Downgrading to v2.1.98 can restore about 40% of the usage limit
· Before: Writing code requires 8 hours a day
· After: Only 2–3 hours for review, with the rest handled by the system
· Setup time: 15–20 minutes
· Daily savings: 5–6 hours
· Monthly savings: 100–120 hours
If your time is valued at $30 per hour—then you are "invisibly losing" $3,000–3,600 per month.
If it’s $100 per hour—that is $10,000–12,000 being wasted every month, just because you are still manually writing code that Claude could handle on his own.
Most developers will never reach this level—not because they can't, but because they think it's complicated. In fact, you're only three commands and one file away from "full automation."
The engineer I mentioned at the beginning is not a genius or a senior engineer from Google. He just spent one night setting up the system—since then, his work has been completed by the system, and he is only responsible for living.
You can do the same thing tonight. While others are still debating whether AI will replace developers, those who have already set up systems are simply relaxing while earning money.
The choice is actually clear. You are building your own life—so choose the right path.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
