How a Google Engineer Automated 80% of His Work Using Claude Code

Karpathy‑Based CLAUDE.md

Andrej Karpathy identified four recurring coding mistakes made by LLMs: over‑design, ignoring existing patterns, adding unnecessary dependencies, and making unrelated changes. A community‑maintained CLAUDE.md file encodes these observations as four actionable principles:

Think Before Coding      → Prevent false assumptions and missed trade‑offs
Simplicity First         → Avoid over‑design and bloated abstractions
Surgical Changes        → Modify only code that is explicitly required
Goal‑Driven Execution   → Test first, verify success criteria

Applying the file to a project reduces Claude’s rule‑violation rate from ~40 % to ~3 % after a five‑minute setup.

Generate a project‑specific CLAUDE.md with:

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

Everything‑Claude‑Code Repository

The repository https://github.com/affaan-m/everything-claude-code provides a full AI‑operating‑system for product development. It contains:

30+ specialized agents (e.g., planner.md, architect.md, tdd‑guide.md, code‑reviewer.md, security‑reviewer.md, loop‑operator.md)

180+ ready‑to‑use skills covering test‑driven development, security analysis, research, content generation, etc.

Built‑in AgentShield with 1,282 security tests that balance productivity and safety.

Installation options:

/plugin marketplace add affaan-m/everything-claude-code

or manually copy only the required agents into the project’s .claude/ directory to keep token consumption low.

Token Inflation in Claude Code v2.1.100

Network analysis shows that version 2.1.100 sends slightly fewer request bytes (≈168 KB) but consumes about 20 k more tokens (≈70 k tokens) than v2.1.98, inflating billing without adding functionality.

Impact:

Hidden token usage dilutes the effect of CLAUDE.md directives.

Faster quota exhaustion and degraded response quality in long sessions.

Temporary fix (≈30 seconds) until an official Anthropic patch:

npx [email protected]

Case Study: End‑to‑End Automation Workflow

A senior engineer built a three‑stage system using a lightweight .NET application that polls a GitLab issue every 15 minutes.

Step 1 – Classification:
  • .NET app calls GitLab API
  • Claude reads the issue and decides if development is ready
  • If not ready, Claude drafts a reply for review

Step 2 – Execution (when ready):
  • Sub‑agent creates a new branch
  • Opens a pull request for review

Step 3 – PR Workflow:
  • Checks for matching PRs
  • Monitors new comments
  • Applies modifications based on feedback

Results after one week:

Before: 8 hours of coding per day
After:  2–3 hours of review and testing per day
Code quality: consistent, fully reviewed by Claude
Developer status: online, mouse moves automatically each minute
Remaining time: fully discretionary

Setup Checklist

Generate Karpathy CLAUDE.md (≈5 min)

claude -p "Create a CLAUDE.md based on Karpathy's principles for this project" --allowedTools Bash,Write,Read

Install selected agents from Everything‑Claude‑Code (≈10 min)

/plugin marketplace add affaan-m/everything-claude-code

Install only the agents you need; avoid loading the entire repository.

Apply token fix if using v2.1.100 (≈30 s)

npx [email protected]

Benefits Overview

Rule‑violation rate drops from 40 % to 3 % with the Karpathy CLAUDE.md.

Ready‑made agents eliminate weeks of custom AI‑system development and avoid $200–$500 monthly AI‑service fees.

Downgrading to the fixed version recovers roughly 40 % of token quota.

Typical time savings: 5–6 hours per day (≈100–120 hours per month), translating to $3 000–$12 000 in avoided labor costs (based on $30–$100 hourly rates).

References

https://github.com/affaan-m/everything-claude-code

https://x.com/noisyb0y1/status/2043609541477044439