My Vibe-Coding Practice Direction — The NL Glue Layer

The Problem: AI Writes Code Too Fast

Vibe-Coding is changing the rhythm of software development. AI coding tools have accelerated code generation by 10x or even 100x. But this speed brings an obvious problem: code out of control.

The Sweet Trap of Vibe-Coding

The Vibe-Coding experience goes like this: You tell the AI — “Help me build an order system that supports placing orders, payments, and refunds. Payments use webhook callbacks, refunds need an approval workflow, and all operations must have audit logs.”

Orchestrator Mode: How I Saved Myself from Half a Billion Daily Tokens

Five million tokens a day, I stroll with ease; fifty million tokens a day, I push with vigor; half a billion tokens a day, I’m drenched in sweat.

Background and Problem

The Context Cost of a Single Session

Every LLM call requires passing the full conversation history, meaning context grows linearly with the number of tool calls. After 50 tool calls in a complex task, the 51st call must carry the preceding 50 entries of history — not only is this costly, but the LLM also tends to lose focus and miss critical information in an overly long context.

Overview

Since the start of this year, I’ve been using AI intensively (Claude Code in particular) to help with many projects and write a lot of software. Over this time, my understanding of how a programmer should work with AI has gone through stage after stage. There’s a lot in my head, and I want to write it down to help myself reflect and synthesize.

Dimensions of the Journey

Several parallel threads have been evolving in my understanding of AI:

One-Line Summary

A complete requirement-driven development framework that turns the full software lifecycle — requirement analysis, technical design, coding, security audit, cleanup, compliance review, verification, and archival — into 8 executable skills for Claude Code, with checkpoint recovery and formal change management.

Why This Exists

The biggest problem with AI-generated code isn’t that it “can’t write code” — it’s that it writes without discipline:

• Starts coding before understanding requirements
• Makes a bunch of changes but can’t tell if they’re correct
• Interruptions mean starting over
• Security vulnerabilities go unchecked
• Requirements drift silently

This framework turns software development best practices into executable AI skills. Claude Code follows a structured process instead of winging it.

One-Line Summary

16 specialized AI agents analyze a stock from every angle — macro, sector, fundamentals, technicals, sentiment, news, filings — then two agents debate bullish vs bearish cases while a judge moderates. Final buy/hold/sell recommendation backed by recomputable numbers. Supports Chinese A-shares, Hong Kong, and US equities.

System Overview

Core Design: Why 16 Agents?

Ask a single LLM “should I buy this stock?” and you get:

• Hallucination: Made-up financial data
• Tunnel vision: Only technical analysis, or only fundamentals
• Unverifiable: No numeric evidence behind the claim

So we split into 16 specialized agents, each doing one thing, using real data instead of LLM-generated data:

Big Picture

The LLM is the brain, the Harness is the hands, the project code is what gets modified. The LLM never directly touches the filesystem — it only says “I want to do X”, and your code executes it.

The Essence: Three Sentences

1. The LLM is the engine: Feed project code to a language model, let it analyze, suggest improvements, and write code. At its core, it’s just a while loop asking the LLM “what else can be improved?”
2. Tools are the hands: The LLM can’t directly read or write files. It uses Anthropic’s tool_use protocol to tell your code “I want to read this file” / “I want to edit this line”, and your code executes it. You could run the whole thing with just a bash tool.
3. Process restart is the key: Python modules are loaded once at startup and stay frozen in memory. When the LLM modifies its own .py files, the running process still uses the old code. A process restart is the only way to apply improvements. That’s why we have the push → tag → CI deploy → restart loop.

From Simplest to Complete System: Each Layer Solves One Problem

Simplest Version (Conceptual)

1
2
3
4

while True:
    code = read_all_project_files()
    response = LLM("Here's the code, improve it:" + code)
    write_back(response)

This works. But it runs into problems. Each layer below solves the previous layer’s problem: