upgrade instruction to subagent or skill

[botingw]

context: current instruction to LLM is a single custom rule file.
subagent and Claude Skills pros:

The Core Principle

Attention Conservation: LLMs get stupider as context length grows. Modularizing rules stops the AI from "forgetting" instructions by never loading irrelevant ones in the first place.

---

Part 1: Claude Code Skills (The "Tools")

Focus: Execution & Specificity

• Just-in-Time Context Loading
  • The Advantage: Instead of force-feeding the AI 5,000 tokens of rules for every language you use, it loads specific instructions only when triggered (e.g., only load "AWS Best Practices" when you type deploy).
  • Result: Saves token costs and keeps the AI’s "IQ" higher by reducing distraction.
• Deterministic Reality (Fact vs. Hallucination)
  • The Advantage: A text rule asks the AI to "guess" the state of your code. A Skill (via MCP) runs a command (like grep or linter) to verify the state.
  • Result: Moves the agent from "I think this is broken" to "I have executed code and confirmed this is broken."
• Encapsulated Logic
  • The Advantage: You can treat prompts like software functions. You can fix the logic for a "Database Migration" skill without accidentally breaking the logic for a "Frontend Component" skill.

Part 2: Subagents (The "Specialists")

Focus: Reasoning & Context Hygiene

• Context Hygiene (The "Fresh Start")
  • The Advantage: Subagents spin up a brand new context window. They do not see your previous 40 messages of debugging errors.
  • Result: They solve the immediate problem with zero bias or "confusion hangover" from previous failed attempts.
• Role Specialization
  • The Advantage: A single .cursorrules file tries to make one model a Jack-of-all-trades. Subagents allow you to configure one model purely as a "Hostile Security Reviewer" and another as a "Helpful UI Designer."
  • Result: You get depth of expertise rather than shallow breadth.
• Parallelization
  • The Advantage: You can have a "Doc Writer" agent and a "Unit Test" agent working simultaneously on the same code block.
  • Result: Faster wall-clock time for complex tasks, though this is harder to orchestrate.

The Honest Trade-off

• Single File: Fast, stupid, easy to manage. Best for quick scripts.
• Skills/Subagents: Slower, smarter, harder to maintain. Required for enterprise-grade engineering.