BYF

Twelve articles, from the Think-Act-Observe loop all the way down into the 500,000-line engineering abyss. Now, standing at the end of the series and looking back over the whole evolutionary path, a bigger question surfaces: What will AI coding tools ultimately evolve into? A better Copilot? A more powerful IDE plugin? Or something new that we haven’t named yet? Three Evolutionary Paths # Looking back at the development of AI coding tools, you can see three clear stages of evolution:

The claude-code-from-scratch project implements a “mini Claude Code” in roughly 3,000 lines of TypeScript. It has the Think-Act-Observe loop, a tool system, a system prompt, permission control, memory, Skills, multi-agent, MCP — every core concept we tore down over the previous ten posts. And the official Claude Code source? Roughly 500K lines of TypeScript/TSX. That’s a 100x+ gap between 3,000 lines and 500K. What’s in it? The gap isn’t “number of features”. from-scratch already covers every core concept. The gap is the road each concept travels from “it runs” to “it’s usable” — that road is called productionization.

Autoregressive large language models operating in unconstrained environments suffer from an inherent tendency toward myopic mutation and runaway divergence. Claude Code’s Plan Mode attaches an external discrete Finite State Machine (FSM) that projects the action space onto a strictly side-effect-free subset, using an interactive human barrier to prune catastrophic physical backtracking branches from the decision tree. In disciplined software engineering, experienced developers instinctively value structural planning: mapping dependency graphs, deriving type signatures, and scoping blast radiuses before modifying a single line of code. Large language models possess no such innate restraint. Shaped by reinforcement learning to generate immediate answers, an agent granted unrestrained write permissions will instinctively mutate files without global context, precipitating an architectural cascade across complex repositories.

Before MCP, every AI tool integration was a custom build. Want ChatGPT to reach a database? Write a plugin. Want Claude to call the GitHub API? Write a Function Calling definition. Want Copilot to access your internal systems? Write a Custom Tool. Every time, the same thing: define an interface, write adapter code, handle auth, manage errors. The wheel, reinvented countless times. The problem MCP (Model Context Protocol) solves is simple: give AI agents a standard protocol for connecting tools. Just as USB-C gave electronic devices a universal connector, MCP gives AI agents a universal tool interface.

There’s a counterintuitive phenomenon: take the same model, split it into multiple independent instances, and have them collaborate — the result is often better than a single instance. Not more compute (three instances burn more total tokens than one), not more capability (every instance is the same model). The gain comes from role separation — when each agent attends to only one facet of the problem, it works with more focus and more depth than a “do-everything agent”.

If you’ve used Claude Code, you may have noticed the /skills command listing installed “skills”. Writing standards, architecture design methods, TDD workflows, code review strategies — it sounds like a plugin marketplace. Skills are not plugins. That’s the first and most important thing to understand about the Skill system. A plugin is code — it extends what the program can do. A Skill is a prompt — it extends what the model knows.

Last time we covered context compaction — solving “how much the current session can remember”. But a coding assistant needs another kind of memory: long-term memory across sessions. You close Claude Code and reopen it tomorrow. What should it know? You prefer TypeScript over JavaScript You said last time “don’t use the underscore library” This project uses Vitest, not Jest You once corrected one of its misunderstandings Context compaction can’t handle these. The session ends, the history clears, tomorrow is a fresh window. If the AI needs you to re-explain everything each time, it isn’t an assistant — it’s a burden.

Large language models possess no biological memory at the hardware tier; they operate entirely on Key-Value Caches that scale linearly in GPU High-Bandwidth Memory (HBM) alongside attention weight distributions that mathematically dilute as sequence length grows. The notion of long-horizon conversations presented by Claude Code is, at its engineering core, an eviction and compaction pipeline operating within the tight margins of memory capacity and numerical precision, trading away information entropy and prefix caching efficiency.

Any AI agent permission system running in userspace that relies on model self-discipline and application-level regex filtering does not constitute an enforceable security boundary from the perspective of the operating system kernel. Claude Code’s 5-level permission model and interactive approval workflow operate as flow-control protocols to mitigate developer mistakes, lacking sandbox containment against adversarial attacks. Granting an AI coding assistant filesystem write access and terminal execution rights under the belief that confirmation prompts preserve a safety perimeter is a fundamental category error. Confusing interactive human approval with privileged access control collapses into a cosmetic barrier when confronted with supply-chain poisonings, indirect prompt injections, and shell parser differentials.

If you think a system prompt is just “giving the AI a persona”, Claude Code’s system prompt will upend that notion. It isn’t a few sentences. It’s a thousand-line template engine made of multiple layers, each responsible for something different: identity definition, behavioral constraints, tool documentation, project context, memory injection, skill loading. Some layers are static (identical every time); others are dynamic (injected in real time based on the project, the session, user preferences).