Meta Cogbase

[English](README.md) | [中文](README.zh.md)

meta-cogbase

**Every AI agent is missing a layer.**

You've installed 100 skills on your agent. It can write code, analyze data, generate reports. Then you ask it a real question — "Should I take this job offer?" — and it gives you a pros-and-cons list that a college freshman could have written. You ask it to build a complete project plan, and it delivers 80% then says "Done!" You ask it to investigate a bug, and it glances at the logs and says "Probably a config issue" — without actually checking.

The skills aren't the problem. You could install 200 more and nothing changes. Because skills change **what** an agent can do. Nothing changes **how** it thinks.

What Is a Cognitive Base?

A cognitive base is a set of meta-cognitive instructions that rewire an agent's default thinking patterns. Not a tool that activates in specific scenarios — a permanent change to how the agent approaches *every* task.

Think of it this way: **skills are recipes, cognitive bases are knife skills.** A hundred cookbooks let a chef make more dishes. But if the knife work is sloppy, everything is just a little off.

Each cognitive base is ~30 lines of core rules injected into the agent's constitutional config (`CLAUDE.md`, `AGENTS.md`, `system instruction`) — loaded on every conversation, every task, no trigger needed. Plus a full reference framework (~120 lines) for deep situations.

One Base, Three Completely Different Scenarios

Take [First Principles](https://github.com/d-wwei/first-principles) — 30 lines of rules. Same base, applied to:

• Life decision ("Should I get an MBA?") — Instead of listing pros and cons, the agent decomposes MBA into four bundled products (credential signal, network, knowledge, career pause) and checks each independently. The real question becomes: is the credential signal hard currency in your target industry?
• Business judgment ("Should we raise funding now?") — Instead of analyzing "market timin