Cognitive Kernel

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

Cognitive Kernel

**Rules don't enforce themselves.**

One cognitive base, 30 lines of rules, injected into `CLAUDE.md` — it works. The agent's thinking changes. You can see it in the output.

Now install 19 bases. That's 720+ lines of rules, 100+ individual instructions, all competing for the agent's attention in the same context window. What happens? The same thing that happens when you tape 100 checklists to a wall — nobody reads any of them. The agent's attention gets diluted. Critical rules get skipped. The cognitive bases you installed are technically "active" but functionally invisible.

Why Text Rules Fail

Writing "you must do X" in a config file is the weakest possible intervention. It depends entirely on the agent noticing the rule, at the right moment, under cognitive load. That's like depending on a factory worker reading a safety poster while operating heavy machinery.

The 6-layer model replaces hope with structure. Instead of one mechanism (text in a file), each rule is enforced through the strongest viable channel:

Layer	Mechanism	Strength	How It Works
L1	Output templates	Strongest	Agent's response must include specific fields — structurally incomplete without them
L2	Hooks	Strong	Platform hooks auto-inject reminders before key actions — agent gets intercepted
L3	If-Then triggers	Medium	Pattern-matched rules fire automatically — no willpower needed
L4	External verification	Strong	A separate sub-agent reviews the work adversarially — self-checking doesn't work
L5	Core rules	Weak	A few critical rules stay in context — always visible, but still voluntary
L6	Skill reference	Weakest	Full framework loaded on demand — agent has to choose to read it

**The key insight:** L1 and L2 are structural — the agent cannot produce valid output without complying. L5 and L6 are vol