Keep teaching the agent
You do not write one perfect setup once. You keep updating the rules as you learn what helps and what annoys you.
Everyone uses AI. The field is full of people sharing tips, techniques, and tutorials. Karpathy explains how the models work. Dozens of guides teach you how to write better prompts. You can find thousands of CLAUDE.md templates on GitHub.
So why does this site exist?
Four things that aren't out there. Not in the prompt engineering guides. Not in the LLM explainers. Not derivable from first principles without having lived through building an AI operating methodology from scratch.
The Four Things
1. Your constitution is a live document, not a setup step
Everyone is converging on "write a CLAUDE.md" or "build a system prompt." That's the technique. The insight most people miss: the value is in the updating, not the artifact.
A system prompt you write once and leave alone starts decaying immediately. Your context changes. Your vocabulary evolves. You learn what the AI gets wrong about how you work. The frozen version drifts from reality.
The constitution is only valuable as a live document with an active update history. The update history IS the product. What has changed, when, and why — that's what makes it an accurate model of how you work, rather than a snapshot of how you thought you worked when you wrote it.
What this changes: You don't "set up your constitution." You maintain it. The habit is updating, not writing.
2. "Specific enough to be wrong" is the quality test
Most people write constitutions that don't actually constrain anything. "Be concise and helpful." "Use a professional tone." "Think step by step."
These are aspirational. They can't be violated in any way you'd notice. The AI could produce exactly what you hate most and technically comply with "be concise and helpful."
The test: is your instruction specific enough that the AI could clearly violate it?
- "Be helpful" — not specific enough to be wrong. Fails the test.
- "When I ask for options, give me exactly 3, in bullet points, and tell me which one you'd choose." — specific enough to be wrong. Passes.
A constitution full of instructions that pass this test is a constitution that's actually doing something.
What this changes: You evaluate your constitution not by whether it sounds good, but by whether each line could be violated in a way you'd immediately notice.
3. You cannot audit your own constitution from within it
After you've built a working constitution and are running sessions with it, you will have a blind spot: the same session running on your constitution can't reliably find where the constitution is wrong.
The AI in a normal session is optimizing to work within your stated principles. It has been told what you value. It gives you outputs consistent with that framing. If your constitution says "I value rigor," the session will confirm rigor — not question whether rigor is actually what you need right now.
This is a structural property of any system that runs on its own rules. You cannot audit yourself from inside your own methodology.
The solution is architectural: a separate session type with an explicit mandate to find where the primary system is wrong. Different context loaded. The instruction isn't "help me work" — it's "tell me where my working methodology is failing."
Most people never run this session. Their constitutions drift further from useful over time while they feel like the system is working fine.
What this changes: A quarterly session with one job — find what's wrong. This is not the same as your normal sessions. It's architecturally separate.
4. Trust resurfacing over cataloguing
The natural instinct when you see a problem with your AI workflow: add it to a list. "Things to fix eventually." The list grows. Most items sit there for months, get reviewed occasionally with good intentions, and mostly never get addressed.
The alternative: no list. If a problem matters, it will resurface. When it resurfaces, that's your signal to run a loop and fix it. If it doesn't resurface, it wasn't worth fixing.
This sounds risky. What if you miss something important? Here's what's actually happening: the things that stay on your list without resurfacing weren't blocking you. They were noise that felt signal-adjacent once. The signal is in the resurfacing — that's what tells you a problem is load-bearing.
A backlog of AI workflow improvements is a backlog of things you'll feel slightly guilty about and never actually address. Trust the system.
What this changes: When you notice a problem with your methodology, you don't add it to a list. You either fix it now (if it's surfaced) or let it go. The discipline is deciding: now or never.
What Comes Next
These four insights are the reason this site exists. The techniques in the curriculum implement them:
- The constitution lesson teaches how to build something that passes the specificity test and that you'll actually maintain
- Session close makes the constitution durable across sessions
- The synthesis engine is the technique that makes your constitution most useful
- Quality loops are the mechanism behind the resurfacing principle
- The meta-layer audit is the architectural solution to insight 3
Start with lesson 1.