The Shelf

Everything here is yours to steal. Templates, one-liners, and plain-English definitions. No theory, just the tools.

The starter constitution

Five rules, each one earned in a chapter. Copy it, rewrite every rule in your own words (Chapter 7 explains why), and put it wherever your AI tool keeps standing instructions.

RULES FOR WORKING WITH ME

1. Never tell me something is done unless you can show me
   evidence: a record, an ID, a change I can verify.

2. If you are not sure, say you are not sure. A wrong answer
   costs me more than a slow one.

3. Before we build around any constraint, check it. Tell me
   what you verified and how.

4. Every fact lives in exactly one place. Tell me when you
   see the same fact in two places.

5. When something breaks, give me two or three possible
   causes before any fix.

Rules of thumb

Evidence over reasoning.
If you can't verify it, it didn't happen.
Confidence is a writing style. Evidence is a thing in the world.
The same fact in two places is a bug.
Chat is a dead end you can live in for years.
Before you build around a wall, touch it.
Your first guess is a prediction machine too, and it is overconfident too.
The system is working when it needs you less.
The mess is memory. Redirect it, don't fight it.

Plain-English glossary

Model: The machine itself: a text predictor trained on a huge amount of writing.
Agent: An AI with hands: connected to real systems, acting on your behalf, with consequences.
Context window: The machine's working memory. What slides out of it is gone, as if never said.
Hallucination: When plausible text describes a world that doesn't exist. Fabrication as the failure mode of plausibility.
Prompt: What you give the machine to work with. Includes your question, your documents, and your standing rules.
Source of truth: The one home a fact lives in. Everything else is a view, not an original.
Verification: Checking reality instead of accepting a claim. The entire basis of trust with an agent.
Constitution: Your written, durable rules that the AI reads every time. Your judgment, given a body.

What mature doctrine looks like

A glimpse of the destination. After a couple of years, a working constitution grows from five rules into something like this (excerpts from real ones, lightly edited):

Define what "done" means before starting, not after.

Break big work into pieces that can each be checked
on their own. One piece, one risk, one clear finish line.

When the machine fails twice the same way, stop prompting
harder. Change what it sees: rules, examples, data.

Review the work like an editor, not a fan. The machine
does most of the doing; you own the judgment.

Track what failed and why. The log of failures is the
curriculum for the next month's rules.

Where to go next

There is no chapter 9. The next step is not more reading.

Do the practices for real. Especially the Walls tally (Chapter 4) and the constitution week (Chapter 7). Reading them counted for nothing, doing them counts for everything.
Answer the return questions when they find you. They are how this library makes things stick.
Then text me. Tell me which chapter's question wouldn't leave you alone. That conversation is the graduation.