Ungoverned by Default
Your AI is making you average.
The worksheet from the session, and the homework that goes with it. One row per rule. You fill in the rule and the level it is on today; the rest is what you do back at your desk, and it is the part that does the work.
The rule
Five rules your organization follows to get work out the door, then three things a new hire learns in their first month that nobody ever wrote down.
Stuck? How you size a piece of work. How you decide something is finished. How you tell your leadership when it will land.
Level now
Which of the five levels each rule sits on today. Most written rules land on 1, 2 or 3. Everything nobody wrote down is a 0.
This is the one that surprises people. You write it yourself and the answer is right there in your own handwriting.
What your AI actually said
Pick three of your own rules and ask your agent how it does each one. Write down what it really answered, not what you were hoping it would.
Ask about what you wrote, not about an example. The comparison only means something if both sides are about the same rule.
Level needed
Where that rule actually needs to live. The gap between this number and the one you wrote in the room is your list.
Most rules are fine where they are. Only a few earn a wall.
The five levels
Where a rule lives decides whether it holds. This is the scale you are writing numbers against.
Everybody just knows it. The agent never had it, and never will.
A prompt, a chat turn. Dies when the conversation does.
A markdown file somebody has to open. Read once, then outranked by everything else it ever read.
CLAUDE.md, .cursorrules, a system prompt. Survives the session, bends under load, and the next edit can contradict it.
A hook, a linter, a test, a CI gate. Holds. The work does not go forward until the rule is satisfied.
Then compare
Everywhere what your AI said disagrees with what you wrote is a standard you are handing over, and you never decided to. Everything you put at level 0, it never had in the first place. A new hire absorbs your standards by sitting next to people who hold them. An agent has nobody to sit next to.
Those gaps are candidates, not answers. A gap earns a wall if the common way is actively wrong for you rather than just different, if the drift stays invisible until it ships, and if it reaches somebody outside your team. Two of those three is the bar. Everything else stays advice, and advice is fine for the things that do not matter very much.
Take the paper version
One page, landscape, one row per rule. Prints on a single sheet, and it is far easier to fill in by hand than anything on a phone.
Open the printable worksheetWant to see what it did?
This is the argument. The evidence is the other session: six months of data out of one organization, what the whole thing cost while we were figuring it out, and what it did to delivery once it held.
Talk to Liz about your own walls