Your AI ignored a rule this week.
You saw 0.
correct
obey once
forget
correct again
The loop never closes.
4.5%
→
8.3%
8 of every 100 messages were corrections, up from 4 or 5. June to August 2026. The rules file grew the whole time.
A sentence cannot fail. A check can.
CLAUDE.md, your rules file
- answer first, details second
11 rules, 11 lights. A light is a check: a small yes-or-no test the computer runs by itself. Green = yes, the rule held.
$ claude
metacognition RED: 1 of 11 rules broken: answer-first
the verdict: one line, GREEN or RED, and which rule. It is the first line of every new session.
Every 30 minutes, while you sleep.
48
checks a day
1,190
runs since 11 August
No window, no sound, no internet. It prints one line and keeps a record, whether or not you are talking to the AI.
Ten rules broken on purpose.
10 of 10 caught
0 false alarms in 20 runs with nothing broken
test on a clean computer, 12 Sep 2026. The records are on GitHub.
The same correction three times becomes a rule.
"again, answer first"
"redo, the answer is still at the bottom"
"not what I asked. Answer first!"
3 times = a pattern
a new light, a new check. From now on the timer runs it 48 times a day.
This film was made with one.
The way these slides move was learned from corrections on earlier slides.
The people who start recording this month will have a year of it by next September.
What did your AI get wrong twice?
$ npx metacognition
metacognition GREEN: 11 of 11 rules held
github.com/jazzs-thoughts/metacognition
One line, pasted into the terminal where you run Claude Code. It asks nothing.
metacognition
click or → next · ← back · F fullscreen · S captions
how to watch
click→←
click or → = next scene · ← = back one scene
F
F = fullscreen on, F again = fullscreen off
S
S = captions off, S again = captions on
01Your AI ignored a rule this week. You saw zero. Nobody sees it, because a sentence in a rules file has no alarm.
click, or press → for the next step