LLMs skip steps when unsupervised and generate plausible-sounding
justifications for doing so. Calling out the common rationalizations
explicitly makes them harder to act on.

Adapted from https://github.com/ed3dai/ed3d-plugins/blob/25eab6a/plugins/ed3d-extending-claude/skills/writing-claude-directives/SKILL.md#L241-L254

Assisted-by: Claude Opus 4.6 via pi

diff --git a/ai/_AGENTS.md b/ai/_AGENTS.md
index b19f1b61..a8d6a568 100644
--- a/ai/_AGENTS.md
+++ b/ai/_AGENTS.md
@@ -126,3 +126,16 @@ Commit quality:
 - Each commit should be one logical unit of change
 - Every commit must build and pass checks (bisect-able history)
 - Separate formatting and refactoring from feature changes
+
+## Anti-rationalization
+
+These conventions exist because LLMs skip steps when unsupervised.
+If you catch yourself thinking any of these, stop — you're
+rationalizing a shortcut:
+
+- "This is simple enough to skip"
+- "I already tested manually"
+- "The spirit, not the letter"
+- "This case is different"
+
+Follow the process.