Codeuteronomy
A repeated instruction is a missing constraint
Telling the agent the same thing twice is not steering. It is a defect in your setup, paid for by hand, every session, forever.
An agent does not remember yesterday's correction. Whatever you said last time you will say again, unless you change something other than your patience. The first correction is steering. The second is a signal that this will recur, and that re-prompting is a tax you've agreed to keep paying. The way out is to make the mistake structurally impossible: move the rule out of your head and into the system, where it enforces itself.
Encoding takes many forms: a rule in CLAUDE.md loaded every session, a hook that rejects the action before it runs, a permission deny that removes the capability, a type that makes the wrong value fail to compile. The shape matters less than the direction, from “I keep saying” to “it can no longer do.” This is the one habit that compounds. Every correction you encode is one you never issue again.
How to practice it
The second occurrence is the trigger.
Correct once; encode the second time.
Push it as low as it will go.
A type that won't compile, then a PreToolUse hook in .claude/settings.json, then a permission deny, then a written rule, then a line in the prompt. Prefer whichever the agent cannot route around.
Make always-on rules load themselves
from CLAUDE.md and topic files under .claude/, so the correction outlives the conversation.
Turn recurring procedures into skills.
One /command instead of a re-explanation.
Remove dangerous capabilities outright
with a permissions deny, rather than repeatedly asking for restraint.