Prompt instructions are soft constraints that RLHF-trained deference and hedging can override under pressure. Claude Code stop hooks run after generation and before you see the output, and can force a revision — here's the AND-gate pattern and two production hooks I run.

Claude Code hooks run shell commands at lifecycle events — auto-format after edits, block protected files, audit every command. Config, exit codes, recipes.

Prompt instructions are soft constraints that RLHF-trained deference and hedging can override under pressure. Claude Code stop hooks run after generation and before you see the…