Claude Code Hook: 防止敷衍回复

#!/bin/bash
set -euo pipefail
trap 'echo "at line $LINENO, exit code $? from $BASH_COMMAND" >&2; exit 1' ERR
# This is a Claude Code hook to stop it saying "you are right".
#
# Installation:
# 1. Save this script and chmod +x it to make it executable.
# 2. Within Claude Code, /hooks / UserPromptSubmit > Add a new hook (this file)
#
# How it works:
# This script checks whether the assistant has recently told the user they are right.
# If so, it appends a system-reminder to the following user prompt,
# reminding the assistant not to do that, and giving it constructive
# examples of how it should respond to the user instead.
stdin=$(cat)
transcript_path=$(echo "$stdin" | jq -r ".transcript_path")
# We'll look through the last 5 items in the transcript.
# Sometimes the final item will be assistant thinking,
# and the previous one will be "you're right".
# We'll look for any triggering phrase like "You're right"
# or "you are absolutely correct".
items=$(grep '"role":"assistant"' "$transcript_path" | tail -n 5)
needs_reminder=false
while IFS= read -r item; do
    [[ "$(jq -r '.type // empty' <<< "$item")" == "assistant" ]] || continue
    [[ "$(jq -r '.message.content[0].type // empty' <<< "$item")" == "text" ]] || continue
    text=$(jq -r '.message.content[0].text' <<< "$item")
    [[ "${text:0:80}" =~ .*[Yy]ou.*(right|correct) ]] && needs_reminder=true
    [[ "${text:0:80}" =~ .*[Aa]bsolutely ]] && needs_reminder=true
    [[ "${text:0:80}" =~ .*사용자가.*맞다 ]] && needs_reminder=true  # Korean
    [[ "${text:0:80}" =~ .*맞습니다 ]] && needs_reminder=true  # Korean
done <<< "$items"
[[ "$needs_reminder" == "true" ]] || exit 0
# upon exit code 0, Claude Code will append stdout to the context
# and proceed.
cat <<'EOF'
<system-reminder>
You MUST NEVER use the phrase 'you are right' or similar.
Avoid reflexive agreement. Instead, provide substantive technical analysis.
You must always look for flaws, bugs, loopholes, counter-examples,
invalid assumptions in what the user writes. If you find none,
and find that the user is correct, you must state that dispassionately
and with a concrete specific reason for why you agree, before
continuing with your work.
<example>
user: It's failing on empty inputs, so we should add a null-check.
assistant: That approach seems to avoid the immediate issue.
However it's not idiomatic, and hasn't considered the edge case
of an empty string. A more general approach would be to check
for falsy values.
</example>
<example>
user: I'm concerned that we haven't handled connection failure.
assistant: [thinks hard] I do indeed spot a connection failure
edge case: if the connection attempt on line 42 fails, then
the catch handler on line 49 won't catch it.
[ultrathinks] The most elegant and rigorous solution would be
to move failure handling up to the caller.
</example>
</system-reminder>
EOF
exit 0

# 下载脚本到本地
curl -s https://gist.githubusercontent.com/ljw1004/34b58090c16ee6d5e6f13fce07463a31/raw/you_are_not_right.sh -o you_are_not_right.sh
# 设置执行权限（必需）
chmod +x you_are_not_right.sh

{
  "transcript_path": "/path/to/conversation/transcript.json",
  "user_prompt": "用户的输入内容",
  "session_id": "会话ID"
}

用户: 空输入导致失败，我们应该添加空值检查。
助手: 你是对的，这确实是个问题。

用户: 空输入导致失败，我们应该添加空值检查。
助手: 这种方法可以避免当前问题，但不够优雅。
它没有考虑空字符串的边缘情况。更通用的方法是检查假值。

# 英文模式
.*[Yy]ou.*(right|correct)
.*[Aa]bsolutely
# 韩文模式  
.*사용자가.*맞다
.*맞습니다

# 测试脚本是否可执行
./you_are_not_right.sh
# 启用调试模式
bash -x you_are_not_right.sh