PRD 评审与团队对齐机制

PRD 评审会最常见的结局是：所有人说"没问题"，然后开发一半发现理解完全不同。对齐不是"没有反对意见"，是"每个人对关键决策的理解一致"。

PRD 评审的标准流程

评审前：PM 自检清单

发出 PRD 之前，PM 必须先对自己的文档做一轮自检。以下问题你应该能直接回答：

常见对齐偏差模式

AI 产品评审中，以下四种偏差最常出现：

Python 示例：PRD 评审会追踪器

"""
PRD 评审会追踪器
PM 视角：记录评审会的讨论、分歧点和决策，确保对齐可追溯
"""
from dataclasses import dataclass, field
from typing import List, Optional, Dict
from enum import Enum
from datetime import date, timedelta
class AlignmentStatus(Enum):
ALIGNED = "已对齐"
PENDING = "待决策"
DEFERRED = "延后讨论"
BLOCKED = "阻塞需升级"
class StakeholderRole(Enum):
PM = "产品经理"
DESIGN = "设计师"
ENGINEER = "研发工程师"
AI_ENGINEER = "AI 工程师"
QA = "测试工程师"
BUSINESS = "业务方"
@dataclass
class ReviewPoint:
"""评审讨论点"""
point_id: str
section: str                # PRD 章节
description: str
raised_by: StakeholderRole
status: AlignmentStatus
decision: Optional[str] = None
decision_owner: Optional[StakeholderRole] = None
deadline: Optional[date] = None
impact_if_unresolved: str = ""
@dataclass
class ReviewSession:
"""评审会记录"""
prd_name: str
prd_version: str
review_date: date
attendees: List[StakeholderRole]
facilitator: str
review_points: List[ReviewPoint] = field(default_factory=list)
pre_read_sent_days_before: int = 0
@dataclass
class ReviewReport:
"""评审报告"""
session: ReviewSession
alignment_rate: float
critical_pending: List[ReviewPoint]
overdue_points: List[ReviewPoint]
readiness_assessment: str
def calculate_alignment_rate(points: List[ReviewPoint]) -> float:
"""计算对齐率"""
if not points:
return 1.0
aligned = sum(1 for p in points if p.status == AlignmentStatus.ALIGNED)
return aligned / len(points)
def find_critical_pending(points: List[ReviewPoint]) -> List[ReviewPoint]:
"""找出关键未决项（阻塞性）"""
return [
p for p in points
if p.status in (AlignmentStatus.PENDING, AlignmentStatus.BLOCKED)
]
def find_overdue_points(points: List[ReviewPoint],
reference_date: date) -> List[ReviewPoint]:
"""找出已过期未决项"""
return [
p for p in points
if p.status == AlignmentStatus.PENDING
and p.deadline is not None
and p.deadline < reference_date
]
def assess_readiness(session: ReviewSession,
alignment_rate: float,
critical_pending: List[ReviewPoint]) -> str:
"""评估 PRD 是否达到可开发状态"""
issues = []
if session.pre_read_sent_days_before < 1:
issues.append("预读时间不足（建议提前 48 小时发出）")
if StakeholderRole.AI_ENGINEER not in session.attendees:
issues.append("AI 工程师未参加评审，AI 能力边界可能未确认")
if StakeholderRole.QA not in session.attendees:
issues.append("测试工程师未参加评审，验收标准可能未对齐")
blocked_points = [p for p in critical_pending
if p.status == AlignmentStatus.BLOCKED]
if blocked_points:
issues.append(f"存在 {len(blocked_points)} 个阻塞问题，需升级决策后才能继续")
if alignment_rate < 0.6:
issues.append(f"对齐率仅 {alignment_rate:.0%}，分歧过多，建议再次评审")
if issues:
return "未达到开发就绪状态：\n" + "\n".join(f"  - {i}" for i in issues)
if alignment_rate < 0.8:
return f"基本可以开始开发，但有 {len(critical_pending)} 个待决项需在开发启动前解决"
return "已达到开发就绪状态，可以进入开发排期"
def generate_review_report(session: ReviewSession) -> ReviewReport:
"""生成评审报告"""
today = date.today()
alignment_rate = calculate_alignment_rate(session.review_points)
critical_pending = find_critical_pending(session.review_points)
overdue = find_overdue_points(session.review_points, today)
readiness = assess_readiness(session, alignment_rate, critical_pending)
return ReviewReport(
session=session,
alignment_rate=alignment_rate,
critical_pending=critical_pending,
overdue_points=overdue,
readiness_assessment=readiness,
)
def print_review_report(report: ReviewReport):
session = report.session
print(f"\n{'='*55}")
print(f"  PRD 评审报告：{session.prd_name} v{session.prd_version}")
print(f"  评审日期：{session.review_date}  主持人：{session.facilitator}")
print(f"  参会角色：{', '.join(r.value for r in session.attendees)}")
print(f"{'='*55}")
print(f"\n  总体对齐率：{report.alignment_rate:.0%}  "
f"（共 {len(session.review_points)} 个讨论点）")
# 按状态统计
status_counts: Dict[AlignmentStatus, int] = {}
for point in session.review_points:
status_counts[point.status] = status_counts.get(point.status, 0) + 1
for status, count in status_counts.items():
print(f"    {status.value}：{count} 条")
if report.critical_pending:
print(f"\n  待决/阻塞项（{len(report.critical_pending)} 条）：")
for p in report.critical_pending:
deadline_str = f"（截止 {p.deadline}）" if p.deadline else ""
owner_str = f"  负责人：{p.decision_owner.value}" if p.decision_owner else ""
print(f"\n    [{p.point_id}] [{p.status.value}] {p.section}")
print(f"    问题：{p.description}")
print(f"    影响：{p.impact_if_unresolved}{deadline_str}{owner_str}")
if report.overdue_points:
print(f"\n  已过期未决项（{len(report.overdue_points)} 条，需立即处理）：")
for p in report.overdue_points:
print(f"    ⚠ [{p.point_id}] {p.description}（截止 {p.deadline}）")
print(f"\n  开发就绪评估：")
print(f"  {report.readiness_assessment}")
# ── 演示 ──────────────────────────────────────────────
if __name__ == "__main__":
today = date.today()
session = ReviewSession(
prd_name="AI 意图识别与自动回复",
prd_version="1.2",
review_date=today,
attendees=[
StakeholderRole.PM,
StakeholderRole.ENGINEER,
StakeholderRole.AI_ENGINEER,
StakeholderRole.QA,
],
facilitator="张三（PM）",
pre_read_sent_days_before=2,
review_points=[
ReviewPoint(
point_id="R01",
section="03-AI 能力描述",
description="置信度阈值 0.85 的依据是什么？工程侧认为实际很难稳定达到",
raised_by=StakeholderRole.AI_ENGINEER,
status=AlignmentStatus.PENDING,
decision_owner=StakeholderRole.PM,
deadline=today + timedelta(days=2),
impact_if_unresolved="阈值不确定会导致用户体验设计无法定稿",
),
ReviewPoint(
point_id="R02",
section="04-失败机制",
description="超时降级时转人工队列的优先级如何？是否插队？",
raised_by=StakeholderRole.ENGINEER,
status=AlignmentStatus.ALIGNED,
decision="超时转入高优队列，与用户主动转人工同等优先级",
decision_owner=StakeholderRole.PM,
),
ReviewPoint(
point_id="R03",
section="05-验收标准",
description="测试集 500 条由谁标注？标注规范是否已有？",
raised_by=StakeholderRole.QA,
status=AlignmentStatus.PENDING,
decision_owner=StakeholderRole.AI_ENGINEER,
deadline=today + timedelta(days=3),
impact_if_unresolved="验收测试无法启动",
),
ReviewPoint(
point_id="R04",
section="02-用户场景",
description="多图片消息场景未在 PRD 中描述，但前端已有该入口",
raised_by=StakeholderRole.DESIGN,
status=AlignmentStatus.DEFERRED,
decision="本期仅支持文字，图片场景列入 v1.3 讨论",
decision_owner=StakeholderRole.PM,
),
ReviewPoint(
point_id="R05",
section="06-上线计划",
description="Beta 阶段流量分配机制（随机 / 按用户分层）未定义",
raised_by=StakeholderRole.ENGINEER,
status=AlignmentStatus.ALIGNED,
decision="按注册时间随机分组，同用户在整个 Beta 期间保持分组不变",
decision_owner=StakeholderRole.ENGINEER,
),
],
)
report = generate_review_report(session)
print_review_report(report)

本章 checklist

• [ ] 评审文档在会议开始前 48 小时发出，给与会者预读时间
• [ ] AI 工程师和测试工程师都出席了评审会
• [ ] 评审会上每个分歧点都记录了负责人和截止时间
• [ ] 评审后 24 小时内，所有待决项的决策都通知给相关方
• [ ] PRD 版本有明确的"已对齐"确认记录，可以作为后续变更的基准

本章小结

• 评审会的目的不是"过稿"，是让每个角色能独立复述关键决策，做不到就说明没有真正对齐
• AI 产品评审特别需要 AI 工程师在场确认能力边界，PM 不能替代这个角色发言
• 分歧点要公开记录并给出决策截止时间，让"没有明确反对"变成有记录的"已确认"

本章完：下一章进入 10-协作编排，学习如何与设计师、工程师、AI 工程师建立高效的协作机制。