Skip to content

v5.0 — 把"防诱导"从靠人升级为靠机制

Latest
Latest

Choose a tag to compare

View all tags
@CyannSHI CyannSHI released this 03 Jun 03:28
· 6 commits to main since this release
v5.0
5bf04cb

这一版解决了什么

复盘 3500+ 通真实外呼（本次新增 1000+ 通）后我们发现：同一套方法论、同样的业务，不同人写出的 Prompt，封闭诱导问题率能差出 3 倍（33% vs 89–100%）。差距不在业务理解，而在 Prompt 有没有"防诱导补丁"。这一版把这些原本靠经验注意的细节，全部固化成了框架机制。

核心更新

  1. 提问前自检清单（最关键） — AI 生成每个问题前逐条自检"比如/还是/对吗/如果XX"四类诱导信号，命中即重写。
  2. 内部参考保护 — 研究目标里斜杠分隔的分析维度自动标注「严禁念出」，杜绝把分析框架当选择题抛给用户。
  3. 生成时风险扫描（Step 9a） — 自动扫描 5 类高风险问题并补上桥接/止损/快速通道规则。
  4. 6 大方法论全量对齐 — 每个方法论的「严禁示例」从只覆盖"是不是"扩展到 5 种诱导模式全覆盖。
  5. 重复与不耐烦兜底 — 新增"换皮重复"拦截和"不耐烦响应"处理，减少提前挂断。

关键指标

  • 封闭诱导问题率：优质 Prompt 33% vs 缺补丁 89–100%
  • 数据基础：本次新增 1000+ 通，累计 3500+ 通真实外呼复盘
  • 防护层级：4 层机制化补丁，覆盖全部 6 种方法论

What this release fixes (EN)

After reviewing 3,500+ real outbound calls (1,000+ new in this round), we found the closed/leading-question rate can differ by 3× (33% vs 89–100%) depending on who writes the prompt — not on domain knowledge. This release hard-wires the "anti-leading patches" into the framework.

Highlights

  1. Pre-question self-check — the AI audits every question against 4 leading signals ("for example / or / right? / what if X") before sending; any hit triggers a rewrite.
  2. Internal-reference protection — slash-separated analysis dimensions are auto-flagged as "never read aloud."
  3. Risk scan at generation time (Step 9a) — scans 5 high-risk question types and auto-appends bridge / stop-loss / quick-exit rules.
  4. All 6 methodologies aligned — "forbidden examples" expanded to cover all 5 leading patterns.
  5. Repetition & impatience fallback — fewer early hang-ups.

Key Metrics

  • Closed/leading-question rate: 33% (strong prompt) vs 89–100% (un-patched)
  • Data basis: 1,000+ new calls reviewed (3,500+ total)
  • Protection layers: 4 mechanized patches across all 6 methodologies
Assets 2
Loading