summersec Semantic Trap Word Detector (Shot Mode)
A concentrated, single-file reference for detecting semantic trap words in LLM instruction files. No plugin installation required — just read and apply.
Source of truth（与完整插件的关系）
本文件中的表格是便携速查。权威词表、严重等级与失控场景说明以仓库内 references/semantic-trap-lexicon.md 为准；安装完整插件时，运行时数据由 npm run build-lexicon 从该 MD 生成 lib/lexicon-data.js。若速查表与 MD 不一致，以 MD / 生成结果为准。
Gotchas
- 表格用于人工扫读，不保证与当前分支 MD 逐字同步；发版前以 MD 为准核对 ID 与宽/窄词对。
- 替换时保持原意图：目标是收窄语义边界，不是消灭所有抽象词。
- 示例句式若写入你自己的 Skill，可能触发 linter；生产文案请用窄边界改写后的版本。
What Are Semantic Traps?
Semantic trap words are vocabulary with wide semantic boundaries that cause LLMs to produce outputs far beyond intended scope. Empirical note (see linked article in repo README): swapping a narrow defect-focused noun for a broad finance-adjacent noun in an otherwise identical Skill can materially reduce task accuracy.
Core principle: Replace wide-boundary words with narrow-boundary alternatives to constrain LLM output.
Trap Word Reference Table
以下宽词表为速查对照；为免在本仓库触发 linter，表体放在围栏块内（阅读时照常查看即可）。
Chinese (T01-T17)
| ID | Trap Word (Wide) | Replacement (Narrow) | Severity | Why It's Dangerous |
|----|-------------------|----------------------|----------|-------------------|
| T01 | 风险 | 漏洞 | critical | Activates finance/health/legal associations |
| T02 | 审查 | 检查 | high | Triggers subjective evaluation |
| T03 | 描述 | 列出 | high | Triggers explanatory prose |
| T04 | 问题 | 缺陷 | high | Extremely wide scope |
| T05 | 分析 | 总结 | medium-high | Triggers inference and hypothesis |
| T06 | 建议 | 要求 | medium-high | Triggers divergent thinking |
| T07 | 异常 | 错误 | medium | Fuzzy boundary |
| T08 | 改善 | 修复 | high | Triggers optimization suggestions |
| T09 | 参考 | 遵循 | medium-high | Implies flexibility |
| T10 | 评估 | 统计 | high | Triggers subjective judgment |
| T11 | 理解 | 提取 | high | Triggers inference |
| T12 | 关联 | 匹配 | medium | Triggers indirect reasoning |
| T13 | 转化 | 复制 | medium-high | Implies modification allowed |
| T14 | 洞察 | 报告 | high | Triggers creative inference |
| T15 | 评价 | 验证 | medium-high | Triggers subjective scoring |
| T16 | 原则 | 规则 | medium | Implies flexibility |
| T17 | 方法 | 步骤 | medium | Implies choice of paths |
English (E01-E10)
| ID | Trap Word (Wide) | Replacement (Narrow) | Severity | Why It's Dangerous |
|----|-------------------|----------------------|----------|-------------------|
| E01 | Risk | Vulnerability | critical | Same as T01 |
| E02 | Review / Audit | Check | high | Triggers comprehensive evaluation |
| E03 | Describe / Explain | List | high | Triggers narrative expansion |
| E04 | Issue / Problem | Bug / Defect | high | Overly broad coverage |
| E05 | Analyze / Assess | Summarize | medium-high | Triggers deep reasoning |
| E06 | Suggestion / Recommendation | Requirement | medium-high | Triggers divergent thinking |
| E07 | Anomaly / Concern | Error | medium | Nearly boundless scope |
| E08 | Interpret | Extract | high | Triggers subjective interpretation |
| E09 | Evaluate / Judge | Verify | medium-high | Triggers multi-dimensional evaluation |
| E10 | Should / Could | Must / Shall | medium-high | Implies optional, not mandatory |
Structural drift patterns
Beyond individual words, watch for these 4 patterns:
1. Open-Ended Verbs: verb + object without scope → add explicit dimensions or defect types.
2. Abstract Targets: vague safety/quality goals → replace with checkable criteria per endpoint.
3. Modal Downgrades: weak modals in constraints → use mandatory wording where rules must hold.
4. Missing Negation Lists: high-severity wide words without exclusions → add an explicit NOT-in-scope list.
How to Apply
When writing or editing instruction files (.md files in /skills/, /agents/, /rules/, /prompts/):
- Scan each key noun and verb against the table above
- If a wide-boundary word is found, suggest the narrow-boundary replacement
- Check for the four structural drift patterns above
- When replacing, preserve the original intent — the goal is precision, not restriction
Full Plugin
For automated detection with Pre/Post hooks, CLI scanning, escalation tracking, and deep semantic analysis, install the full semantic-linter plugin.