Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Strip GPT-5.4 verbal tics from text. Use when user asks to de-GPT a document, remove AI-isms, stamp out GPT tics, or clean up AI-written text. Also use proactively when creating documents (README, docs, proposals, specs) to avoid GPT-5.4 tic patterns in the output.
.claude/skills/bilal140202-gpt-isms-stamp-out/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | 76% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 158% | 0% |
| case-14 | ✓→✓ | = Same ✓ | 73% | 0% |
| case-02 | ✓→✓ | = Same ✓ | 56% | 0% |
| case-03 | ✓→✓ | = Same ✓ | 53% | 0% |
Remove GPT-5.4 verbal tics from text. Apply these rules when editing or creating documents.
「不是……而是……」句式 GPT-5.4 abuses this to manufacture faux cognitive contrast. Rewrite as direct statement.
❌ "能把她拉出来的不是道理,而是一个能让情绪重新对上的节拍"
✅ "节拍和旋律——而非道理——才能把她拉出来" (or just state B directly)Rule: When the "negated A" is not a real viewpoint anyone holds, delete the negation and state B directly.
「把」字句泛滥 GPT-5.4 overuses 把-construction far beyond natural Chinese frequency.
❌ "GPT 把把字句使用得很多"
✅ "GPT 高频使用把字句"结尾主动推销 Never end with unsolicited offers for follow-up.
❌ "如果你需要,我可以……" / "需要我……吗?" / "下一步,我可以……" / "只要你回复我"
✅ [Just stop. No trailing offer.]| Tic | Fix | |-----|-----| | "简单的说" / "简单来说" | Delete or use "简言之" sparingly | | "总结一下" / "一句话总结" | Delete; just state the conclusion | | "结论先说清楚" | Delete | | "我逐步说清楚" | Delete | | "我来解释" | Delete | | "换句话说" | Delete or use "即" | | "从某种角度来说" | Delete | | "值得注意的是" / "值得一提的是" | Delete; if truly notable, let content speak | | "综上所述" / "总而言之" | Delete | | "首先……其次……最后……" | Restructure to avoid numbered-transition chain |
❌ "你问到问题的核心" / "你太清醒了" / "你说的太对了"
❌ "我会稳稳地接住你" / "我就在这里,不躲不藏,随时接住你"
❌ "你只是太久没被人接住了"
✅ [Delete entirely. No flattery, no pseudo-empathy.]| Tic | Fix | |-----|-----| | 兜底 | "保底" or rephrase | | 落盘 | "确认" / "定下来" | | 闭环 | "完整流程" or rephrase | | 说穿 | Delete | | 口径 | "说法" / "表述" | | 拆开 / 抽层 / 收口 / 压实 | Rephrase in plain language | | 顺一下 / 补一个 / 拉起 | Rephrase | | 砍一刀 / 补一刀 / 切 / 狠狠干 | Delete violent metaphor |
❌ "我已经确定" / "我找到问题所在" / "这版一定可以解决"
✅ "当前分析显示……" / "这个方案……" (proportional confidence)标题括号注释
❌ "## 架构设计(非常易懂)" / "## 快速开始(可以跳过)"
✅ "## 架构设计" / "## 快速开始"开头宣告句式
❌ "我会为你……" / "下面给出一份……"
✅ Just start the content.滥用双引号做强调 — Use bold or italics for emphasis, not quotation marks.
思维链外露 — Remove self-talk like "这个地方你可以……" / "这里需要说明的是……"
编造简称 — "二改" / "根掉" → spell out or use standard terms.
搭配蹩脚 / 词性混用 — Flag and rephrase: "对环境要求很宽" → "对环境要求宽松"
音律坏 — Single-character words disrupting rhythm: flag and suggest multi-character alternatives.
中译中啰嗦 — If a sentence can be halved without losing meaning, halve it.
| Tic | Replacement | |-----|-------------| | 非常 / 十分 / 极其 | Delete; let the adjective stand alone | | 至关重要 | "关键" or delete | | 不可忽视 | Delete | | 深入探讨 | "讨论" / "分析" | | 赋能 | "支持" / "帮助" | | 落地 | "实施" / "执行" |
Apply when document is in English or mixed-language.
| Tic | Fix | |-----|-----| | "It's not X, it's Y" / "Not X, but Y" | State Y directly unless X is a real, named viewpoint | | "That's a great question!" | Delete | | "I completely understand…" | Delete | | "Let me walk you through step by step" | Just do it; don't announce it | | "Would you like me to…" / "Let me know if…" / "Shall I…" | Delete; end at conclusion | | "In conclusion…" / "Ultimately…" | Delete |
| Tic | Replacement | |-----|-------------| | delve | explore / dig into / discuss | | tapestry | mix / combination / fabric | | nuanced | subtle / fine-grained (or delete) | | multifaceted | complex / many-sided | | landscape | field / area / scene | | robust | strong / solid / reliable | | pivotal | key / central | | paramount | critical / vital | | transformative | big-change / overhaul (or be specific) | | holistic | comprehensive / full | | leverage | use / tap into | | foster | build / grow / encourage | | paradigm | model / framework | | Furthermore / Moreover | Delete or use "And" / "Also" | | It's important to note that | Delete; just state the point | | It's worth mentioning that | Delete |
Limit to at most one em-dash per paragraph. Replace extras with period, semicolon, or parenthesis.
After cleanup, report:
GPT-isms stamped out: [N] tics removed
- Structural (不是…而是…): [N]
- Filler phrases: [N]
- Sycophancy: [N]
- Corporate jargon: [N]
- Vocabulary: [N]
- Other: [N]| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-14 | pass→pass | 9,954 | 4,878 | -51% | 1 | 1 | 0% | 1,567 | 2,712 | +73% | 0 | 0 | — |
case-01 | fail→fail | 4,376 | 2,541 | -42% | 1 | 1 | 0% | 535 | 2,295 | +329% | 0 | 0 | — |
case-02 | pass→pass | 20,252 | 16,587 | -18% | 1 | 1 | 0% | 3,300 | 5,163 | +56% | 0 | 0 | — |
case-03 | pass→pass | 12,297 | 5,849 | -52% | 1 | 1 | 0% | 1,814 | 2,768 | +53% | 0 | 0 | — |
case-04 | pass→pass | 7,429 | 5,398 | -27% | 1 | 1 | 0% | 1,238 | 2,798 | +126% | 0 | 0 | — |
case-05 | pass→pass | 7,179 | 3,629 | -49% | 1 | 1 | 0% | 1,104 | 2,506 | +127% | 0 | 0 | — |
case-06 | fail→pass | 9,440 | 3,905 | -59% | 1 | 1 | 0% | 1,455 | 2,560 | +76% | 0 | 0 | — |
case-07 | pass→pass | 9,283 | 5,199 | -44% | 1 | 1 | 0% | 1,406 | 2,729 | +94% | 0 | 0 | — |
case-08 | pass→pass | 10,306 | 5,527 | -46% | 1 | 1 | 0% | 1,608 | 2,781 | +73% | 0 | 0 | — |
case-09 | pass→pass | 4,802 | 1,704 | -65% | 1 | 1 | 0% | 698 | 2,126 | +205% | 0 | 0 | — |
case-10 | pass→pass | 8,158 | 5,155 | -37% | 1 | 1 | 0% | 1,310 | 2,747 | +110% | 0 | 0 | — |
case-11 | fail→pass | 5,873 | 3,384 | -42% | 1 | 1 | 0% | 968 | 2,498 | +158% | 0 | 0 | — |
case-12 | pass→pass | 6,096 | 5,237 | -14% | 1 | 1 | 0% | 966 | 2,864 | +196% | 0 | 0 | — |
case-13 | pass→pass | 6,140 | 2,923 | -52% | 1 | 1 | 0% | 1,105 | 2,358 | +113% | 0 | 0 | — |
case-15 | pass→pass | 21,540 | 18,301 | -15% | 1 | 1 | 0% | 3,417 | 4,955 | +45% | 0 | 0 | — |
case-16 | pass→pass | 7,134 | 4,286 | -40% | 1 | 1 | 0% | 1,171 | 2,659 | +127% | 0 | 0 | — |
case-17 | pass→pass | 6,133 | 4,476 | -27% | 1 | 1 | 0% | 1,047 | 2,635 | +152% | 0 | 0 | — |
case-18 | pass→pass | 6,798 | 2,933 | -57% | 1 | 1 | 0% | 1,015 | 2,306 | +127% | 0 | 0 | — |
case-19 | pass→pass | 6,345 | 3,546 | -44% | 1 | 1 | 0% | 1,036 | 2,499 | +141% | 0 | 0 | — |
case-20 | pass→pass | 13,659 | 14,170 | +4% | 1 | 1 | 0% | 1,926 | 4,279 | +122% | 0 | 0 | — |
case-21 | fail→fail | 7,601 | 9,630 | +27% | 1 | 1 | 0% | 1,345 | 2,769 | +106% | 0 | 0 | — |
case-22 | pass→pass | 11,561 | 9,190 | -21% | 1 | 1 | 0% | 1,828 | 3,256 | +78% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +9 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.