Install any skill in seconds. Free to start, no credit card required.
Get Started Free →当用户学习或接触一个新概念/新技术/新算法/新领域时使用(尤其感到陌生或有点难时)。用「跨界原则」拿用户已掌握的知识快速撬动新知识——指出他其实已经学过的同一个东西(换了名字)、结构同构的旧知识、能解释新知识的已有知识,并点出新概念体现的跨领域元知识模式。让「学新东西」变成「发现你已经会了一半」。触发场景:学 X、接触 X、这个好难、X 是什么、帮我理解 X。
.claude/skills/li-evan-learn-crossover/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | 87% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 36% | 0% |
| case-19 | ✗→✓ | ▲ Improved | 55% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 29% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 41% | 0% |
> 核心信条:真正的快速学会,其实是「你已经学过了」。 跨界匹配的是结构,不是名词。
用户在学 / 接触一个新概念 X(新技术、新算法、新理论、新领域……),尤其觉得"陌生 / 有点难"的时候。难,往往不是智商问题,是它相对用户还存在"没接上的旧知识"。
用一两句话说清 X 到底在干什么——它的核心机制 / 结构是什么。剥掉术语外壳,留下"它本质是一个 ___"。只有先拿到结构,才能去匹配用户学过的东西。
主动询问,建立"用户已掌握知识"的清单:
拿不准就直接问「你学过 ___ 吗?」,绝不从正在讲的材料 / 文章作者背景推断用户会什么。
X 体现了哪个反复出现的底层模式?(分治、自举 / bootstrap、阻尼-负反馈、探索 vs 应用、量变质变、控制变量、状态机……)。告诉用户"这个模式你在 ___、___ 也见过",把 X 挂到他的元知识网上。
一句话收尾,降低学习恐惧 + 指明剩下要新学的最小部分:
> "所以 X 你已经会了 ___ 部分,真正全新、需要从头学的只有 ___。"
> ⚠️ 铁律·只用确证的已会知识:判断用户「已经会什么」只能用他确证学过的知识(亲口确认或可靠背景);严禁把「正在讲的材料 / 文章作者背景 / 对话里别人的知识」当成用户会的。拿不准 → 直接问「⚠️ 你学过 ___ 吗?」,绝不替他假设。(最常见的翻车点:把材料作者的背景错安到学习者头上,整段跨界作废。)
learn-occam,要"系统建图"用 learn-graph,要"动手迭代"用 learn-prototype,要"自查懂没懂"用 learn-feynman。| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 21,652 | 27,784 | +28% | 1 | 1 | 0% | 2,932 | 4,040 | +38% | 0 | 0 | — |
case-02 | pass→pass | 12,996 | 17,964 | +38% | 1 | 1 | 0% | 2,052 | 3,649 | +78% | 0 | 0 | — |
case-03 | pass→fail | 53,786 | 18,504 | -66% | 1 | 1 | 0% | 8,221 | 3,620 | -56% | 0 | 0 | — |
case-04 | fail→fail | 4,224 | 8,616 | +104% | 1 | 1 | 0% | 606 | 2,184 | +260% | 0 | 0 | — |
case-05 | fail→pass | 11,898 | 18,580 | +56% | 1 | 1 | 0% | 2,010 | 3,752 | +87% | 0 | 0 | — |
case-17 | pass→pass | 18,841 | 22,391 | +19% | 1 | 1 | 0% | 2,888 | 4,319 | +50% | 0 | 0 | — |
case-18 | fail→pass | 14,414 | 15,518 | +8% | 1 | 1 | 0% | 2,243 | 3,052 | +36% | 0 | 0 | — |
case-19 | fail→pass | 18,486 | 20,071 | +9% | 1 | 1 | 0% | 2,431 | 3,775 | +55% | 0 | 0 | — |
case-06 | pass→pass | 18,888 | 16,140 | -15% | 1 | 1 | 0% | 3,239 | 3,610 | +11% | 0 | 0 | — |
case-07 | fail→pass | 18,785 | 18,001 | -4% | 1 | 1 | 0% | 2,892 | 3,717 | +29% | 0 | 0 | — |
case-08 | pass→pass | 18,886 | 20,996 | +11% | 1 | 1 | 0% | 3,279 | 3,932 | +20% | 0 | 0 | — |
case-09 | fail→pass | 16,469 | 14,344 | -13% | 1 | 1 | 0% | 2,265 | 3,183 | +41% | 0 | 0 | — |
case-10 | fail→fail | 20,295 | 23,682 | +17% | 1 | 1 | 0% | 2,898 | 4,223 | +46% | 0 | 0 | — |
case-11 | fail→fail | 17,261 | 18,716 | +8% | 1 | 1 | 0% | 2,963 | 3,706 | +25% | 0 | 0 | — |
case-12 | fail→pass | 19,016 | 21,591 | +14% | 1 | 1 | 0% | 2,479 | 3,868 | +56% | 0 | 0 | — |
case-13 | pass→pass | 23,803 | 20,933 | -12% | 1 | 1 | 0% | 3,246 | 3,977 | +23% | 0 | 0 | — |
case-14 | fail→fail | 17,447 | 20,297 | +16% | 1 | 1 | 0% | 2,728 | 3,950 | +45% | 0 | 0 | — |
case-15 | pass→pass | 16,687 | 15,705 | -6% | 1 | 1 | 0% | 2,224 | 3,437 | +55% | 0 | 0 | — |
case-16 | fail→fail | 19,252 | 22,497 | +17% | 1 | 1 | 0% | 2,984 | 3,979 | +33% | 0 | 0 | — |
case-20 | fail→fail | 15,284 | 14,690 | -4% | 1 | 1 | 0% | 2,360 | 3,168 | +34% | 0 | 0 | — |
case-21 | pass→pass | 16,531 | 15,326 | -7% | 1 | 1 | 0% | 2,197 | 3,161 | +44% | 0 | 0 | — |
case-22 | fail→pass | 12,668 | 18,476 | +46% | 1 | 1 | 0% | 2,187 | 3,926 | +80% | 0 | 0 | — |
case-23 | fail→pass | 19,837 | 18,476 | -7% | 1 | 1 | 0% | 2,628 | 3,711 | +41% | 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. 23 cases were attempted. The headline lift of +30 percentage points is the difference between those two pass rates over the 23 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
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.