Install any skill in seconds. Free to start, no credit card required.
Get Started Free →用户学任何新概念/新技术/新理论的默认深度入口——一次性用五个视角把概念讲透并帮他选深入方向:crossover 用已会的撬动、occam 框定该学多深、graph 建知识地图、prototype 最小原型迭代、feynman 拷问检验。触发场景:我想学 X、理解 X、X 是什么、讲讲 X、搞懂 X、学一下 X、深入 X、给我讲讲 X。除非用户明确只要某一个视角(那时改用对应的单个 learn-* skill)。
.claude/skills/li-evan-learn-deep/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -47% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 79% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 26% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 10% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 6% | 0% |
> 把 learn-crossover / learn-occam / learn-graph / learn-prototype / learn-feynman 五个视角编排成一遍全景,给用户学任何概念的"一次扫透 + 选方向"。
用户说"想学 / 理解 / 搞懂 / 讲讲一个概念 X"时——这是默认入口,一次跑完五视角,用户再选深入哪个。 例外:用户明确只要某一个角度("用跨界讲""帮我建图谱""考考我")→ 直接用对应的单个 learn-* skill,别全跑。
先问清用户的背景:学过哪些相关领域、做过什么、熟悉哪些工具 / 理论。后面 crossover / occam / graph 都要用到。只采纳用户亲口确认学过的。
抓住 X 的本质结构(剥术语),按三猜想给 🎁其实已学过 / 🔗结构同构(字段级对应表)/ 🧩可用已有知识解释,点出元知识。先激发信心,再谈深入。
定位"既定问题"(学 X 解决什么)、现有知识够不够、X 的贬值速度与 ROI,给"够用就停 / 只学最小那块 / 值得深挖"的深度边界。不是劝退,是防止一上来过度钻。
X 在所属领域的知识图谱骨架(概念/用途/父子节点),标复用价值最高的节点 + 从常识能入门的点,给学习路径。引导用户补节点(自己建图才学得到)。
给"最垃圾但能跑的原型"起点 + 引导式提问(让用户自己洞察缺陷),预告会撞到的坑。不替他做。
让用户用自己的话答,答不顺处 = 没真懂的洞。最后一个问题尽量打在 X 的根本局限上(真懂的试金石)。
明确推荐往哪 1–2 个方向深入(综合 occam 的 ROI 判断 + 用户的目标 + 哪个视角最戳中他),并指出对应该接哪个单 skill(要动手→learn-prototype,要验收→learn-feynman)。
> ⚠️ 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。
learn-crossover learn-occam learn-graph learn-prototype 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 | 26,563 | 55,472 | +109% | 1 | 1 | 0% | 3,728 | 6,884 | +85% | 0 | 0 | — |
case-02 | fail→pass | 25,194 | 9,486 | -62% | 1 | 1 | 0% | 4,108 | 2,159 | -47% | 0 | 0 | — |
case-03 | fail→pass | 24,201 | 40,985 | +69% | 1 | 1 | 0% | 4,010 | 7,190 | +79% | 0 | 0 | — |
case-04 | pass→pass | 19,320 | 18,229 | -6% | 1 | 1 | 0% | 2,772 | 3,580 | +29% | 0 | 0 | — |
case-05 | pass→pass | 19,366 | 16,737 | -14% | 1 | 1 | 0% | 2,641 | 3,589 | +36% | 0 | 0 | — |
case-06 | pass→pass | 8,609 | 11,835 | +37% | 1 | 1 | 0% | 1,388 | 2,753 | +98% | 0 | 0 | — |
case-07 | pass→pass | 17,149 | 22,213 | +30% | 1 | 1 | 0% | 2,725 | 4,112 | +51% | 0 | 0 | — |
case-08 | pass→pass | 35,477 | 22,081 | -38% | 1 | 1 | 0% | 4,974 | 4,219 | -15% | 0 | 0 | — |
case-09 | pass→pass | 21,779 | 23,354 | +7% | 1 | 1 | 0% | 3,055 | 4,767 | +56% | 0 | 0 | — |
case-10 | fail→pass | 25,087 | 27,451 | +9% | 1 | 1 | 0% | 3,915 | 4,944 | +26% | 0 | 0 | — |
case-11 | fail→fail | 26,974 | 25,777 | -4% | 1 | 1 | 0% | 3,960 | 4,868 | +23% | 0 | 0 | — |
case-12 | fail→pass | 29,137 | 24,544 | -16% | 1 | 1 | 0% | 4,179 | 4,592 | +10% | 0 | 0 | — |
case-13 | fail→pass | 27,159 | 22,042 | -19% | 1 | 1 | 0% | 3,728 | 3,958 | +6% | 0 | 0 | — |
case-14 | fail→pass | 23,535 | 22,318 | -5% | 1 | 1 | 0% | 3,666 | 4,024 | +10% | 0 | 0 | — |
case-15 | fail→fail | 29,603 | 21,137 | -29% | 1 | 1 | 0% | 4,174 | 4,316 | +3% | 0 | 0 | — |
case-16 | fail→fail | 20,317 | 26,184 | +29% | 1 | 1 | 0% | 3,368 | 5,129 | +52% | 0 | 0 | — |
case-17 | fail→pass | 24,977 | 19,014 | -24% | 1 | 1 | 0% | 3,896 | 3,706 | -5% | 0 | 0 | — |
case-18 | fail→pass | 27,405 | 36,380 | +33% | 1 | 1 | 0% | 4,108 | 6,539 | +59% | 0 | 0 | — |
case-19 | fail→pass | 17,557 | 23,355 | +33% | 1 | 1 | 0% | 3,121 | 4,559 | +46% | 0 | 0 | — |
case-20 | fail→fail | 31,195 | 24,534 | -21% | 1 | 1 | 0% | 4,587 | 4,570 | -0% | 0 | 0 | — |
case-21 | fail→pass | 30,186 | 26,545 | -12% | 1 | 1 | 0% | 4,610 | 4,910 | +7% | 0 | 0 | — |
case-22 | fail→fail | 34,829 | 27,829 | -20% | 1 | 1 | 0% | 5,231 | 5,360 | +2% | 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 +45 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.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 8/28/2026 | +23% |
Other measured skills in the registry, with their headline benchmark lift.