Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →当用户要系统学一个新领域、不知道从哪入手、或担心「学得不够系统」时使用。用「知识图谱学习法」和用户一起构建该领域的概念/用途/父子节点图谱（自己建图的过程本身就是学习），标出复用价值最高的节点和「从常识就能入门的点」，给出有效学习路径并回答「学到哪算够」。触发场景：系统学 X 领域、从哪开始学、学得不系统、想要 X 的全貌、规划学习路径、这个领域有多大。
.claude/skills/li-evan-learn-graph/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | -7% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -27% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 30% | 0% |
| case-17 | ✗→✓ | ▲ Improved | 9% | 0% |
| case-19 | ✓→✗ | ▼ Worse | 3% | 0% |
> 核心信条：自己一步步建图谱的过程，本身就是最有效的学习——不要直接套用别人给的图谱。 绝大部分知识，都有一个从常识就能入门的点。
用户要系统进入一个新领域，或焦虑"学得不够系统 / 不知何时算够"。
用户为什么学 X？（接 learn-occam 的"既定问题"）目的决定图谱画到多细。
概念/名称 · 用途 · 上下文关系（父子节点）：
从入门点出发、沿父子关系排一条有效路径。颗粒度按需自由切换（领域图 → 细分学科图）。"学到哪算够"= 覆盖到能解决第一步那个目的的节点即可，不必学满。
learn-prototype（在图上找"最垃圾原型"的起点）。> ⚠️ 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-occam(该不该学) learn-crossover(已会什么) learn-prototype(动手) learn-feynman(自查)。| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-02 | fail→fail | 10,593 | 11,692 | +10% | 1 | 1 | 0% | 1,402 | 2,279 | +63% | 0 | 0 | — |
case-01 | fail→fail | 14,992 | 15,854 | +6% | 1 | 1 | 0% | 2,327 | 2,472 | +6% | 0 | 0 | — |
case-03 | fail→fail | 15,587 | 11,397 | -27% | 1 | 1 | 0% | 2,174 | 2,298 | +6% | 0 | 0 | — |
case-04 | pass→pass | 21,307 | 16,574 | -22% | 1 | 1 | 0% | 3,457 | 3,015 | -13% | 0 | 0 | — |
case-05 | pass→pass | 21,872 | 14,325 | -35% | 1 | 1 | 0% | 3,381 | 2,670 | -21% | 0 | 0 | — |
case-06 | pass→pass | 18,713 | 18,662 | -0% | 1 | 1 | 0% | 2,817 | 3,475 | +23% | 0 | 0 | — |
case-07 | fail→pass | 22,726 | 16,365 | -28% | 1 | 1 | 0% | 3,204 | 2,985 | -7% | 0 | 0 | — |
case-08 | pass→pass | 17,615 | 17,755 | +1% | 1 | 1 | 0% | 2,929 | 3,303 | +13% | 0 | 0 | — |
case-09 | pass→pass | 15,582 | 11,624 | -25% | 1 | 1 | 0% | 2,287 | 2,110 | -8% | 0 | 0 | — |
case-10 | fail→pass | 13,438 | 5,587 | -58% | 1 | 1 | 0% | 1,818 | 1,326 | -27% | 0 | 0 | — |
case-11 | fail→pass | 8,985 | 7,367 | -18% | 1 | 1 | 0% | 1,216 | 1,583 | +30% | 0 | 0 | — |
case-12 | pass→pass | 15,794 | 11,278 | -29% | 1 | 1 | 0% | 2,453 | 2,124 | -13% | 0 | 0 | — |
case-13 | pass→pass | 18,403 | 16,596 | -10% | 1 | 1 | 0% | 2,672 | 3,025 | +13% | 0 | 0 | — |
case-14 | pass→pass | 13,365 | 3,710 | -72% | 1 | 1 | 0% | 1,863 | 1,055 | -43% | 0 | 0 | — |
case-15 | pass→pass | 21,193 | 15,873 | -25% | 1 | 1 | 0% | 3,110 | 2,714 | -13% | 0 | 0 | — |
case-16 | fail→fail | 15,752 | 12,952 | -18% | 1 | 1 | 0% | 2,269 | 2,360 | +4% | 0 | 0 | — |
case-17 | fail→pass | 20,340 | 17,827 | -12% | 1 | 1 | 0% | 2,767 | 3,022 | +9% | 0 | 0 | — |
case-18 | pass→pass | 18,899 | 13,948 | -26% | 1 | 1 | 0% | 2,447 | 2,520 | +3% | 0 | 0 | — |
case-19 | pass→fail | 16,431 | 13,358 | -19% | 1 | 1 | 0% | 2,352 | 2,428 | +3% | 0 | 0 | — |
case-20 | pass→fail | 16,705 | 15,635 | -6% | 1 | 1 | 0% | 3,029 | 2,902 | -4% | 0 | 0 | — |
case-21 | pass→pass | 9,139 | 9,931 | +9% | 1 | 1 | 0% | 1,180 | 1,901 | +61% | 0 | 0 | — |
case-22 | pass→fail | 21,048 | 15,490 | -26% | 1 | 1 | 0% | 3,167 | 2,919 | -8% | 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 +5 percentage points is the difference between those two pass rates over the 22 comparable cases. 3 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.