Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →当用户学完一个东西想自查是否真懂、或觉得「好像懂了」但不确定时使用。用「费曼学习法」让他用自己的话把概念讲出来，你扮好奇学生专挑他含糊/跳过的地方追问，把「讲不顺的模糊处」揪出来作为没真懂的漏洞，定位是缺前置知识还是没想透，判断理解是否闭环。触发场景：我学完了考考我、自查一下、我好像懂了、我讲讲你看对不对、检验我的理解、这个我真懂了吗。
.claude/skills/li-evan-learn-feynman/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-08 | ✗→✓ | ▲ Improved | 76% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 107% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 34% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 10% | 0% |
| case-18 | ✗→✓ | ▲ Improved | -35% | 0% |
> 核心信条：能简单地教会别人，才算真懂。 讲的时候被简化或跳过的地方，正是"我以为我懂了"的幻觉核心点。
用户学完一个东西想验真伪，或隐约觉得"好像懂了但不踏实"。也是"重输入轻输出"的一次强制输出。
请他用自己的话、把你当外行，把概念讲一遍。别让他背定义——要他解释、打比方。
专挑他含糊带过、用术语糊弄、跳过的环节追问："为什么？""那这个是怎么来的？""举个例子？"命中他答不上来或开始绕的地方。
明确指出哪几处他没真懂（不是责备，是定位）。这些就是幻觉核心点。
每个漏洞是：① 缺前置知识（→转 learn-graph 定位 / learn-crossover 看是否其实已会）还是 ② 有料但没想透（→当场再讲一轮，直到讲顺）？
能顺畅讲通、追问不倒 = 闭环。否则明确指出还差哪一环。
> ⚠️ 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 learn-crossover learn-graph learn-prototype。| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-08 | fail→pass | 7,724 | 11,213 | +45% | 1 | 1 | 0% | 1,341 | 2,355 | +76% | 0 | 0 | — |
case-09 | fail→pass | 5,457 | 7,815 | +43% | 1 | 1 | 0% | 847 | 1,753 | +107% | 0 | 0 | — |
case-01 | pass→pass | 13,340 | 16,933 | +27% | 1 | 1 | 0% | 2,376 | 3,349 | +41% | 0 | 0 | — |
case-02 | fail→fail | 1,609 | 4,114 | +156% | 1 | 1 | 0% | 231 | 1,152 | +399% | 0 | 0 | — |
case-15 | fail→pass | 10,304 | 11,154 | +8% | 1 | 1 | 0% | 1,610 | 2,153 | +34% | 0 | 0 | — |
case-03 | fail→fail | 2,390 | 12,066 | +405% | 1 | 1 | 0% | 378 | 2,402 | +535% | 0 | 0 | — |
case-04 | fail→pass | 8,703 | 5,657 | -35% | 1 | 1 | 0% | 1,304 | 1,434 | +10% | 0 | 0 | — |
case-05 | pass→pass | 7,265 | 7,035 | -3% | 1 | 1 | 0% | 1,181 | 1,526 | +29% | 0 | 0 | — |
case-06 | fail→fail | 3,966 | 6,028 | +52% | 1 | 1 | 0% | 643 | 1,511 | +135% | 0 | 0 | — |
case-07 | pass→pass | 3,780 | 6,555 | +73% | 1 | 1 | 0% | 619 | 1,475 | +138% | 0 | 0 | — |
case-10 | fail→fail | 6,396 | 4,664 | -27% | 1 | 1 | 0% | 1,056 | 1,185 | +12% | 0 | 0 | — |
case-11 | pass→pass | 6,519 | 3,710 | -43% | 1 | 1 | 0% | 1,015 | 1,199 | +18% | 0 | 0 | — |
case-12 | pass→pass | 8,511 | 8,204 | -4% | 1 | 1 | 0% | 1,487 | 1,888 | +27% | 0 | 0 | — |
case-13 | pass→pass | 5,529 | 6,454 | +17% | 1 | 1 | 0% | 781 | 1,350 | +73% | 0 | 0 | — |
case-14 | pass→pass | 8,404 | 13,271 | +58% | 1 | 1 | 0% | 1,374 | 2,345 | +71% | 0 | 0 | — |
case-16 | pass→pass | 9,113 | 6,400 | -30% | 1 | 1 | 0% | 1,491 | 1,461 | -2% | 0 | 0 | — |
case-17 | pass→pass | 5,725 | 5,048 | -12% | 1 | 1 | 0% | 903 | 1,304 | +44% | 0 | 0 | — |
case-18 | fail→pass | 12,706 | 5,223 | -59% | 1 | 1 | 0% | 2,028 | 1,328 | -35% | 0 | 0 | — |
case-19 | fail→pass | 4,958 | 5,733 | +16% | 1 | 1 | 0% | 755 | 1,473 | +95% | 0 | 0 | — |
case-20 | pass→pass | 9,503 | 9,196 | -3% | 1 | 1 | 0% | 1,559 | 1,900 | +22% | 0 | 0 | — |
case-21 | pass→pass | 7,519 | 6,474 | -14% | 1 | 1 | 0% | 1,261 | 1,524 | +21% | 0 | 0 | — |
case-22 | fail→pass | 13,723 | 6,954 | -49% | 1 | 1 | 0% | 2,068 | 1,508 | -27% | 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 +32 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.