Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →SOP: Deep-read the original north-star context, distill this ARA's overall direction, and align it with the user via the reused present-and-ask / present-candidates dialogue SOPs
.claude/skills/yogsoth-ai-north-star-align/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 54% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -5% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -46% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -56% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -29% | 0% |
Key question: 这篇 ARA 的大方向是什么?和用户对齐了吗?
context-exploring 定位的那篇)。提炼:这条研究弧最初要回答的问题、它的范围边界、它自认的核心贡献。
约束。这段会进 compiler 的 $ARGUMENTS,用来约束 PAPER.md 的 title/abstract, 防止逆向抽取出的 ARA 跑题。
Skill load present-candidates —— 把 context-exploring 聚出的 arc 候选作排序候选呈现给用户(默认整目录,但给手挑口子)。
Skill load present-and-ask —— 把"这篇 ARA 用哪些 context / 大方向是否如此理解"摆给用户确认,收回用户的取舍。
### 大方向,把用户手挑的范围写进### arc 范围。
完成的投喂计划(大方向已填、arc 范围已定)+ 对齐过的大方向,交给 compile-and-review tactic。
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 19,879 | 25,178 | +27% | 1 | 1 | 0% | 2,231 | 3,433 | +54% | 0 | 0 | — |
case-02 | fail→pass | 23,723 | 33,502 | +41% | 1 | 1 | 0% | 2,663 | 2,530 | -5% | 0 | 0 | — |
case-03 | fail→fail | 18,111 | 20,995 | +16% | 1 | 1 | 0% | 2,512 | 2,428 | -3% | 0 | 0 | — |
case-04 | pass→fail | 51,602 | 28,908 | -44% | 1 | 1 | 0% | 8,229 | 530 | -94% | 0 | 0 | — |
case-05 | pass→pass | 21,344 | 26,238 | +23% | 1 | 1 | 0% | 2,906 | 2,546 | -12% | 0 | 0 | — |
case-06 | pass→pass | 49,058 | 47,503 | -3% | 1 | 1 | 0% | 8,217 | 8,541 | +4% | 0 | 0 | — |
case-07 | fail→fail | 32,930 | 18,511 | -44% | 1 | 1 | 0% | 2,896 | 2,433 | -16% | 0 | 0 | — |
case-08 | fail→pass | 18,364 | 9,430 | -49% | 1 | 1 | 0% | 2,051 | 1,115 | -46% | 0 | 0 | — |
case-09 | fail→fail | 12,874 | 7,343 | -43% | 1 | 1 | 0% | 1,294 | 701 | -46% | 0 | 0 | — |
case-10 | fail→fail | 8,445 | 7,247 | -14% | 1 | 1 | 0% | 464 | 691 | +49% | 0 | 0 | — |
case-11 | fail→pass | 16,253 | 7,559 | -53% | 1 | 1 | 0% | 1,663 | 735 | -56% | 0 | 0 | — |
case-12 | fail→pass | 11,698 | 7,499 | -36% | 1 | 1 | 0% | 1,024 | 730 | -29% | 0 | 0 | — |
case-13 | fail→pass | 8,626 | 2,077 | -76% | 1 | 1 | 0% | 1,325 | 595 | -55% | 0 | 0 | — |
case-14 | fail→pass | 15,609 | 8,034 | -49% | 1 | 1 | 0% | 1,380 | 775 | -44% | 0 | 0 | — |
case-15 | fail→pass | 16,802 | 2,321 | -86% | 1 | 1 | 0% | 1,715 | 703 | -59% | 0 | 0 | — |
case-16 | fail→fail | 12,050 | 2,295 | -81% | 1 | 1 | 0% | 1,054 | 735 | -30% | 0 | 0 | — |
case-17 | fail→pass | 15,827 | 2,332 | -85% | 1 | 1 | 0% | 2,673 | 663 | -75% | 0 | 0 | — |
case-18 | fail→pass | 8,008 | 3,170 | -60% | 1 | 1 | 0% | 1,191 | 843 | -29% | 0 | 0 | — |
case-23 | fail→pass | 11,998 | 3,349 | -72% | 1 | 1 | 0% | 1,657 | 1,036 | -37% | 0 | 0 | — |
case-19 | fail→pass | 14,505 | 7,035 | -51% | 1 | 1 | 0% | 2,152 | 639 | -70% | 0 | 0 | — |
case-20 | fail→pass | 8,213 | 9,861 | +20% | 1 | 1 | 0% | 1,245 | 1,117 | -10% | 0 | 0 | — |
case-21 | fail→pass | 30,679 | 11,347 | -63% | 1 | 1 | 0% | 1,922 | 2,036 | +6% | 0 | 0 | — |
case-22 | fail→pass | 17,643 | 9,277 | -47% | 1 | 1 | 0% | 2,694 | 1,852 | -31% | 0 | 0 | — |
case-24 | pass→pass | 7,590 | 2,252 | -70% | 1 | 1 | 0% | 1,155 | 687 | -41% | 0 | 0 | — |
case-25 | fail→pass | 5,834 | 2,869 | -51% | 1 | 1 | 0% | 849 | 738 | -13% | 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. 25 cases were attempted, and 24 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +60 percentage points is the difference between those two pass rates over the 24 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.