Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →DoWhy (Microsoft) — causal inference library. Causal graph modeling, identification (back-door, front-door, IV), estimation (matching, IPW, double-ML), and refutation/robustness checks for causal claims.
.claude/skills/mkurman-dowhy/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -60% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 2% | 0% |
| case-03 | ✓→✓ | ＝ Same ✓ | -22% | 0% |
| case-04 | ✓→✓ | ＝ Same ✓ | -19% | 0% |
| case-05 | ✓→✓ | ＝ Same ✓ | -26% | 0% |
DoWhy (Microsoft/py-why) provides end-to-end causal inference: causal graph modeling (DAG), identification strategies (back-door, front-door, instrumental variables), estimation (linear regression, matching, IV, double-ML), and refutation tests (placebo, bootstrap, random common cause, data subset).
bashuv pip install dowhy
pythonfrom dowhy import CausalModel model = CausalModel( data=df, treatment="treatment", outcome="outcome", common_causes=["age", "gender", "income"], ) # 1. Identify identified = model.identify_effect(proceed_when_unidentifiable=True) # 2. Estimate estimate = model.estimate_effect(identified, method_name="backdoor.linear_regression") print(f"ATE: {estimate.value:.4f} (p={estimate.p_value:.4f})") # 3. Refute refute = model.refute_estimate(identified, estimate, method_name="placebo_treatment_refuter") print(f"Refutation passed: {refute.refutation_result}")
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 16,372 | 13,756 | -16% | 1 | 1 | 0% | 3,338 | 3,258 | -2% | 0 | 0 | — |
case-02 | fail→pass | 11,002 | 3,509 | -68% | 1 | 1 | 0% | 1,970 | 786 | -60% | 0 | 0 | — |
case-03 | pass→pass | 5,937 | 2,772 | -53% | 1 | 1 | 0% | 959 | 748 | -22% | 0 | 0 | — |
case-04 | pass→pass | 6,705 | 3,827 | -43% | 1 | 1 | 0% | 914 | 741 | -19% | 0 | 0 | — |
case-05 | pass→pass | 8,681 | 4,447 | -49% | 1 | 1 | 0% | 1,218 | 897 | -26% | 0 | 0 | — |
case-06 | pass→pass | 3,857 | 2,412 | -37% | 1 | 1 | 0% | 664 | 688 | +4% | 0 | 0 | — |
case-07 | pass→pass | 4,033 | 5,379 | +33% | 1 | 1 | 0% | 499 | 823 | +65% | 0 | 0 | — |
case-08 | pass→pass | 5,015 | 2,531 | -50% | 1 | 1 | 0% | 801 | 733 | -8% | 0 | 0 | — |
case-09 | pass→pass | 5,779 | 2,101 | -64% | 1 | 1 | 0% | 927 | 625 | -33% | 0 | 0 | — |
case-10 | pass→pass | 5,658 | 2,082 | -63% | 1 | 1 | 0% | 872 | 609 | -30% | 0 | 0 | — |
case-11 | pass→pass | 12,813 | 10,129 | -21% | 1 | 1 | 0% | 2,157 | 2,068 | -4% | 0 | 0 | — |
case-12 | pass→pass | 12,046 | 10,479 | -13% | 1 | 1 | 0% | 2,333 | 2,341 | +0% | 0 | 0 | — |
case-13 | pass→pass | 9,001 | 7,110 | -21% | 1 | 1 | 0% | 1,661 | 1,682 | +1% | 0 | 0 | — |
case-14 | pass→pass | 9,863 | 3,706 | -62% | 1 | 1 | 0% | 1,642 | 967 | -41% | 0 | 0 | — |
case-15 | fail→pass | 13,802 | 11,500 | -17% | 1 | 1 | 0% | 2,236 | 2,289 | +2% | 0 | 0 | — |
case-16 | fail→fail | 12,569 | 7,986 | -36% | 1 | 1 | 0% | 2,205 | 1,623 | -26% | 0 | 0 | — |
case-17 | pass→pass | 7,113 | 3,141 | -56% | 1 | 1 | 0% | 1,100 | 818 | -26% | 0 | 0 | — |
case-18 | pass→pass | 6,087 | 3,266 | -46% | 1 | 1 | 0% | 1,011 | 854 | -16% | 0 | 0 | — |
case-19 | pass→pass | 2,558 | 2,283 | -11% | 1 | 1 | 0% | 372 | 634 | +70% | 0 | 0 | — |
case-20 | pass→pass | 4,295 | 3,556 | -17% | 1 | 1 | 0% | 721 | 941 | +31% | 0 | 0 | — |
case-21 | pass→pass | 2,445 | 3,180 | +30% | 1 | 1 | 0% | 401 | 780 | +95% | 0 | 0 | — |
case-22 | pass→pass | 17,980 | 12,021 | -33% | 1 | 1 | 0% | 3,570 | 2,809 | -21% | 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 +9 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.