Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Feast — open-source feature store. Online and offline serving, point-in-time joins, feature validation, and streaming ingestion. Standardizes feature management across training and production.
.claude/skills/mkurman-feast/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-22 | ✗→✓ | ▲ Improved | -59% | 0% |
| case-02 | ✓→✓ | ＝ Same ✓ | 18% | 0% |
| case-03 | ✓→✓ | ＝ Same ✓ | 17% | 0% |
| case-04 | ✓→✓ | ＝ Same ✓ | -13% | 0% |
| case-05 | ✓→✓ | ＝ Same ✓ | -5% | 0% |
Feast is an open-source feature store for production ML, providing offline (batch training data via SQL queries) and online (low-latency serving via Redis, Firestore, or DynamoDB) feature retrieval with point-in-time correctness. Features are versioned, validated, and governed through a registry.
bashuv pip install feast
pythonfrom feast import Entity, FeatureView, FileSource, ValueType from datetime import timedelta driver = Entity(name="driver_id", value_type=ValueType.INT64, description="Driver identifier") source = FileSource(path="data/driver_stats.parquet", timestamp_field="event_timestamp") feature_view = FeatureView( name="driver_hourly_stats", entities=[driver], ttl=timedelta(hours=2), source=source, )
bashfeast apply # register in registry feast serve # online serving at localhost:6566
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 7,790 | 6,502 | -17% | 1 | 1 | 0% | 1,486 | 1,460 | -2% | 0 | 0 | — |
case-02 | pass→pass | 2,569 | 1,370 | -47% | 1 | 1 | 0% | 333 | 394 | +18% | 0 | 0 | — |
case-03 | pass→pass | 4,074 | 3,215 | -21% | 1 | 1 | 0% | 735 | 858 | +17% | 0 | 0 | — |
case-04 | pass→pass | 3,972 | 2,112 | -47% | 1 | 1 | 0% | 656 | 571 | -13% | 0 | 0 | — |
case-05 | pass→pass | 7,248 | 5,618 | -22% | 1 | 1 | 0% | 1,421 | 1,347 | -5% | 0 | 0 | — |
case-06 | pass→pass | 3,282 | 1,283 | -61% | 1 | 1 | 0% | 488 | 455 | -7% | 0 | 0 | — |
case-07 | pass→pass | 3,018 | 1,548 | -49% | 1 | 1 | 0% | 480 | 431 | -10% | 0 | 0 | — |
case-08 | pass→pass | 5,069 | 1,565 | -69% | 1 | 1 | 0% | 876 | 485 | -45% | 0 | 0 | — |
case-09 | pass→pass | 7,639 | 2,401 | -69% | 1 | 1 | 0% | 1,444 | 664 | -54% | 0 | 0 | — |
case-10 | pass→pass | 3,664 | 1,851 | -49% | 1 | 1 | 0% | 617 | 580 | -6% | 0 | 0 | — |
case-11 | pass→pass | 4,267 | 2,716 | -36% | 1 | 1 | 0% | 739 | 697 | -6% | 0 | 0 | — |
case-12 | pass→pass | 8,855 | 5,899 | -33% | 1 | 1 | 0% | 1,526 | 1,347 | -12% | 0 | 0 | — |
case-13 | pass→pass | 2,947 | 1,958 | -34% | 1 | 1 | 0% | 481 | 565 | +17% | 0 | 0 | — |
case-14 | pass→pass | 3,128 | 1,784 | -43% | 1 | 1 | 0% | 526 | 526 | 0% | 0 | 0 | — |
case-15 | pass→pass | 6,438 | 2,761 | -57% | 1 | 1 | 0% | 1,094 | 674 | -38% | 0 | 0 | — |
case-16 | pass→pass | 2,380 | 1,701 | -29% | 1 | 1 | 0% | 378 | 502 | +33% | 0 | 0 | — |
case-17 | pass→pass | 4,772 | 1,983 | -58% | 1 | 1 | 0% | 833 | 544 | -35% | 0 | 0 | — |
case-18 | pass→pass | 2,852 | 1,787 | -37% | 1 | 1 | 0% | 449 | 559 | +24% | 0 | 0 | — |
case-19 | fail→fail | 9,074 | 7,155 | -21% | 1 | 1 | 0% | 1,785 | 1,618 | -9% | 0 | 0 | — |
case-20 | fail→fail | 8,747 | 6,717 | -23% | 1 | 1 | 0% | 1,651 | 1,434 | -13% | 0 | 0 | — |
case-21 | fail→fail | 15,595 | 12,936 | -17% | 1 | 1 | 0% | 3,018 | 2,823 | -6% | 0 | 0 | — |
case-22 | fail→pass | 7,713 | 1,911 | -75% | 1 | 1 | 0% | 1,467 | 602 | -59% | 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 0 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is 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.