Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Weights & Biases — ML experiment tracking and visualization. Log metrics, hyperparameters, model checkpoints, and artifacts. Collaborative dashboards, sweep hyperparameter search, and model registry.
.claude/skills/mkurman-wandb/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -31% | 0% |
| case-03 | ✓→✓ | ＝ Same ✓ | 20% | 0% |
| case-04 | ✓→✓ | ＝ Same ✓ | -13% | 0% |
| case-05 | ✓→✓ | ＝ Same ✓ | 34% | 0% |
| case-06 | ✓→✓ | ＝ Same ✓ | 49% | 0% |
Weights & Biases (wandb) tracks ML experiments with rich visualizations, hyperparameter sweeps, dataset versioning, model registry, and collaborative dashboards. Industry standard for experiment tracking across ML teams.
bashuv pip install wandb wandb login # authenticate with API key
pythonimport wandb wandb.init(project="my_project", config={ "learning_rate": 0.001, "batch_size": 32, "architecture": "transformer", }) for epoch in range(10): loss = train_one_epoch() wandb.log({"train_loss": loss, "val_loss": val_loss, "epoch": epoch}) wandb.finish()
pythonsweep_config = { "method": "bayes", "metric": {"name": "val_loss", "goal": "minimize"}, "parameters": {"lr": {"min": 1e-5, "max": 1e-2}}, } sweep_id = wandb.sweep(sweep_config, project="my_project") wandb.agent(sweep_id, function=train_function, count=20)
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 9,217 | 9,020 | -2% | 1 | 1 | 0% | 1,817 | 2,052 | +13% | 0 | 0 | — |
case-02 | fail→pass | 5,958 | 2,514 | -58% | 1 | 1 | 0% | 1,055 | 732 | -31% | 0 | 0 | — |
case-03 | pass→pass | 3,631 | 2,098 | -42% | 1 | 1 | 0% | 581 | 698 | +20% | 0 | 0 | — |
case-04 | pass→pass | 8,282 | 5,874 | -29% | 1 | 1 | 0% | 1,565 | 1,355 | -13% | 0 | 0 | — |
case-05 | pass→pass | 3,320 | 2,898 | -13% | 1 | 1 | 0% | 565 | 759 | +34% | 0 | 0 | — |
case-06 | pass→pass | 3,715 | 3,880 | +4% | 1 | 1 | 0% | 594 | 886 | +49% | 0 | 0 | — |
case-07 | pass→pass | 8,900 | 7,390 | -17% | 1 | 1 | 0% | 1,737 | 1,701 | -2% | 0 | 0 | — |
case-08 | pass→pass | 3,448 | 2,002 | -42% | 1 | 1 | 0% | 563 | 618 | +10% | 0 | 0 | — |
case-09 | pass→pass | 5,786 | 5,524 | -5% | 1 | 1 | 0% | 1,052 | 1,257 | +19% | 0 | 0 | — |
case-10 | pass→pass | 13,125 | 11,695 | -11% | 1 | 1 | 0% | 2,475 | 2,452 | -1% | 0 | 0 | — |
case-11 | pass→pass | 19,704 | 5,471 | -72% | 1 | 1 | 0% | 1,842 | 1,291 | -30% | 0 | 0 | — |
case-12 | pass→pass | 6,265 | 3,619 | -42% | 1 | 1 | 0% | 1,143 | 979 | -14% | 0 | 0 | — |
case-13 | pass→pass | 5,633 | 2,391 | -58% | 1 | 1 | 0% | 949 | 750 | -21% | 0 | 0 | — |
case-14 | pass→pass | 7,746 | 5,898 | -24% | 1 | 1 | 0% | 1,508 | 1,398 | -7% | 0 | 0 | — |
case-15 | pass→pass | 2,921 | 2,132 | -27% | 1 | 1 | 0% | 487 | 682 | +40% | 0 | 0 | — |
case-16 | pass→pass | 5,535 | 3,474 | -37% | 1 | 1 | 0% | 951 | 940 | -1% | 0 | 0 | — |
case-17 | pass→pass | 3,125 | 1,886 | -40% | 1 | 1 | 0% | 461 | 606 | +31% | 0 | 0 | — |
case-18 | pass→pass | 2,722 | 1,841 | -32% | 1 | 1 | 0% | 410 | 576 | +40% | 0 | 0 | — |
case-19 | pass→pass | 3,071 | 2,492 | -19% | 1 | 1 | 0% | 397 | 644 | +62% | 0 | 0 | — |
case-20 | pass→pass | 7,796 | 5,896 | -24% | 1 | 1 | 0% | 1,503 | 1,477 | -2% | 0 | 0 | — |
case-21 | pass→pass | 7,846 | 5,235 | -33% | 1 | 1 | 0% | 1,480 | 1,197 | -19% | 0 | 0 | — |
case-22 | pass→pass | 2,679 | 2,603 | -3% | 1 | 1 | 0% | 489 | 752 | +54% | 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 -100 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.