Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →OpenAI CLIP — contrastive language-image pre-training. Zero-shot image classification, image-text similarity, concept search, and cross-modal retrieval. Embed images and text into shared space.
.claude/skills/mkurman-clip/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | -48% | 0% |
| case-01 | ✗→✓ | ▲ Improved | -36% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -29% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -5% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -30% | 0% |
OpenAI CLIP (Contrastive Language-Image Pre-training) learns joint text-image representations. Enables zero-shot image classification, image-text similarity, cross-modal search, and image captioning without task-specific training.
bashuv pip install openai-clip
pythonimport clip import torch model, preprocess = clip.load("ViT-B/32") image = preprocess(load_image("photo.jpg")).unsqueeze(0) text = clip.tokenize(["a dog", "a cat", "a bird"]) with torch.no_grad(): logits, _ = model(image, text) probs = logits.softmax(dim=-1) print(f"Predicted: class {probs.argmax().item()} with {probs.max():.2%} confidence")
pythonimages = torch.stack([preprocess(img) for img in [load_image("a.jpg"), load_image("b.jpg")]]) texts = clip.tokenize(["sunset", "ocean", "mountain"]) with torch.no_grad(): similarity = model(images, texts)[0].softmax(dim=-1)
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-05 | fail→pass | 11,585 | 4,545 | -61% | 1 | 1 | 0% | 2,146 | 1,118 | -48% | 0 | 0 | — |
case-10 | pass→pass | 5,463 | 4,799 | -12% | 1 | 1 | 0% | 884 | 710 | -20% | 0 | 0 | — |
case-01 | fail→pass | 10,819 | 5,044 | -53% | 1 | 1 | 0% | 2,175 | 1,392 | -36% | 0 | 0 | — |
case-02 | fail→fail | 11,157 | 9,336 | -16% | 1 | 1 | 0% | 2,237 | 2,161 | -3% | 0 | 0 | — |
case-03 | fail→pass | 9,302 | 5,069 | -46% | 1 | 1 | 0% | 1,811 | 1,278 | -29% | 0 | 0 | — |
case-04 | fail→pass | 5,530 | 3,777 | -32% | 1 | 1 | 0% | 955 | 903 | -5% | 0 | 0 | — |
case-06 | pass→pass | 8,151 | 3,242 | -60% | 1 | 1 | 0% | 1,466 | 756 | -48% | 0 | 0 | — |
case-07 | pass→pass | 8,701 | 3,843 | -56% | 1 | 1 | 0% | 1,673 | 926 | -45% | 0 | 0 | — |
case-08 | fail→fail | 10,413 | 5,876 | -44% | 1 | 1 | 0% | 1,862 | 1,301 | -30% | 0 | 0 | — |
case-09 | fail→pass | 11,460 | 5,568 | -51% | 1 | 1 | 0% | 1,994 | 1,399 | -30% | 0 | 0 | — |
case-11 | pass→pass | 5,649 | 3,103 | -45% | 1 | 1 | 0% | 1,025 | 850 | -17% | 0 | 0 | — |
case-12 | fail→fail | 11,900 | 7,634 | -36% | 1 | 1 | 0% | 2,206 | 1,869 | -15% | 0 | 0 | — |
case-13 | fail→pass | 8,754 | 4,346 | -50% | 1 | 1 | 0% | 1,611 | 1,056 | -34% | 0 | 0 | — |
case-14 | pass→pass | 7,551 | 2,549 | -66% | 1 | 1 | 0% | 1,234 | 663 | -46% | 0 | 0 | — |
case-15 | pass→pass | 8,070 | 7,391 | -8% | 1 | 1 | 0% | 1,465 | 1,730 | +18% | 0 | 0 | — |
case-16 | pass→pass | 10,485 | 4,261 | -59% | 1 | 1 | 0% | 1,899 | 1,090 | -43% | 0 | 0 | — |
case-17 | pass→pass | 9,304 | 4,469 | -52% | 1 | 1 | 0% | 1,722 | 1,200 | -30% | 0 | 0 | — |
case-18 | fail→fail | 18,468 | 9,065 | -51% | 1 | 1 | 0% | 3,624 | 2,091 | -42% | 0 | 0 | — |
case-19 | fail→pass | 5,974 | 3,965 | -34% | 1 | 1 | 0% | 1,005 | 1,050 | +4% | 0 | 0 | — |
case-20 | fail→pass | 4,702 | 3,934 | -16% | 1 | 1 | 0% | 925 | 1,034 | +12% | 0 | 0 | — |
case-21 | pass→pass | 7,701 | 2,060 | -73% | 1 | 1 | 0% | 1,398 | 644 | -54% | 0 | 0 | — |
case-22 | pass→pass | 18,827 | 18,751 | -0% | 1 | 1 | 0% | 3,766 | 4,089 | +9% | 0 | 0 | — |
case-23 | pass→pass | 11,030 | 6,729 | -39% | 1 | 1 | 0% | 2,163 | 1,596 | -26% | 0 | 0 | — |
case-24 | pass→pass | 15,087 | 11,769 | -22% | 1 | 1 | 0% | 3,027 | 2,769 | -9% | 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. 24 cases were attempted. The headline lift of +33 percentage points is the difference between those two pass rates over the 24 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.