More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    CLIP: Zero-Shot Image Classification and Search

    1

    Label images and search them by description with no training data: write captions, not labels, calibrate on images you eyeballed, and ship a shortlist with a threshold — never a forced verdict.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    CLIP: Zero-Shot Image Classification and Search

    CLIP: Zero-Shot Image Classification and Search

    Example session with this skill installed

    We have 2,000 construction site photos and no labeled examples. Flag the ones showing workers without hard hats for a human safety review.

    • Read your context and instructions
    • Compiled the clip

    LABELS: Three descriptive labels, each with three templates averaged: "a photo of a worker wearing a hard hat", "a photo of a worker without a hard hat", "a photo of construction equipment with no people" — plus an explicit "a photo of something else entirely" escape option.

    CALIBRATION: 40 images eyeballed by hand. The model confuses "without a hard hat" with "wearing a hard hat" on distant shots where the head is small — recorded as a known failure mode. On images matching none of the labels (a delivery truck, a lunch break), the model still picks a label at 40–55% probability — which is why the escape option and the threshold exist. Threshold set at 65%: below it, the image goes to "unsure, needs human review."

    SEARCH: The collection embedded once with ViT-B/32 (2,000 images, fast). The query "a photo of a worker without a hard hat" ranks the collection; the top 50 go to a human reviewer, re-ranked with ViT-L/14 for the final ordering.

    RESULT: Of 2,000 photos, 63 score above the threshold as "without a hard hat" and go to review; 41 are confirmed true positives, 22 are distant-shot confusions. The 18% of images below the threshold go to "unsure" rather than being forced into a wrong label. The human reviews 113 images instead of 2,000, and every flagged image carries its probability and its label — the review is a check, not a rubber stamp.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    CLIP learned to connect images and text from hundreds of millions of image-caption pairs, so you can classify images and search them by description without training anything — but the quality of your text determines the quality of your results. This skill is about writing text that makes CLIP see clearly: full descriptive phrases instead of single words ("a photo of a worker wearing a hard hat" beats "safety"), multiple templates per class averaged together, visual specifics instead of abstract concepts, and disambiguation where a word has two meanings.

    The workflow embeds a collection once and searches by typing, ranks by cosine similarity over normalized embeddings, and uses an ANN index past a few thousand images. Calibration is mandatory: run the label set over a few dozen eyeballed images, record the confusions, include an explicit "none of these" option (CLIP always picks a label otherwise), and set the unsure threshold from your calibration set, not a round number. Model selection is explicit — B/32 as the workhorse, B/16 for detail, L/14 for re-ranking a shortlist. A worked example flags hard-hat violations across 2,000 site photos, cutting human review from 2,000 images to 113.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    • Fine-grained distinctions the model never saw captioned (model years, spec-sheet differences) are out of reach.
    • Counting, precise spatial relationships, and small text are weak points.
    • Images matching none of the labels still get a label without the escape option and threshold.
    • Zero-shot probabilities are not calibrated confidences until you calibrate them on your own images.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 3 days ago

    Permissions

    No special permissions declared or detected

    Listed3 days ago

    What's inside

    Frequently Asked Questions

    CLIP trained on captions, not class names. "A photo of a damaged package" matches what it saw during training; "damaged" does not. The phrasing gap is measurable — descriptive phrases plus averaged templates are the cheapest accuracy win in the whole workflow.

    Check three things: is the label a caption or a single word, does the image match none of your labels (no escape option), and is the probability below your calibrated threshold? Most wrong labels are one of those three, and all three are fixable without retraining anything.

    Embed and filter with B/32, then re-rank the shortlist with L/14. Running the biggest model over the whole collection is a schedule problem, not a quality win.

    New in LLM Ops