More screenshots
Works with the AI Agents you already use
LLaVA: Run Vision-Language Models That See Images
Run LLaVA to caption images, answer questions about them, and hold multi-turn image dialogues — with workarounds for its classic failures.
$5
LLaVA: Run Vision-Language Models That See Images
Example session with this skill installed
I have product photos and I want LLaVA to describe them and answer questions like 'is there a scratch on the left edge?' Show me the setup and how to keep it from inventing things.
- Read your context and instructions
- Compiled the llava
Setup: loaded LLaVA with 4-bit quantization on a single consumer GPU, which fits the model and leaves room for a batch of images; verified the load with a known photo and a caption request before touching the real set. Prompting: used the grounded pattern — 'Describe only what is visible in this image' followed by the question — and asked for the scratch question as a yes/no with a location, not an open essay. On the test photo the model correctly said no scratch, and on a deliberately scratched sample it said yes and located it at the left edge. Counting: the first count of screws in a photo came back wrong (7 instead of 5), so the workaround kicked in — ask for each item to be located one at a time, then count the locations, instead of asking for a number directly. The recount came back 5, matching the hand count. Tiny text: the model misread a serial label, so the fix was to crop the label region and ask again at full resolution rather than trusting the downscaled read. Dialogue: a three-turn conversation about one photo kept the image in context throughout, with each follow-up referring back to the same image instead of re-uploading. Evaluation: ran 20 photos with known answers — 17 correct, 2 honest 'cannot tell', 1 invented detail — and the invented one was a reflection the model read as an object. The rule going forward: any claim about a defect gets a crop-and-recheck before it is reported.
Connects securely to your tools. The creator never sees your data.
About this skill
A vision-language model that hallucinates a nonexistent object is worse than one that says nothing, so this skill is built around the failure modes, not just the setup. It covers loading and quantizing LLaVA, the prompting pattern that keeps answers grounded in what is actually visible, multi-turn image dialogues that keep the image in context, serving patterns for single images and batches, and honest evaluation against a small labeled set. The workarounds are the point: bad counting, tiny text, and confident invention each get a specific counter-move, and the skill is explicit about when a vision-language model is the wrong tool entirely.
LLaVA counts badly, misreads tiny text, and invents details with confidence — the skill gives workarounds, not a cure.
It does not cover training or fine-tuning a vision-language model, only running and prompting an existing one.
High-resolution images cost memory and time; the skill shows when to crop instead of feeding the whole frame.
How to use
- 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
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
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