Abilities
A single preconfigured vision task that returns structured output
An Ability is a preconfigured AI task — detect objects, classify a scene, extract fields from a document, understand an event in video — that takes media and returns structured output. Abilities let you add visual intelligence without training or deploying your own models.
Browse ready-made ones, with example output, in the Abilities Hub. Create your own with a prompt in the dashboard, or with eyepop create ability from the CLI section in this site's navigation. Publishing an ability gives it an alias under your namespace, <your-namespace>.<task>.<name>:latest, and that alias is how the CLI, a Pop, and a deployment reference it.
Chain abilities together into a pipeline with a Pop.
Anatomy of an Ability
An Ability is a small set of parameters that control how media is analyzed, how much compute it uses, and how often frames are sampled.
name
Unique identifier used in API calls, e.g. <your-namespace>.structured-ocr.read-drivers-license.
description
What the Ability does. Also what the Prompt Creation Agent reads to generate or refine prompts, so make it specific and task-oriented.
config.image_size
Resolution media is resized to before inference. The single biggest lever on cost and speed.
text_prompt
The instruction given to the vision-language model (--prompt on the CLI). Define the task and restrict the possible outputs.
worker_release
The backing vision-language model, currently the shared default qwen3-instruct. Not selectable when you create an Ability.
transform_into.classes
The fixed set of output labels raw model output is mapped into (--class on the CLI).
config.fps
Frames per second sampled from video and livestreams.
image_size and fps are part of the Ability's config object; the other fields sit at the top level.
A useful mental model: Ability = model + prompt + media sampling. The model sets reasoning capability, the prompt defines the task, and fps plus image_size control performance and compute.
A good prompt
Determine whether a person in the frame is wearing a safety helmet.
Return exactly one label from: ["helmet", "no_helmet"].It names the task and constrains the output. Avoid open-ended prompts like Analyze the image — they produce inconsistent labels, verbose text, and higher cost.
Typical settings
Use case
image_size
fps
Object detection
512–640
—
Find event in video
512–640
2–5
Sports analytics
512–640
5–10
Industrial monitoring
512–640
1–3
Document analysis
768–1024
—
Lower image_size and fps reduce compute while still capturing most events.
Cost
Abilities consume Compute Units based on the model, the image_size you set, and — for video — the frame rate. See Pricing for plans and rates.
Ability Prompt Creation Agent
Vision models are sensitive to how instructions are written; small wording changes shift accuracy, consistency, and cost. The Prompt Creation Agent takes the task you want to perform and generates a prompt that follows tested patterns, avoiding inconsistent classifications, verbose output, hallucinations, and unpredictable formatting.
Coming soon to the dashboard.
Next steps
Models — the pretrained catalog you can run today, and how custom models differ
Pops — chain abilities into a pipeline
Train Your Own Model — build an ability from your own data
Last updated