sprezzature-vision
Vision — alt text
World Wide Web Consortium (W3C)-compliant alt-text drafting via a local Ollama vision model: a per-image decision tree, output in any detected language, deterministic on-disk cache. The one model is qwen3-vl:8b. Nothing leaves the machine.
Work in progress: every output is a draft requiring human review, not finished alt text.
Makes
- Alt-text drafts for informative / decorative / functional / text / complex / group images
- Output in any detected language, native phrasing for common ones
- Surrounding-text and project-vocabulary biasing
- A deterministic on-disk cache: same image never hits the model twice
Audits
- Flags images with no alt
- Drafts are starting points; verify before committing
Say one of these
The skill triggers itself on phrasing like this, no command to memorize.
alt text
describe this image
draft alt
image description
img has no alt
decorative image
figure / chart description
batch alt text
Reference library
The knowledge the skill draws on: read them straight from the repo.