Codex Skill Watchlist — Media Generation and UI Quality

The July 29 Codex skill-ranking screenshot is useful as a discovery signal, not as a permanent ranking. Skill directories move quickly, names drift, and some entries are wrappers around larger repos. Use this page as a router: what Q already has, what is actionable now, and what needs source review before installation.

Sources checked: local skill-folder checks in ~/.codex/skills, ~/.claude/skills, and ~/.agents/skills on 2026-08-08; Skills directory; inference-sh/skills; ai-video-generation upstream folder; ai-image-generation upstream folder; inference.sh CLI install notes.

Bottom Line

The practical stack is:

fuzzy idea
  -> grill-me
  -> media asset generation: ai-video-generation / ai-image-generation
  -> frontend design pass: Taste / design-guide candidate
  -> final UI critique: anti-ui-slop candidate / design review
  -> prose cleanup: stop-slop / avoid-ai-writing

ai-video-generation

This is the strongest item in the screenshot because it is already installed locally and has a real upstream source. The local SKILL.md describes an inference.sh/Belt CLI workflow for more than forty video models, including Veo, Seedance, HappyHorse, Wan, Grok Imagine Video, OmniHuman, Fabric, and HunyuanVideo Foley.

The skill documentation describes these capability groups:

Use it when you want to generate media, not when you want to understand a video. For video analysis, transcript extraction, or cutting existing footage, video-use is the more relevant workflow.

Practical caveats:

• It depends on the belt CLI and inference.sh account/auth.
• Model availability, price, and quality can change quickly.
• Generated media still needs human review for accuracy, rights, and visual fit.
• Treat model lists as documentation claims unless you have run the exact app ID.

coll_ai-image-generation

The screenshot label is coll_ai-image-generation, but the verified upstream package I found is ai-image-generation inside inference-sh/skills. That upstream skill is real and currently names 50+ image models through the same Belt CLI, including GPT-Image-2, FLUX, Gemini image models, Grok Imagine, Seedream, Reve, image editing, inpainting, LoRA, upscaling, and text rendering.

That does not prove coll_ai-image-generation is the same skill. The exact coll_ name was not installed locally, so the correct status is candidate, not recommendation.

When it becomes worth installing:

• You want one CLI route across multiple image models.
• You need image generation as input to video generation.
• Existing posts such as ImageLens or Awesome GPT Image prompt library are not enough because you need execution, not prompt references.

Before installing, verify the exact owner/repo and inspect SKILL.md. Do not install from a screenshot label alone.

design-guide

design-guide is interesting because the pain is real: agents can build a working frontend and still ship generic layouts. But Q already has stronger local coverage for this area:

• Taste Skill for anti-generic frontend rules.
• Agents with Taste for encoding design judgment into skill files.
• ui.sh for AI-native interface tooling.
• Pretext & Refero for frontend-design context and feedback loops.

The exact design-guide skill was not installed locally. I also found a local web-design-guidelines skill under Agents and an upstream inference.sh landing-page-design guide, but those are not the same thing. Keep the names separate.

Use the design-guide idea as a checklist before installing anything:

1. Does it read the project’s existing design system instead of inventing a new one?
2. Does it produce implementation guidance, or only generic design advice?
3. Does it support local/private code without uploading the app?
4. Does it enforce responsive checks, visual states, typography, spacing, and empty/error states?
5. Does it reduce duplicate prompts compared with Taste?

If the answer is yes, it may deserve a later standalone page. For now, the current wiki should point students to Taste first.

anti-ui-slop

This should not be confused with writing cleanup. Q already has Killing AI Slop and Editing AI-Sounding Writing for prose. anti-ui-slop is a frontend-quality idea: catch default AI UI patterns before they ship.

The exact skill is not installed locally, so treat it as a candidate. The evaluation rubric is still useful:

For LearnAI web apps, run this after the main design pass, not before. First pick the audience and design language; then use anti-slop checks to reject generic output.

grill-me

grill-me is already covered well in the wiki. Do not duplicate it just because it appears in a ranking screenshot.

Use it before choosing a media or UI skill:

Use grill-me to stress-test this LearnAI course asset plan.
Ask one question at a time.
Recommend an answer for each question so I can approve or correct it.

That matters because media generation and frontend polishing can burn time fast. grill-me should force the asset to justify itself before you generate it.

Default Student-Friendly Workflow

For student projects and LearnAI course assets, keep the workflow small:

The wiki should not chase every trending skill as a standalone page. Add a page only when the skill either changes a workflow Q actually uses, or teaches students a reusable idea they can apply outside the tool.