Recent AI news · Original report: 2026-09-17 (KST)

New AI agent skills on GitHub: video shot analysis, 3D scenes, code blueprints (Sep 17, 2026)

This article was compiled with AI. This English translation preserves the original report's collection date, findings, and limitations. Please consult the linked sources for current information. Originally collected: 2026-09-17 10:11 (KST, UTC+09:00).

Translated into English: 2026-09-28 (KST). The translation date does not indicate a new source check.

At a glance

We selected four relatively highly starred AI skill repositories created this week, September 10–12 KST, and summarized what their READMEs say they do, along with their limitations. We did not install or run them.

Repository overview

These figures were retrieved through the GitHub API on September 17 at 10:11 a.m. KST. Commit dates are the latest commit dates in Korean Standard Time. “Noncommercial” below means a license restricted to personal, noncommercial use.

reelbench-skills: breaking a finished video into shots

The Chinese README introduces two video skills for Claude Code and Codex. video-shots breaks a finished video into an analysis table describing each shot’s duration, shot size, camera movement, and rhythm. video-sync combines the video with shot information displayed alongside it to produce a new video.

• The README says ffmpeg measures cut points and durations, leaving only items that require judgment to the model. It lists 15 quality checks.
• It says Node.js 18 or later and ffmpeg are the only requirements, with no separate API key needed.
• The repository includes an example analysis of a 202.9-second AI short film, divided into 53 shots averaging 3.83 seconds each.
• The opening of the README advertises a paid AI video community run by the author.

3dviz-pro-max: a 3D scene from a sentence

The README describes it as “A standalone agent skill for creative 3D visualization.” It provides a ten-step workflow, a recipe collection, and Three.js templates. The workflow has the agent capture and inspect screenshots of the scene it creates.

• The README says that, as of September 10, it contained 223 recipes, 440 knowledge records, and 37 examples.
• It reports verifying plugin checks and loading in Claude Code on macOS on September 9, 2026, but says several other installation routes, including installation in the Codex app, had not been tested.
• It also says its comparison with and without the skill ran only once per condition, so it does not establish that the skill produces better scenes.

birdview: a diagram before code changes

The README’s summary is “See AI changes before they happen.” Before an AI modifies code, it creates a diagram of the project, highlights the parts it is about to touch, and then makes the changes. The output is a single HTML file that opens without a server or internet connection.

• The README says the default automatic mode requires the agent to check the diagram before each code-changing task. Mode-setting commands manage their own section in the project’s AGENTS.md or CLAUDE.md.
• It says the activity log shown on screen is self-reported by the agent, and the system provides guidance rather than enforcing a block on changes.
• The example image at the top of the README uses fictional data, according to the document.

vivid-figures-skill: research charts from data

The Chinese README describes a scientific charting skill that takes data, has the AI choose a chart type, draw and inspect it, and deliver both the image and the plotting code. It targets papers, mathematical modeling, and experiment reports, and provides seven color palettes.

• It says Python 3.10 or later is required, and ordinary charts do not require an image-generation API key.
• License restriction: The LICENSE file permits personal, noncommercial use only. It states that modifications, derivative development, redistribution, and commercial use, including internal company work, require the copyright holder’s prior written permission.
• It says results depend on how well the model follows instructions, and example images do not guarantee the same quality on every run.

What we could not verify

• Actual operation and security: We did not install, run, or conduct a code security review of any of the four repositories.
• How the stars were accumulated: The star counts in the table are cumulative totals at the time of the query. The repositories were less than a week old, and we do not know how they attracted those stars.
• The template count in vivid-figures-skill: The README lists 143 recipes and three combination templates, totaling 146 explanatory cards. The repository description still lists 108 recipes, apparently the count from its initial release. We did not count them ourselves to establish which reflects the current state.
• The birdview version: The README badge says 0.2.0, while the feature list in the body is written for v0.1.
• The 3dviz-pro-max counts: The recipe and knowledge-record figures are the README’s counts as of September 10. We did not count them ourselves.

Sources

• GitHub, eternityspring/reelbench-skills https://github.com/eternityspring/reelbench-skills
• GitHub, viettranx/3dviz-pro-max https://github.com/viettranx/3dviz-pro-max
• GitHub, Qiuner/birdview https://github.com/Qiuner/birdview
• GitHub, yjz211/vivid-figures-skill https://github.com/yjz211/vivid-figures-skill
• vivid-figures-skill LICENSE https://github.com/yjz211/vivid-figures-skill/blob/main/LICENSE