Recent AI news · Original report: 2026-09-19 (KST)

New AI agent skills on GitHub: video edit drafts, code gates, skill library (Sep 19, 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: 2026-09-19 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 AI agent skill and tool repositories with high star counts among those created on September 14–18, Korea Standard Time (KST), and summarized what each README says it does and where its limits lie. We did not install or run them.

Repositories at a glance

These figures came from the GitHub API at 10:11 a.m. KST on September 19, 2026. The commit column gives the date of the latest commit in KST. “Noncommercial” is shorthand for a license limited to personal learning and noncommercial use.

jianying-headless: turn an AI editing plan into an editor draft

The Chinese README describes a local automation tool for the macOS version of Jianying Pro, the video editor. It creates a Jianying draft from an editing plan file so a person can continue editing. It modifies existing drafts only through copies and exports an MP4 only when requested.

• It states that it is not an official Jianying developer tool and is not an editor that works without Jianying.
• In one example, it says it converted an approximately 50-second Instagram tutorial into 23 tracks and 154 segments, with text remaining editable in 109 of those segments.
• The README says the result is editable rather than a visually identical conversion. Fonts and cropping differ, and it specifically notes a difference between the two images at the 37-second mark.
• It requires an Apple silicon Mac, macOS 26.0 or later and a separately installed copy of Jianying 11.5.0, with compatibility for 11.4.2. Support for 11.5.0 and the example above were added to the README in a commit three minutes before the check. The author says installation validation on other machines is not yet complete.
• License caveat: The LICENSE file permits only personal learning and noncommercial use. It says commercial use, including client deliverables and company work, as well as public redistribution, requires the copyright holder's prior written permission.

gap-trap: put checks around AI-written code

The README summarizes the project as “Turns vibe coding into high quality code.” It says running the tool once in a repository makes it read the code, write rules tailored to the project and attach checks that fail a commit or CI run when a rule is violated.

• It introduces four things: contracts defining permitted and prohibited approaches for each area; checks that fail when a new test also passes against the old code; a mechanism that allows the count of known problems only to decrease; and a record of lessons the project has learned.
• The README says it runs only with the harness's top coding model and stops on smaller models.
• It separately emphasizes that a person should carefully review the generated rules and contracts.

skillbox: a self-hosted library for skills

The README describes it as “A self-hosted, versioned skills library for AI agents.” It stores versioned skills in a database on the user's own server and lets agents find and load them through MCP.

• It says each change leaves a history that can be rolled back, and access keys can be issued and revoked separately.
• It states that the server does not execute uploaded skill code. Installation requires Docker.
• Recommending skills for a task is optional. When enabled, the README says the task description and skill descriptions are sent to an external AI provider, with charges billed to the user's account. Recommendation scores use a 0–4 scale and are not probabilities.
• The top and bottom of the README contain lengthy material about the author's other products and sponsors.

EvoOntology: definitions for data agents

This is an implementation of a paper by researchers at Renmin University of China. The README says it gives agents that work with tables, files and databases a semantic layer through MCP tools. That layer contains metric definitions, relationships between fields and business constraints, and is revised using execution records.

• A revision is reportedly released as a new version only if it outperforms the previous version in a comparison under the same conditions.
• Compared with having no semantic layer, the README's performance table reports DDR-Bench improving from 69.5 to 89.5, BIRD from 63.6 to 72.4 and InsightBench from 53.2 to 54.2. These figures combine results across four models and come from the authors' paper.
• We opened the cited paper, arXiv 2609.15779, and confirmed that its title and four authors matched the README.

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.
• EvoOntology's installation address: The README's installation command points to a different repository, MeiduoChong/EvoOntology, rather than ruc-datalab. That address returned 404, or not found, at the time of the check.
• How the stars accumulated: Star counts are cumulative totals at the time of the check. jianying-headless had 1,033 stars and 504 forks, a high proportion of forks. We could not establish why.
• Background to skillbox's recommendation feature: Among the 237 results in this search, six new repositories, including skillbox, mentioned the same recommendation model, Jev, in their descriptions. All six were created on September 17–18 KST. We could not establish whether an event or sponsorship was involved.

Sources

• GitHub, mcncarl/jianying-headless https://github.com/mcncarl/jianying-headless
• jianying-headless LICENSE https://github.com/mcncarl/jianying-headless/blob/main/LICENSE
• GitHub, pliablepixels/gap-trap https://github.com/pliablepixels/gap-trap
• GitHub, kitze/skillbox https://github.com/kitze/skillbox
• GitHub, ruc-datalab/EvoOntology https://github.com/ruc-datalab/EvoOntology
• arXiv 2609.15779, EvoOntology: A Self-Evolving Ontology Layer for Data Agents https://arxiv.org/abs/2609.15779