Recent AI news · Original report: 2026-09-23 (KST)

New AI agent skills on GitHub: model routing, photo geolocation, promo video (Sep 23, 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: 2026-09-23 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 repositories with high star counts among those created on September 18–19, Korea Standard Time (KST), and summarized what their READMEs say they do and where their 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 23, 2026. The commit column gives the date of the latest commit in KST.

hermes-jev-skills: delegate small decisions to a cheaper model

The README opens with “Give your agent a fast, cheap second brain for the small decisions.” The idea is to hand choices, such as which model should handle a turn or which skill to load, to TypeSafe AI's decision model Jev, leaving the writing to the more expensive model. It describes Jev as responding only with choices, scores or yes/no answers, rather than writing text.

• The README reports roughly 0.4 seconds to route a model for one turn and about 2.8 seconds to choose among 377 installed skills. These are the author's measurements.
• For session handoffs, the author says that passing the full conversation preserved more recall than a summary selected by Jev. As a result, Jev is not used by default for this task.
• Because Jev is a cloud API, the README has a separate section on what leaves the machine. For routing, it says up to 2,500 characters of the message are sent after personal information is redacted; conversation history, tool results and files are not sent.
• The project states that Jev and TypeSafe are TypeSafe AI products and that this project is independent.

geo-sleuth: find where a photo was taken

The README describes it as “An agent skill that finds where a photo was taken — and shows its work.” It says it narrows down locations using OpenStreetMap geometry, elevation data, satellite tiles and street views, even without text, license plates or landmarks. The same folder reportedly works with six agents, including Claude Code and Codex.

• One example is shown step by step. Starting with a photo of farmland and an elevated railway with its location metadata removed, it narrows 27,335 candidate railway bridge segments to 171 locations, then 14,372 camera positions to 22, then 3, then 1, with a stated error of ±2 m. The process reportedly took about 72 minutes.
• Twenty scripts handle searching, scoring and sorting; the model judges only a handful of top candidates.
• It requires Python 3.10 or later and uv, and uses the MIT license. The README ends by asking users to use only their own photos or photos they have permission to investigate, and not to locate people who do not want to be found.

guizang-product-video-skill: make update videos from code

The Chinese README opens with a sentence meaning that a software update should come with a promotional video. The author says the skill packages a workflow refined while making promotional videos for CodePilot. Given a code repository and a description of the changes, the AI reportedly uses actual product components to produce an MP4 and an editable project.

• Defaults are a 45–60-second video, landscape format and Chinese. The README says music is synthesized in code and sound effects come from 11 bundled WAV files. It requires Node.js 22 or later, Python 3.9 or later, FFmpeg and Playwright Chromium.
• License caveat: The main project is AGPL-3.0, but the README says CodePilot-derived components in the assets/fallback folder remain under Business Source License 1.1. It also states that generated videos do not become AGPL-licensed merely because the tool was used to create them.

cli-faq-shortcuts: turn long daily requests into short commands

The README opens with “Stop typing the same long request to your coding agent every day.” This skill reads the user's Claude Code and Codex conversation history, identifies recurring requests and turns each into a short command such as /sent.

• In the example project, an identical sentence appeared at most four times. Grouping requests with the same meaning despite different wording raised the highest count to 127.
• The author says the extractor reads only local files and makes no network calls. Because prompts may include names, email addresses or tokens, the resulting list is written to a temporary folder. The license is Apache-2.0.

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. All measurements above are values reported by the authors in their READMEs.
• Why Jev is spreading: In our September 19 article, we found six new repositories using Jev. This time, GitHub search returned 2,133 repositories created since September 16 containing typesafe or jev. We could not establish whether an event or sponsorship was involved.
• The number of hermes-jev-skills skills: The README said nine, but its skills folder contained ten at the time of the check. We could not determine whether the count excluded a configuration skill.

Sources

• GitHub, kerpopule/hermes-jev-skills https://github.com/kerpopule/hermes-jev-skills
• GitHub, Oldcircle/geo-sleuth https://github.com/Oldcircle/geo-sleuth
• GitHub, op7418/guizang-product-video-skill https://github.com/op7418/guizang-product-video-skill
• guizang-product-video-skill LICENSE (AGPL-3.0) https://github.com/op7418/guizang-product-video-skill/blob/main/LICENSE
• guizang-product-video-skill assets/fallback LICENSE.CodePilot (BSL 1.1) https://github.com/op7418/guizang-product-video-skill/blob/main/assets/fallback/LICENSE.CodePilot
• GitHub, kishormorol/cli-faq-shortcuts https://github.com/kishormorol/cli-faq-shortcuts