Recent AI news · Original report: 2026-09-29 (KST)

New AI agent skills on GitHub: logo design, Blender 3D, travel planner, charts (Sep 29, 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: 2026-09-29 10:11 (KST, UTC+09:00).

Translated into English: 2026-09-29 (KST). The translation date does not indicate a new source check.

At a glance

We picked four AI agent skill repositories created between September 23 and 26, KST, and summarized what their READMEs say they do, along with their stated limitations. What stood out this time was skills that measure and filter their own output before showing it to a person. We did not install or run them.

Repository overview

These figures were retrieved from the GitHub API at 10:11 a.m. on September 29, 2026, KST. The commit column shows the date of the latest commit in KST.

logo-design-skill: logo work from brief to SVG files

The README describes a skill that turns an agent into an identity designer, working "from the first brief to production-ready SVG files and brand guidelines."

• It includes a reference library of more than 1,400 real-world logos, classified by mark type, technique and industry. The README says the library is for studying construction and avoiding look-alikes, "never to copy."
• Concepts are tested at 16 pixels, in one colour and on reversed backgrounds. After showing the concepts, the skill stops until a person picks a direction.
• The README shows 18 fictional brands run end to end. The tools use only the standard library of Python 3.8 or later.
• Note: the README states separately that the reference logo files are trademarks of their owners and are not covered by the MIT license.

blender-game-skills: artwork into game-ready 3D models

The README describes a skill that takes concept art, photos or sketches and builds game-ready 3D models in Blender, such as characters, vehicles, buildings and props. The repository currently holds one skill, blender-image-to-3d.

• The work is split into phases 0 to 10, from the brief through rigging to export. Each phase compares renders taken from the same angle as the reference, using numbers. The first blockout phase passes at a silhouette overlap of 0.85 or more and proportions within 5 percent.
• The README says judging the viewport by eye is not accepted. Exported files are re-imported into an empty scene and checked that the height is within 1 percent.
• It needs Blender 4.2 or newer and no GPU. The author says Claude Code is the only agent it has been tested with.
• Among its limits, the README says results depend on the model, the quality of the references and a person's review at each phase, and that a single image cannot show the back or sides, so those parts are marked as inferred.

jianhao-travel-planner: a travel plan that cross-checks sources

The Chinese-language README describes a workflow that takes "I want to go on a trip," gathers requirements, looks up real information online, and produces an itinerary and food information cards.

• It says key facts come with sources, unverified items are marked, and fares carry the time they were checked. It cross-checks three Chinese services: Dianping, Amap (Gaode) and Xiaohongshu.
• It mechanically checks that every day includes all three meals. The author added this check after finding that "in a 5-day itinerary, not a single day had all three meals."
• The README says no paid API or platform logins are needed. For photos, it uses a free Pexels API key, or AI-generated images labelled as illustrations instead.

evident-charts: checking charts before you see them

The README says it teaches AI coding agents to make "clear, honest charts" and checks each one before the user sees it.

• It first checks the data for totals mixed in with their parts and for duplicate rows. The title states what the data shows rather than the topic, and the source line names the real publisher instead of a file name.
• For matplotlib charts, code checks for overlapping text, bars that do not start at zero and colours that are hard to tell apart for colour-blind readers. A separate reviewer that sees only the finished image then fixes issues, up to three rounds.
• Each rule is tagged by what it rests on: experimental evidence, practitioner consensus or taste. A project style guide can override taste rules but not integrity rules such as bars starting at zero.

What we could not confirm

• Actual behavior and security: we did not install or run any of the four repositories, and we did not review their code for security. All thresholds and counts are figures the authors wrote in their READMEs.
• blender-game-skills pass thresholds: these are the author's own criteria. We could not confirm whether a model that passes them is good enough for a real game.
• jianhao-travel-planner accuracy: it says it cross-checks Chinese services, but we did not check whether the itineraries it produces are correct. The README also does not say how well it works for destinations in Korea.
• Excluded repositories: several video skills, including onetake with 842 stars, were left out because their licenses were unclear. golive-skill, 3dicon and zh-tech-writing were covered in earlier posts.

Sources

• GitHub, kaankiziltug/logo-design-skill https://github.com/kaankiziltug/logo-design-skill
• logo-design-skill trademark notice (TRADEMARKS.md) https://github.com/kaankiziltug/logo-design-skill/blob/main/TRADEMARKS.md
• GitHub, majidmanzarpour/blender-game-skills https://github.com/majidmanzarpour/blender-game-skills
• GitHub, awangwang123/jianhao-travel-planner https://github.com/awangwang123/jianhao-travel-planner
• GitHub, rhiever/evident-charts https://github.com/rhiever/evident-charts
• GitHub, feitangyuan/onetake (why it was excluded) https://github.com/feitangyuan/onetake