This week: The security ecosystem went into overdrive after Hugging Face's July AI intrusion incident โ three new scanner/sandbox projects surged into the top ranks. On the coding side, "less is more" philosophy dominated with ponytail's senior-dev minimalism taking the crown. And Google DeepMind quietly dropped a scientific research skill that's already changing how labs approach reproducibility.
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | ponytail | Dietrich Gebert | โญ 11.1k | SKILL.md | Senior-dev minimalism as installable behavior |
| ๐ฅ | improve | shadcn | โญ 4.6k | SKILL.md | Use your best model to audit, cheaper models to execute |
| ๐ฅ | Waza | tw93 | โญ 5.7k | SKILL.md | Familiar engineering habits packaged as skills |
Why ponytail took #1: Dietrich Gebert's ponytail encodes a philosophy that's resonated across X and Reddit this week โ "favor less code, fewer moving parts, fewer unnecessary changes." In an ecosystem where 144k people starred Karpathy's anti-overengineering guidelines, ponytail goes further: it's a complete behavioral style that makes Claude Code ship smaller, cleaner diffs by default. The repo jumped from 8k to 11.1k stars in the last month.
Why improve at #2: shadcn's improve flips the standard coding model on its head. Instead of using your most capable (and expensive) model for everything, it uses that model to audit your codebase and write detailed implementation plans โ then dispatches cheaper models to execute them. At current Claude Opus pricing, that's the difference between $0.15 and $0.015 per task.
Why Waza at #3: tw93's Waza collection packages familiar engineering habits (conventional commits, clean diffs, review discipline) into skills that feel natural rather than constraining. 5.7k stars and growing steadily as more teams standardize on it for onboarding new AI-assisted developers.
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | academic-research-skills | Imbad0202 | โญ 30.2k | SKILL.md | End-to-end academic research pipeline |
| ๐ฅ | claude-scientific-skills | K-Dense AI | โญ 28k | SKILL.md | Ready-to-use scientific research toolkit |
| ๐ฅ | science-skills | Google DeepMind | โญ 2k | SKILL.md | Reproducible scientific experiment workflows |
Why academic-research-skills at #1: The researchโwriteโreviewโreviseโfinalize pipeline encoded here has become the de facto standard for AI-assisted academic writing. 30.2k stars with a spike of 3k new stars in the past two weeks as the fall academic season kicks in. It's not just a prompt โ it's a structured methodology that mirrors how real research labs operate.
Why claude-scientific-skills at #2: K-Dense AI's collection is broader than academic-research-skills, covering engineering, analysis, finance, and writing alongside pure science. 28k stars and the most forked research skill collection on GitHub.
Why science-skills at #3: Google DeepMind's entry is newer (2k stars) but carries the weight of DeepMind's reputation. The focus on reproducible technical work โ experiment tracking, analysis pipelines, verified results โ fills a gap most research skills ignore. This is the one to watch for next month.
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | SkillWard | Fangcun-AI | โญ New | SKILL.md + Docker | Escalation pipeline: scan โ semantic review โ sandbox execution |
| ๐ฅ | Cisco AI Defense Skill Scanner | Cisco | โญ New | SKILL.md | Multi-engine static analysis with policy gates |
| ๐ฅ | Claude-BugHunter | elementalsouls | โญ 2.5k | SKILL.md | 71 offensive security skills + attack matrices |
Why SkillWard at #1: The Hugging Face security incident changed everything. SkillWard is the first skill-security tool that doesn't just scan โ it escalates. Stage A runs YARA and regex checks. Stage B asks a model to classify intent. Stage C launches the skill in a Docker sandbox with decoys to expose malicious behavior. It's the closest thing we have to "dynamic analysis for agent skills," and it arrived at exactly the right moment.
Why Cisco Skill Scanner at #2: Cisco's entry brings AppSec-grade tooling to skills: YAML signatures, YARA rules, Python bytecode integrity checks, shell-pipeline taint analysis, and optional VirusTotal integration. The --policy strict --fail-on-severity high CI gate makes it immediately usable in production pipelines.
Why Claude-BugHunter at #3: 71 curated offensive security skills, slash commands, report patterns, and attack matrices โ elementalsouls built what might be the most comprehensive red-team skill bundle in the ecosystem. 2.5k stars with strong community contributions.
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | Generative-Media-Skills | SamurAIGPT | โญ 3.5k | SKILL.md + scripts | Multi-modal: images, video, audio, music |
| ๐ฅ | video-shotcraft | Vincentwei1021 | โญ 2.2k | SKILL.md + Remotion | Cinematic product videos from prompts |
| ๐ฅ | manim_skill | adithya-s-k | โญ 907 | SKILL.md + Python | 3Blue1Brown-style math animations |
Why Generative-Media-Skills at #1: SamurAIGPT's collection is the Swiss Army knife of media skills โ image generation, video creation, audio production, and music composition in one package. 3.5k stars and the highest "media generation" install count on skills.sh. Works across Claude Code, Cursor, and Gemini CLI.
Why video-shotcraft at #2: Vincentwei1021 solved a real pain point: turning product screenshots into cinematic videos without a video editor. Built on Remotion with shot recipes, motion previews, and production templates. The Remotion community has embraced it, and 2.2k stars in under 3 months is exceptional velocity.
Why manim_skill at #3: If you've ever wanted to create 3Blue1Brown-quality math animations without spending hours in Manim, this is it. 907 stars is modest, but the niche is deep โ educators, researchers, and content creators who need explanatory math visuals. Quality over quantity.
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | drawio-skill | Agents365-ai | โญ 2.4k | SKILL.md + scripts | 10,000+ shapes, 321 AI brand logos, vision self-check |
| ๐ฅ | lottie | diffusionstudio | โญ 2.1k | SKILL.md + scripts | Production-ready Lottie animations from code |
| ๐ฅ | skill.color-expert | meodai | โญ 505 | SKILL.md | Color science: spaces, contrast, palettes, pigment mixing |
Why drawio-skill at #1: Generating diagrams has been one of the weakest areas for AI coding agents โ Mermaid is limited, and hand-coding SVG is painful. drawio-skill changes that: 6 presets, vision-based self-check with up to 5 rounds of refinement, codebase-to-diagram conversion, and access to 10,000+ official shapes plus 321 AI/LLM brand logos. It's the kind of skill that makes you wonder how you worked without it.
Why lottie at #2: diffusionstudio made Lottie animation generation a first-class skill. Instead of describing animations in prose and hoping Claude figures it out, the skill encodes the Lottie format, easing curves, and composition patterns so agents produce animations that actually work in production. 2.1k stars from a design community that rarely stars GitHub repos.
Why skill.color-expert at #3: meodai's color science skill is niche but deep: color spaces, accessibility contrast ratios, palette generation, pigment mixing, and historical color theory. For UI work, it catches contrast issues before they ship. 505 stars, but every one is from a designer who genuinely needed it.
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | humanizer | blader | โญ 23.7k | SKILL.md | Remove AI fingerprints from writing |
| ๐ฅ | codebase-to-course | zarazhangrui | โญ 4.6k | SKILL.md | Turn any codebase into interactive HTML course |
| ๐ฅ | translate-book | deusyu | โญ 794 | SKILL.md + scripts | Translate entire books via parallel subagents |
Why humanizer at #1: blader's humanizer is the anti-AI-slop skill that everyone needs. 23.7k stars โ it removes the telltale signs of AI writing: predictable sentence structures, overused transitions, corporate-speak patterns. In a world where AI-generated content is everywhere, making it sound human is the new differentiator. The companion "napkin" skill (persistent per-repo scratchpad for recording Claude mistakes) is equally clever.
Why codebase-to-course at #2: zarazhangrui built something genuinely innovative โ feed it a codebase, get back an interactive single-page HTML course. Perfect for onboarding, documentation, and teaching. 4.6k stars in under 4 months. This is the kind of skill that creates entirely new workflows rather than just optimizing existing ones.
Why translate-book at #3: deusyu's skill translates entire books (PDF/DOCX/EPUB) into any language using parallel subagents working on different chapters simultaneously. 794 stars is modest but the technical achievement is significant โ parallel translation with consistency guarantees across chapters is a hard problem.
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | skillsbench | benchflow-ai | โญ 1.4k | SKILL.md | Benchmark suite for skill effectiveness |
| ๐ฅ | evals-skills | hamelsmu | โญ 1.4k | SKILL.md | Golden test sets + LLM-as-judge workflows |
| ๐ฅ | guard-skills | amElnagdy | โญ 881 | SKILL.md | Quality gates for AI-generated code |
Why skillsbench at #1: As the skill ecosystem matures, the question "does this skill actually work?" becomes critical. skillsbench is the first serious attempt at benchmarking skill effectiveness โ testing not just the skill code but how well agents use it. 1.4k stars and growing as skill marketplaces start requiring benchmark scores.
Why evals-skills at #2: hamelsmu's evaluation skills encode best practices for testing AI systems: golden test sets, LLM-as-judge evaluation, repeatable eval workflows. 1.4k stars from the ML engineering community that takes evaluation seriously.
Why guard-skills at #3: amElnagdy's quality-gate skills catch common AI-generated failure modes in code, tests, and documentation before they reach production. 881 stars and one of the most practical testing tools โ it's not about measuring quality, it's about enforcing it.
Ranked by AgentSkillReview. Data sourced from GitHub, skills.sh, agenticskills.io, scriptbyai.com, and community forums. Stars and install counts as of September 4, 2026. Only modular SKILL.md skills considered โ no SaaS platforms, no coding IDEs, no research-only frameworks.
