Best agent skills: the top 20 on GitHub this week

The best agent skills right now, ranked by new GitHub stars: anatomy, a skill that draws ideas as working machines, leads week 3 so far, a whisker ahead of ai-job-search and Anthropic's official plugin directory. The chart updates every Thursday. Below it, every skill that has made a weekly top 20 since 1 October gets a plain-English profile: what it does, who it's for, which AI agents it works in and how to install it safely.

Last updated 21 October 2026. Written by Annie, a search and digital marketing specialist. Star counts from gittrend's daily figures and OSS Insight; facts from each project's README and website. See the editorial policy.

Use this chart. You are welcome to quote the figures, embed a screenshot or republish the top 20 in a newsletter, as long as you link to this page: https://aiagents.org.uk/best-agent-skills/. Suggested credit: "Source: aiagents.org.uk weekly agent-skills chart". For the raw weekly figures, email Annie.

On this page

• This week's top 20: 15-21 October (so far)
• What the chart is telling us
• Earlier weeks
• What are agent skills?
• The skills guide: all 28 profiled
• Also tracked
• New launches to watch
• skills.sh and the copycat "superpowers" repos
• How this chart works
• Frequently asked questions

This week's top 20: 15-21 October (so far)

anatomy leads by just 42 stars from ai-job-search, with Anthropic's plugin directory third. Five skills are new to the top 20 this week: ai-job-search, claude-plugins-official, K-Dense's science pack, caveman and graphify. These figures cover 15-17 October, so the order can still change before the week closes on Wednesday.

1. 1anatomy ↑1Explains ideas as working isometric machines. Works in Claude Code, Codex, OpenCode +2.+4,108
2. 2ai-job-search NewRuns a whole job search. Works in Claude Code.+4,066
3. 3claude-plugins-official NewAnthropic's official plugin directory. Works in Claude Code.+3,909
4. 4mattpocock/skills ↓3Engineering workflow skills from Matt Pocock. Works in Claude Code, Codex, Copilot, Gemini CLI.+3,786
5. 5ECC ↑1All-in-one harness kit: skills, memory, security. Works in Claude Code, Codex, Cursor +6.+3,260
6. 6scientific-agent-skills (K-Dense) New177 science and research skills. Works in Claude Code, Codex, Cursor +2.+2,786
7. 7ponytail ↓3Makes the agent write less, better code. Works in Claude Code, Codex, Cursor +8.+2,770*
8. 8hyperframes ↓3Write HTML, render MP4 video. Works in Claude Code, Codex, Cursor, Gemini CLI.+2,275
9. 9superpowers ↑3A full spec-plan-build method with TDD. Works in 16 agents, Claude Code to Qwen.+2,160
10. 10taste-skill ↑4Art direction for frontend agents. Works in Claude Code, Codex, Cursor.+1,873
11. 11diagram-design ↓8Editorial diagrams as HTML + SVG. Works in Claude Code, Codex, Cursor +3.+1,725
12. 12Agent-Reach ↓4Lets agents read X, Reddit, YouTube and more. Works in any agent with a shell.+1,424*
13. 13caveman NewCuts tokens with terse replies and a proxy. Works in Claude Code, Codex, Gemini CLI +30.+1,413
14. 14graphify NewTurns a codebase into a knowledge graph. Works in Claude Code, Codex, Cursor +20.+1,343
15. 15archify ↓8Interactive diagrams from ideas or code. Works in Claude Code, Codex, Cursor +2.+1,336
16. 16addyosmani/agent-skills ↓6Production engineering lifecycle skills. Works in Claude Code, Codex, Cursor +6.+1,195
17. 17impeccable ↓1Design vocabulary and anti-slop checks. Works in Claude Code, Codex, Cursor +6.+1,170
18. 18security-audit-skill (Cloudflare) ↓7Security audit with verified findings. Works in any Agent Skills agent.+1,137
19. 19ppt-master ↓10Native, editable PowerPoint decks. Works in Claude Code, Codex, Cursor +4.+1,083
20. 20ui-ux-pro-max-skill =Searchable UI styles, palettes and fonts. Works in Claude Code, Cursor, Codex +8.+1,060

Stars gained 15-17 October 2026. Arrows show the move from last week; New means outside last week's top 20. * from OSS Insight's star history; other figures from gittrend's daily counts.

What the chart is telling us

• A new number one every week. ponytail led week 1, mattpocock/skills week 2, and anatomy leads week 3 so far. Only Matt Pocock's pack has stayed in the top 4 all month.
• Visual skills rule. 9 of week 2's top 20 make diagrams, video, slides or better-looking pages, from anatomy and diagram-design to HyperFrames and ppt-master.
• Useful-for-life skills are breaking through. ai-job-search jumped from outside the top 20 to #2 this week, and K-Dense's science pack went straight in at #6.
• Anthropic is climbing its own chart. knowledge-work-plugins entered at #18 in week 2, and the official plugin directory is #3 this week.
• Fast risers fade fast. image-blaster went from #4 in week 1 to outside the top 20; universal-modder slid from #7 to #19 and out.
• Check before you install. Copycat "superpowers" repos lead skills.sh's 24-hour list. Our safety guide explains how to spot them.

Earlier weeks

What are agent skills?

Agent skills are folders of instructions, and sometimes scripts, that teach an AI agent how to do a particular job. Each is built around a file called SKILL.md. Because they follow an open format, the same skill works in Claude Code, OpenAI's Codex, Cursor, GitHub Copilot, Gemini CLI and dozens of other agent apps.

Most install with one command, npx skills add owner/repo, or from a store. We review the three biggest: skills.sh, Anthropic's Claude plugin directory and the Cursor Marketplace. A skill runs with your agent's access to your files and keys, so check it first: our guide to whether agent skills are safe has a five-point check.

The skills guide: all 28 profiled

Every repo that has made a weekly top 20 since 1 October, in order of total stars gained from 1 to 17 October. Click any picture to see it full size.

• Helps with: the job it does for you
• Use case: a moment when you'd reach for it
• Works in: the agent apps it supports. Most use the open Agent Skills format, so npx skills add installs them almost anywhere
• Local AI: whether it needs a paid cloud service, or works with any model your agent runs, including a local one in LM Studio or Ollama
• October: stars gained 1-17 October, and the weekly ranks (week 1 · week 2 · week 3 so far; out means outside the top 20). * marks OSS Insight figures

1. mattpocock/skills

The engineer's toolkit: small, sharp skills that make the agent ask before it builds.

• Helps with: keeping AI coding disciplined (planning, tests, triage, refactoring) without handing control to a big framework.
• Use case: before building a feature, run /grill-me and the agent questions your plan until the gaps show; then /tdd builds it test-first.
• Works in: Claude Code (an official plugin), Codex, GitHub Copilot (CLI and VS Code) and Gemini CLI; any Agent Skills agent through npx skills.
• Site: aihero.dev/skills
• Local AI: plain instructions, no API keys; works with whatever model your agent runs.
• October: +23,104 stars · ranks #2 · #1 · #4
• Repo: github.com/mattpocock/skills · 295,639 stars · 24,803 forks · MIT · created 3 Feb 2026

Matt Pocock, best known for teaching TypeScript, built these 38 skills for his own daily work. He pitches them against heavyweight methods such as GSD, BMAD and Spec-Kit, which he says take away your control and make bugs harder to trace. Each skill fixes one agent failure: grilling sessions make the agent question you before it writes code; others tackle verbose agents, broken code and codebases that drift into a "ball of mud".

It is also the most-installed collection on skills.sh: 12 of the all-time top 25 skills are Pocock's, led by grill-me at 1.4 million installs. Version 1.3, in early October, added implement-spec, which works through a spec's tickets with sub-agents in separate worktrees.

Install

claude plugin install mattpocock-skills@claude-plugins-official
npx skills@latest add mattpocock/skills

Then run /setup-matt-pocock-skills once.

2. DietrichGebert/ponytail

A terse senior developer inside your agent: less code, fewer tokens, the same results.

• Helps with: stopping over-engineering. The agent reuses what already exists and writes the one line that works instead of fifty that don't.
• Use case: a long refactor that is burning through tokens; then ponytail-review gives a blunt review of the diff before you merge.
• Works in: about 20 agents, including Claude Code, Codex, Cursor, Windsurf, Kiro, Gemini CLI, OpenCode, Devin and Qoder.
• Site: ponytail.dev
• Local AI: a ruleset plus hooks, no API keys; works with any model.
• October: +18,210 stars* · ranks #1 · #4 · #7
• Repo: github.com/DietrichGebert/ponytail · 166,600 stars · 9,018 forks · MIT · created 12 Jun 2026

ponytail won week 1 on a simple promise: your agent writes too much. The author's own benchmark (39 tasks, five runs each) reports 53% less code, 41% less time, 26% lower cost and 45% fewer tokens, with 98% of risky logic shipping with a test against 68% without it. Treat those as the maker's figures, not an independent test.

Version 5.3 adds "the band", a strip above the prompt in Claude Code that shows each task's cost, tokens, agent time, how often code was reused and whether tests passed. Six skills cover reviews, audits, technical debt and savings, with an intensity dial from lite to ultra.

Install

/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail

3. wheresryan22/anatomy

Name an idea in one line, and Claude draws the machine that does it.

• Helps with: explaining technical ideas visually: not boxes and arrows, but an object whose moving parts behave the way the idea does.
• Use case: explaining caching to a new colleague. A cache key that skips unchanged work becomes a pin-tumbler lock they can turn and press.
• Works in: Claude Code, plus Antigravity, OpenCode, Codex and MiniMax Code.
• Site: skills.wheresryan.sh/anatomy, with 33 live figures
• Local AI: no API keys and nothing to install beyond Node 20.10+; your agent's model draws the figure.
• October: +15,338 stars · ranks out · #2 · #1
• Repo: github.com/wheresryan22/anatomy · 15,596 stars · 1,191 forks · MIT · created 8 Oct 2026

The launch of the month. anatomy did not exist in week 1, added 11,230 stars in its first week and leads week 3 so far. It passed 10,000 stars six days after launch.

The output is a detailed isometric illustration in SVG, made of dozens to hundreds of small parts, with optional WebGL effects for fire, water and light. Every number a figure shows comes from the same model that moves it, and it exports standalone HTML, SVG or React components. The limits are honest: isometric view only, with true 3D still in beta.

Install

npx skills add wheresryan22/anatomy -g

Then ask: /anatomy explain <idea>.

4. cathrynlavery/diagram-design

Editorial diagrams your designer won't hate.

• Helps with: architecture diagrams, flowcharts and data graphics that look designed, delivered as one self-contained HTML file with inline SVG.
• Use case: type "draw our deploy flow" and get a brand-matched diagram for docs, a pitch or a blog post in about a minute, with no Figma and no Mermaid.
• Works in: Claude Code, Codex, GitHub Copilot, Cursor, Factory Droid and Pi.
• Site: diagramdesign.dev
• Local AI: no keys and no build step; the author notes that output quality varies by model.
• October: +14,479 stars · ranks #17 · #3 · #11
• Repo: github.com/cathrynlavery/diagram-design · 57,353 stars · 3,656 forks · MIT · created 16 Apr 2026 by Cathryn Lavery

The biggest climber of week 2, up 14 places to #3. It offers 44 diagram types (the README now says 45), from architecture and flowcharts to Sankey and Wardley maps, each in light, dark and full-editorial styles. Point it at your own website and it matches your fonts and colours; it can also redraw existing Mermaid, draw.io and Excalidraw files.

The author is upfront about the trade-off: there is no diffable source, so Mermaid stays the better choice for documentation that changes every sprint. Recent releases added a WCAG contrast check and an offline system-font option.

Install

npx skills add cathrynlavery/diagram-design

5. affaan-m/ECC

The operating layer for agent harnesses: skills, memory, planning and security in one kit.

• Helps with: giving an agent a complete working setup (300+ skills plus agents, hooks, memory and language rules) instead of assembling one yourself.
• Use case: a team that wants Claude Code, Codex and Cursor to work the same way across every repo, with a shared security scanner (AgentShield).
• Works in: Claude Code, Codex, Cursor, OpenCode, Gemini CLI, Kimi Code, Kiro, Qoder, GitHub Copilot and more.
• Site: ecc.tools
• Local AI: the skills and rules are plain files that work with any model; the GitHub App is an optional hosted extra.
• October: +13,504 stars · ranks #6 · #6 · #5
• Repo: github.com/affaan-m/ECC · 282,953 stars · 42,017 forks · MIT · created 18 Jan 2026

The steadiest performer in the chart: sixth, sixth and fifth, at around 5,000 stars a week. ECC was called "Everything Claude Code" until this month and now pitches itself as tool-neutral. Its site counts 1.2 million skill installs on skills.sh and 7,621 GitHub App installs.

One warning from the README is worth repeating: install only from the official repo, the ecc-universal npm package or the plugin slug ecc@ecc. Re-uploads elsewhere may carry malware.

Install

npx ecc-universal install --guided

6. Panniantong/Agent-Reach

Eyes on the open web: X, Reddit, YouTube, GitHub and China's big platforms, through one command-line tool.

• Helps with: letting an agent read and search social and video platforms without paying for their APIs.
• Use case: "What are developers on X and Reddit saying about this new tool this week?", answered with links, from inside your agent.
• Works in: Claude Code, Cursor, Windsurf, OpenClaw and any agent that can run shell commands.
• Site: none; the README is mainly in Chinese, with an English version in docs/.
• Local AI: a local CLI with no API fees; some platforms read your browser cookies.
• October: +12,688 stars* · ranks #3 · #8 · #12
• Repo: github.com/Panniantong/Agent-Reach · 98,903 stars · 8,710 forks · MIT · created 24 Feb 2026

Agent-Reach covers X, Reddit, YouTube, GitHub, Bilibili and XiaoHongShu. Each platform has a preferred route and fallbacks, so it keeps working when one route breaks. It ships as a skill that teaches the agent the CLI and installs it if missing.

Its momentum is fading: third in week 1, twelfth so far in week 3. Check two things before installing. Some platforms use your logged-in browser cookies, and the quick start asks your agent to follow an install document and run shell commands.

Install

npx skills add Panniantong/Agent-Reach@agent-reach

7. heygen-com/hyperframes

HeyGen's open-source way for agents to make video: write HTML, render MP4.

• Helps with: making video from code (launch clips, changelogs, explainers) where the same input always gives the same frames.
• Use case: turn a merged pull request into a short release video automatically with the pr-to-video skill, ready to post.
• Works in: Claude Code, Codex, Cursor and Gemini CLI; the CLI works from any agent. HyperFrames Studio, a desktop editor for macOS, sits on top.
• Site: hyperframes.dev
• Local AI: renders on your machine with headless Chrome and FFmpeg; no cloud service is needed to render.
• October: +11,483 stars · ranks #9 · #5 · #8
• Repo: github.com/heygen-com/hyperframes · 66,307 stars · 5,989 forks · Apache-2.0 · created 10 Mar 2026

Agents write a video as plain HTML with data attributes, and HyperFrames renders it frame by frame. Because the output is deterministic, it suits CI pipelines and automated rendering. It is pitched as an HTML-first alternative to Remotion, which uses React components.

The install numbers are serious: 7 of the all-time top 25 skills on skills.sh are HyperFrames skills, and its npm package shows 2.7 million downloads a month. There are 21 skills in the main folder and 7 more in its registry, from hyperframes-animation to ready-made product-launch-video and talking-head-recut workflows.

Install

claude plugin marketplace add heygen-com/hyperframes
claude plugin install hyperframes@hyperframes
npx skills add heygen-com/hyperframes

8. tt-a1i/archify

Draw diagrams with your AI agent: interactive, animated and easy to share.

• Helps with: turning an idea, a plan or a whole repository into an interactive HTML diagram you can click through.
• Use case: getting to grips with an unfamiliar codebase. Ask the agent to read it and draw the architecture, then send the HTML file to the team.
• Works in: Claude Code, Codex, Cursor, OpenCode and GitHub Copilot.
• Site: archify.si
• Local AI: no keys; the result is a local HTML file.
• October: +9,919 stars · ranks #8 · #7 · #15
• Repo: github.com/tt-a1i/archify · 85,162 stars · 5,771 forks · MIT · created 15 Apr 2026

archify is diagram-design's natural rival. Where diagram-design makes static, editorial graphics, archify makes animated diagrams you can explore. Version 3.1.0 (14 October) brought ten types: architecture, workflow, sequence, data flow and lifecycle, now joined by entity-relationship, class, tree, timeline and waterfall.

Five skills ship together: archify itself, archify-review, archify-tuning, simplify-codebase and test-value. Its README claims more than 130,000 installs on skills.sh.

Install

npx skills add tt-a1i/archify -g

9. pbakaus/impeccable

The missing design vocabulary for agents.

• Helps with: getting AI-built interfaces past the generic look, with 24 design commands and 61 rules that spot the tells of AI-made frontends.
• Use case: run /impeccable audit on a landing page your agent built, then /impeccable polish or /impeccable quieter to fix what it finds.
• Works in: Claude Code, Cursor, Codex, Gemini CLI, Kiro, Grok, OpenCode, Pi and more.
• Site: impeccable.style
• Local AI: the CLI and browser extension run the 61 detector rules with no LLM and no API key.
• October: +9,532 stars · ranks #5 · #16 · #17
• Repo: github.com/pbakaus/impeccable · 82,530 stars · 4,971 forks · Apache-2.0 · created 16 Nov 2025 by Paul Bakaus

The tells it hunts are familiar to anyone who has seen a lot of AI-made sites: Inter everywhere, purple-to-blue gradients, cards nested in cards and grey text on coloured backgrounds. Start a project with /impeccable init, which records your audience, purpose and voice in a PRODUCT.md file so later commands keep to it.

It began from Anthropic's frontend-design skill and has grown well beyond it. Its weekly rank has slipped since a strong first week, but it remains one of the most-starred design skills.

Install

npx impeccable install

Then run /impeccable init inside your agent.

10. obra/superpowers

The most-starred skills repo on GitHub: a complete software method for coding agents.

• Helps with: long builds that stay on track. The agent draws out a spec, writes a plan, then runs sub-agents task by task with red-green testing and review.
• Use case: hand over a feature in the morning. The agent brainstorms the spec with you, writes the plan, then works through it for hours, tests first.
• Works in: 16 install targets, including Claude Code, Codex (app and CLI), Cursor, Gemini CLI, GitHub Copilot CLI, Antigravity, Factory Droid, Kimi Code, OpenCode and Qwen Code.
• Site: none; everything lives on GitHub.
• Local AI: plain skills with no API keys.
• October: +8,714 stars · ranks #10 · #12 · #9
• Repo: github.com/obra/superpowers · 301,420 stars · 26,913 forks · MIT · by Jesse Vincent

Superpowers is the reference point for the whole category: 15 skills plus a few instructions that make sure the agent actually uses them. It is the only repo in this guide with more than 300,000 stars.

Version 7.0.0 shipped on 9 October. Brainstorming now asks one open question first and sizes the work; a fresh "builder" sub-agent checks the design; there is a native Antigravity plugin; and Windows hook fixes include one that stopped a repo planting a script that ran at session start.

Install (Claude Code)

/plugin install superpowers@claude-plugins-official

11. addyosmani/agent-skills

Production-grade engineering skills from Addy Osmani, from first idea to live deploy.

• Helps with: making the agent follow a senior engineer's workflow (spec, plan, build, test, review, simplify, ship) with quality gates at each step.
• Use case: /build auto plans and implements every task after a single approval, still test-first; /ship then runs the launch checklist.
• Works in: Claude Code, Codex, Cursor, GitHub Copilot, Windsurf, Cline, Antigravity, OpenCode and Kiro; the site says 70+ agents.
• Site: skills.addy.ie
• Local AI: plain skills, no keys.
• October: +8,016 stars · ranks #15 · #10 · #16
• Repo: github.com/addyosmani/agent-skills · 108,008 stars · 11,259 forks · MIT · created 15 Feb 2026

There are 24 lifecycle skills plus a meta-skill that picks the right one. Each is a workflow with steps, verification gates and tables that counter the excuses agents make for cutting corners. Coverage runs from spec- and test-driven development to API design, security hardening, performance, accessibility, observability, CI/CD and migrations.

Nine slash commands map to the lifecycle: /spec, /plan, /build, /test, /constraints, /review, /webperf, /code-simplify and /ship.

Install

npx skills add addyosmani/agent-skills

12. rehan-remade/universal-modder

Point any coding agent at a PC game you own, and mod it.

• Helps with: game modding from start to finish: finding the game, working out its engine, reading the real code, building the mod, testing it in the running game and cutting a showcase video.
• Use case: the README's own demo puts Minecraft's look into GTA V; you can also port an existing mod to another game.
• Works in: Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot and OpenCode; any agent that reads AGENTS.md.
• Site: none yet; a mod hub at universalmodder.com is promised.
• Local AI: needs a fal API key (a paid cloud service) for art, 3D and sound; the key is kept in your operating system's keyring.
• October: +7,535 stars* · ranks #7 · #19 · out
• Repo: github.com/rehan-remade/universal-modder · 8,292 stars · 903 forks · MIT · created 30 Sep 2026

The most playful repo in the chart, and one of the fastest to fade: seventh in week 1, out of the top 20 by week 3. Twelve skills cover the stages, including game-recon, reverse-engineering, mod-any-game, port-a-mod and showcase-video, alongside a fal MCP server and its own um command-line tool.

Each agent writes field notes to a shared knowledge base, so the next agent on the same game starts with what the last one learned.

Install

/plugin marketplace add rehan-remade/universal-modder
/plugin install universal-modder@universal-modder

13. calesthio/OpenMontage

The first open-source agentic video production system: your coding agent becomes a studio.

• Helps with: turning a plain-language brief into a finished video: research, script, assets, editing and the final cut.
• Use case: describe a 60-second explainer. The agent gathers real footage from free stock libraries and open archives, cuts it to a timeline and renders it, rather than animating a few stills.
• Works in: Claude Code, Codex, Cursor, GitHub Copilot and Windsurf; OpenMontage Studio, a hosted workspace, has a waitlist.
• Site: openmontage.video
• Local AI: every API key is optional. Free stock media works with no keys, and local-model support through Ollama and LM Studio is listed as coming soon; paid providers (ElevenLabs, FLUX, Kling and others) unlock more tools.
• October: +7,424 stars · ranks #11 · #13 · out
• Repo: github.com/calesthio/OpenMontage · 69,351 stars · 8,937 forks · AGPL-3.0 · created 29 Mar 2026

The scale is large: 12 production pipelines, more than 100 tools and 90 distinct skills. Check the licence before commercial use, because AGPL-3.0 requires you to share changes if you run a modified version as a service.

Install: clone it, run npm install, then open the folder in your agent.

git clone https://github.com/calesthio/OpenMontage.git

14. cloudflare/security-audit-skill

Cloudflare's security auditor in a single skill, and the seed of its own vulnerability harness.

• Helps with: finding real security bugs in a codebase without drowning in false alarms.
• Use case: audit a repo before a release. Every candidate finding goes to a fresh verifier that tries to disprove it, and you get records marked confirmed, needs validation or rejected.
• Works in: any agent that supports Agent Skills and sub-agents, installed with npx skills.
• Site: Cloudflare's blog post Build your own vulnerability harness
• Local AI: no keys, but it runs several isolated sub-agents through six phases, so allow for a long run.
• October: +7,387 stars · ranks #16 · #11 · #18
• Repo: github.com/cloudflare/security-audit-skill · 30,902 stars · 1,935 forks · MIT · created 18 Jun 2026

The six phases are reconnaissance, coverage-led hunting, candidate validation, structured output, independent verification and a neutral report. Findings land in schema-checked JSON plus a readable REPORT.md, and repeat runs build on earlier ones rather than starting again.

Its growth is word of mouth rather than new releases: the repo has changed little since September, yet it topped GitHub's monthly trending list. Cloudflare says this skill grew into the fleet-wide system described in its blog.

Install

npx skills add https://github.com/cloudflare/security-audit-skill

15. Leonxlnx/taste-skill

The anti-slop frontend framework for AI agents.

• Helps with: art direction, so agent-built pages stop looking like every other AI-made site.
• Use case: building a client website that feels bespoke. The agent first lists what is unique about the client, picks one of 20 researched design directions, then builds around one idea and one interactive element.
• Works in: Claude Code, Codex and Cursor; plus TasteCode, the team's new desktop design agent, which works with Codex, Claude Code and Grok.
• Site: tasteskill.dev
• Local AI: plain skill files with no keys.
• October: +7,227 stars · ranks #18 · #14 · #10
• Repo: github.com/Leonxlnx/taste-skill · 98,644 stars · 6,700 forks · MIT

One of the few repos climbing every week. Version 2.0.0 (13 October) rebuilt the main skill from a study of 247 curated websites. The other skills (14 in all) handle redesigning an existing interface, image-to-code, a stricter variant for GPT and Codex, and minimalist, brutalist and brand-kit styles.

Install

npx skills add Leonxlnx/taste-skill

16. thedotmack/claude-mem

Memory that follows your agent from one session to the next.

• Helps with: stopping the agent forgetting. It records what happens in each session, compresses it into summaries and feeds the relevant parts back later.
• Use case: open Monday's session and the agent already knows Friday's decisions, fixes and dead ends, without you pasting notes back in.
• Works in: Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Antigravity, Pi and more.
• Site: claude-mem.ai, which now also sells CMEM Pro for teams
• Local AI: the open-source engine is free and runs a worker service alongside your agent; the paid Pro tier adds syncing across tools and teammates.
• October: +6,738 stars · ranks #14 · #15 · out
• Repo: github.com/thedotmack/claude-mem · 101,516 stars · 8,905 forks · Apache-2.0 · created 31 Aug 2025

One of the older repos here, and a steady seller rather than a spike. It has 31 skills, including mem-search, make-plan, handoff, standup and learn-codebase, plus search tools exposed over MCP. Version 14.0.1 shipped on 17 October, and it is part of Vercel's 2026 open-source software programme.

Install

npx claude-mem install

17. neilsonnn/image-blaster

One image in, a 3D world out.

• Helps with: jump-starting 3D work. A single picture becomes an explorable 3D environment, meshes and sound effects in under five minutes, according to the README.
• Use case: game-level concepts, film location scouting, architectural renders or robot training environments, starting from one reference photo.
• Works in: Claude Code, run inside the cloned repo.
• Site: none.
• Local AI: needs World Labs and fal API keys, both paid cloud services.
• October: +6,585 stars · ranks #4 · out · out
• Repo: github.com/neilsonnn/image-blaster · 11,467 stars · 1,147 forks · MIT · created 21 Apr 2026

The purest spike of the month: fourth in week 1, then gone from the top 20, with no code changes since May. Eight skills split the job, from image-blast-world and image-blast-3d to image-blast-sfx and image-blast-uncover.

The quick start is refreshingly plain: clone the repo, run Claude Code inside it, give it your two API keys, drop an image into input/ and ask Claude to "blast it and confirm each step with me".

Install

git clone https://github.com/neilsonnn/image-blaster

18. MadsLorentzen/ai-job-search

The job search that runs on your machine.

• Helps with: the grind of job hunting: finding and ranking postings, tailoring CVs and cover letters, preparing for interviews and tracking outcomes.
• Use case: each morning run /scrape and /rank, then /apply for the best matches. A separate reviewer checks every draft, and follow-ups are drafted but never sent without you.
• Works in: Claude Code; the README also mentions Codex, Gemini CLI and Antigravity.
• Site: none.
• Local AI: runs on your own machine inside your agent; CVs are LaTeX files you keep.
• October: +6,421 stars · ranks out · out · #2
• Repo: github.com/MadsLorentzen/ai-job-search · 51,392 stars · 17,072 forks · MIT · created 18 Mar 2026

The breakout of week 3. Weekly gains went 673, then 1,682, then 4,066 in just three days, and it was #2 on GitHub's daily list on 18 October. Its fork count is unusually high because the setup is to fork it and make it yours.

The story sells it. The author, a geophysicist whose job was cut in late 2025, used this exact workflow for 69 tailored applications, 20 first interviews and one contract, and started as an AI engineer in June 2026. Nine skills include job searches for Danish sites and LinkedIn; the six commands are /setup, /scrape, /rank, /apply, /interview and /outcome.

Install: fork and clone the repo, run python3 tools/install_portal_clis.py, then /setup in Claude Code.

19. hugohe3/ppt-master

Real, editable PowerPoint from any document: native shapes, not pictures of slides.

• Helps with: turning reports, papers or a bare topic into a .pptx you can actually edit, with native charts, tables, transitions and animations.
• Use case: feed it a long PDF report and get a deck on your own company template, with data-backed charts and narration generated from the speaker notes.
• Works in: Claude Code, Codex, Cursor, GitHub Copilot, Windsurf, Cline and others.
• Site: example gallery: 70 examples, 987 pages, 70 templates
• Local AI: builds the file on your machine with Python 3.10+.
• October: +6,195 stars · ranks out · #9 · #19
• Repo: github.com/hugohe3/ppt-master · 63,346 stars · 5,011 forks · MIT · created 10 Dec 2025 · v6.7.2 (17 Oct)

Its pitch is "real PowerPoint shapes, not images": every element is a native object, so you can keep editing the deck in PowerPoint. The author, Hugo He, describes himself as a finance professional, and the project has 36 releases behind it.

Install

npx skills add hugohe3/ppt-master

20. Graphify-Labs/graphify

Turn any codebase into a knowledge graph your agent can query instead of grepping.

• Helps with: big or unfamiliar projects. It maps code, docs, PDFs, images and video into one graph, so the agent follows real connections instead of reading file after file.
• Use case: type /graphify . in a large repo, then ask for the path between two modules or an explanation of one concept, without the agent opening every file.
• Works in: Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot and 20+ more, on macOS, Linux and Windows.
• Site: graphify.com, home of Graphify Cloud
• Local AI: code is parsed on your machine with tree-sitter, with no LLM and nothing leaving your computer. Docs and media use your assistant's model, a configured key or a local Ollama model.
• October: +5,600 stars · ranks #20 · out · #14
• Repo: github.com/Graphify-Labs/graphify · 128,029 stars · 12,344 forks · Apache-2.0 and MIT

Graphify is not a vector database: it builds a real graph across about 40 languages, and tags every edge as either read directly from the source or inferred, so you know which links to trust. It is backed by Y Combinator, and its site counts 8.8 million PyPI downloads.

The company also sells Graphify Cloud, an always-on version for monorepos and teams that it says cuts agent token spend by 80%. That is the vendor's figure.

Install

uv tool install graphifyy
graphify install

Then run /graphify . in your agent.

21. multica-ai/andrej-karpathy-skills

Andrej Karpathy's complaints about AI coding, turned into four rules in one file.

• Helps with: curbing the classic agent failures: wrong assumptions, bloated abstractions and edits to code it was not asked to touch.
• Use case: drop the CLAUDE.md file into a project so every session starts from four principles: think before coding, simplicity first, surgical changes and goal-driven execution.
• Works in: Claude Code (as a plugin or a CLAUDE.md file) and Cursor (as a rules file).
• Site: none.
• Local AI: a single text file that works with any model.
• October: +5,359 stars* · ranks out · #17 · out
• Repo: github.com/multica-ai/andrej-karpathy-skills · 220,824 stars · 22,176 forks · no licence listed · formerly forrestchang/andrej-karpathy-skills

One of the smallest repos in this guide, and one of the most starred. It is not written by Karpathy: it turns his widely shared observations about LLM coding (models that run with wrong assumptions, overcomplicate code and change things they don't understand) into instructions. The repo has barely changed in months, yet it keeps gaining thousands of stars a week.

Install

/plugin marketplace add forrestchang/andrej-karpathy-skills
/plugin install andrej-karpathy-skills@karpathy-skills

22. nextlevelbuilder/ui-ux-pro-max-skill

A searchable design brain for agents: styles, palettes, font pairings and UX rules.

• Helps with: choosing a coherent design system before the agent builds any interface.
• Use case: ask for a SaaS landing page and the skill searches its catalogue for a matching style, colour palette and font pairing, then hands the agent a design system to build from.
• Works in: Claude Code, Cursor, Windsurf, Antigravity, GitHub Copilot, Kiro, Codex, Qoder, Roo Code, Trae and OpenCode.
• Site: uupm.cc
• Local AI: the design catalogue is searched on your machine by a Python script, with no keys.
• October: +5,355 stars · ranks out · #20 · #20
• Repo: github.com/nextlevelbuilder/ui-ux-pro-max-skill · 137,044 stars · 14,561 forks · MIT

The site lists 57 UI styles, 95 colour palettes, 56 font pairings, 24 chart types and 29 landing-page patterns across 8 tech stacks; the README now counts 79 searchable styles. Seven skills ship together, including design-system, brand, banner-design, slides and ui-styling, and its command-line installer shows 72,000 downloads a month.

It has sat at exactly #20 two weeks running, a quiet sign of steady demand rather than a launch spike.

Install

npm install -g ui-ux-pro-max-cli
uipro init --ai claude

Swap claude for cursor, windsurf, codex, copilot or another agent.

23. ayghri/i-have-adhd

ADHD-friendly answers: action first, steps numbered, no padding.

• Helps with: replies you can act on at a glance: the answer up front, numbered steps, and no "Great question!" or "Hope this helps!".
• Use case: debugging when you are tired. Instead of three paragraphs about your auth flow, you get "Edit src/auth.ts:42" and two numbered steps.
• Works in: Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Qwen, Kimi and Pi.
• Site: none.
• Local AI: plain instructions that work with any model.
• October: +5,333 stars · ranks #13 · out · out
• Repo: github.com/ayghri/i-have-adhd · 57,605 stars · 3,327 forks · MIT · created 13 May 2026

No diagnosis needed to benefit: it simply stops a coding agent burying the answer. The README's before-and-after table makes the case in seconds, it ships evals that measure the skill's own cost and token overhead, and the README comes in 11 languages.

Install: paste this line into your agent.

Install the i-have-adhd skill/plugin from https://github.com/ayghri/i-have-adhd, refer to the repo's AGENTS.md for instructions.

24. K-Dense-AI/scientific-agent-skills

177 science skills that turn a coding agent into a research assistant.

• Helps with: scientific work in biology, chemistry, medicine, drug discovery, Earth science, physics and data analysis, with the conventions and checks each field expects.
• Use case: ask for compounds similar to a lead molecule; the agent queries PubChem and ChEMBL, runs RDKit and writes up the results with sources.
• Works in: Cursor, Claude Code, Codex, Gemini CLI, GitHub Copilot, Antigravity and Pi; also K-Dense BYOK, a free desktop co-scientist built on the same skills.
• Site: the arXiv paper describing the collection
• Local AI: K-Dense BYOK runs on your desktop with your own keys and a choice of 40+ models.
• October: +5,291 stars · ranks out · out · #6
• Repo: github.com/K-Dense-AI/scientific-agent-skills · 52,556 stars · 4,652 forks · MIT

A mid-October surge took it straight to #6 in week 3. The collection was renamed from Claude Scientific Skills to Scientific Agent Skills to reflect that it now follows the open standard. A Database Lookup skill documents 80 sources, from PubChem, ChEMBL and UniProt to ClinicalTrials.gov and FRED, and more than 70 skills cover Python packages such as RDKit, Scanpy and scikit-learn.

The repo claims more than 250,000 scientists use it. Version 2.72.0 dropped four document skills it had copied from Anthropic and added its own scientific-slides skill.

Install

npx skills add K-Dense-AI/scientific-agent-skills

25. JuliusBrussee/caveman

"Why use many token when few token do trick."

• Helps with: cutting token spend twice over: the agent says less (a terse "caveman" voice) and reads less (a local proxy that trims what goes into the model).
• Use case: long Claude Code sessions on a Pro or Max plan that keep hitting usage limits.
• Works in: Claude Code, Codex, Gemini CLI, OpenCode and 30+ more, plus the Claude app.
• Site: docs.caveman.so
• Local AI: the skill and proxy both run on your machine with your own keys or login; no hosted account is needed.
• October: +5,112 stars · ranks #19 · out · #13
• Repo: github.com/JuliusBrussee/caveman · 113,373 stars · 6,560 forks · Apache-2.0 and MIT

The joke name hides a serious idea. The voice is modelled on Simplified Technical English, the controlled language of aircraft maintenance manuals: answer first, one idea per sentence, code and error messages left untouched. The README reports 33% fewer input tokens across 54 Claude Code sessions and says its style cuts output costs 1.4 to 2.4 times across eight models, citing Adobe Research.

Version 4.0 split it into six parts, each with its own switch, and 22 skills including caveman-commit, caveman-review and caveman-compress. Treat the savings as the maker's own figures.

Install

npx skills add JuliusBrussee/caveman -g

That installs the skill only; npx @caveman-ai/cli@latest adds the proxy too.

26. tigerless-labs/autoharness

A skill library that writes, merges and prunes itself from your own sessions.

• Helps with: capturing what you teach the agent. Each session is distilled into a skill, near-duplicates are folded together, and skills that stop being used are pruned.
• Use case: after you have corrected the agent on your release steps a couple of times, the lesson becomes a reusable skill; /learn distils one on demand.
• Works in: Claude Code only, on Linux and macOS.
• Site: none.
• Local AI: runs entirely in Python 3.11+ with no third-party dependencies, and only touches the skills it wrote itself.
• October: +5,089 stars · ranks #12 · out · out
• Repo: github.com/tigerless-labs/autoharness · 11,136 stars · 629 forks · MIT

The README makes a neat argument: changing the harness alone took one model from 42% to 78% on the CORE-Bench benchmark, yet harnesses are still rebuilt by hand. autoharness bets that one slice of the harness, the skill layer, can maintain itself. Skills are judged by whether they get used in later sessions, not by a benchmark, and every change is logged to a per-skill ledger.

Install

/plugin marketplace add tigerless-labs/autoharness
/plugin install autoharness@autoharness

27. anthropics/claude-plugins-official

Anthropic's official plugin directory, the catalogue behind Claude Code's /plugin menu.

• Helps with: finding and installing plugins, each a bundle of skills, slash commands, MCP servers and hooks.
• Use case: type /plugin, open Discover, and install Anthropic's frontend-design or Pocock's skills pack in one step.
• Works in: Claude Code.
• Site: Claude Code plugin docs
• Local AI: depends on the plugin you install.
• October: +4,874 stars · ranks out · out · #3
• Repo: github.com/anthropics/claude-plugins-official · 42,275 stars · 4,503 forks · Apache-2.0

The surprise of week 3: from outside the top 20 to #3, after reaching the top two of GitHub's daily trending list on 18 October. Its catalogue lists 315 plugins: 53 live in the repo and 262 come from partners such as Adobe, AWS, Atlassian, Airtable, AMD and Google. Anthropic's own include frontend-design, skill-creator, build-mcp-server and claude-md-improver.

Its own README carries the key caveat: Anthropic does not control or verify what third-party plugins contain, so check each plugin's homepage before installing.

Install

/plugin install <plugin-name>@claude-plugins-official

28. anthropics/knowledge-work-plugins

Anthropic's plugins that turn Claude into a specialist for your role, not just for coders.

• Helps with: everyday knowledge work in sales, marketing, finance, legal, customer support, product, HR, data and operations, with connectors to the tools each team already uses.
• Use case: the marketing plugin drafts content, plans campaigns, enforces brand voice and briefs you on competitors, pulling from tools such as Canva, Figma, HubSpot and Ahrefs.
• Works in: Claude's desktop app and Claude Code.
• Site: none beyond GitHub.
• Local AI: built around Claude and cloud connectors.
• October: +4,194 stars* · ranks out · #18 · out
• Repo: github.com/anthropics/knowledge-work-plugins · 30,078 stars · 3,396 forks · Apache-2.0

The only repo here aimed squarely at people who don't write code. It started with 11 open-sourced plugins and now has folders for 15 roles plus partner-built plugins, 252 skills in all. Each plugin bundles skills, connectors, slash commands and sub-agents, and is meant to be customised with your own company's terms and processes.

Install

claude plugin marketplace add anthropics/knowledge-work-plugins
claude plugin install marketing@knowledge-work-plugins

Also tracked

These skills repos gained stars in October but never made a weekly top 20. They are the ones to watch for next week's chart.

Figures marked * are from OSS Insight. Google's android/skills is small but climbing fast this week.

New launches to watch

GitHub lists 5,165 repos tagged agent-skills created since 4 Oct. The most-starred by 18 Oct:

• zhuyansen/new-term-radar, 1,094 stars, created 12 Oct: a daily radar of new AI model, product and repo names.
• Tejashmakwana/motionmaxxing, 853 stars, 6 Oct: makes AI-generated motion graphics look professionally made.
• juliobsz/skills, 388 stars, 12 Oct: builds polished websites from a one-line brief.
• zhouwei713/luobo-ppt, 381 stars, 8 Oct: generates editable native PowerPoint decks.
• hanzohanzhe/Grandet, 327 stars, 5 Oct: finds the cheapest API for a given model.
• liyupi/ai-promo-video-maker, 315 stars, 9 Oct: researches a product and makes its promo video.

skills.sh and the copycat "superpowers" repos

Treat skills.sh's 24-hour trending list with caution: its #1 entry looks like a renamed copy of another vendor's skills. The all-time chart is credible; the 24-hour one is not.

• All-time leaders (installs): find-skills (Vercel) 3.9M, grill-me (Pocock) 1.4M, grill-with-docs 1.2M, agent-browser (Vercel) 1.2M, then Pocock's tdd and improve-codebase-architecture and HeyGen's hyperframes-cli at about 1.1M each.
• 24-hour #1: ai-image-generation from magentosh/superpowers, 29.3K installs, 116.9K across its four skills. Its README is the inference.sh image and video skill pack, published under the name "superpowers" (the name of obra's genuine repo).
• The pattern repeats: five more …/superpowers accounts (bankai-skills, skills-shell, 101-skills, qu-skills, its-a-skill-issue) sit in the top 40 with near-identical counts of 9.7K-9.9K. Two of them show the same 76-commit history as magentosh.
• Also odd: five gencraft-labs/skills entries land within 15.0K-15.3K of each other.
• Why it matters: a skill runs with your agent's permissions. Check the owner and repo before installing anything from a trending list.

How this chart works

Each week runs from Thursday to Wednesday. I add up the stars each repo gained on GitHub every day of that week and rank the top 20. Daily counts come from each repo's page on gittrend; where gittrend has no daily figures, I use OSS Insight's star history and mark it with *. Where both exist, they agree within about 3%.

• Which repos count: repos that ship agent skills (a SKILL.md file or a plugin of skills) that you install into an agent, found through gittrend's AI-skills list, its weekly digests, GitHub's trending pages and skills.sh.
• Why stars: stars show what developers are excited about this week. They don't prove a skill is good or safe, so every profile says what it needs and how to check it.
• Profiles: facts from each project's README, website and code. Figures for speed or savings are the makers' own unless I say otherwise.
• Screenshots: taken on 18 October 2026, signed out of GitHub.

Frequently asked questions

What are the best agent skills right now?

By new GitHub stars, the top three for 15-17 October 2026 are anatomy, ai-job-search and Anthropic's claude-plugins-official. Over the whole of October so far, Matt Pocock's mattpocock/skills has gained the most, with about 23,100 new stars.

What is the most popular Claude Code skill?

On skills.sh, Vercel's find-skills has the most installs at 3.9 million, followed by Matt Pocock's grill-me at 1.4 million. On GitHub, obra/superpowers is the most-starred skills repo, with more than 300,000 stars.

How often is this chart updated?

Every Thursday morning, covering the seven days from Thursday to Wednesday. The newest week goes at the top, and earlier weeks stay on the page.

Are agent skills safe to install?

Most are, but a skill runs with your agent's access to your files and keys, and fake skills stole SSH keys in 2026. Check the publisher, read every file and scan it with a tool such as NVIDIA's SkillSpector before installing.

Do agent skills work with local AI models?

Most skills are plain instructions, so they work with whatever model your agent runs, including a local one in LM Studio or Ollama. Results depend on how capable the model is, and some skills need paid cloud services for images, video or 3D.

Read next

• Are agent skills safe to install? - the five-point check
• skills.sh review, Claude plugin directory review and Cursor Marketplace review
• AI agent marketplaces: the 23 biggest, reviewed
• How to run AI agents on your Mac

Sources

• gittrend: October 2026 digest, week 41, week 42, AI-skills list and each repo's daily page, for example anatomy
• OSS Insight star history API, used for ponytail, Agent-Reach, universal-modder, andrej-karpathy-skills, knowledge-work-plugins, text-to-cad, emilkowalski/skills, brag, cursor/plugins and anthropics/skills
• GitHub trending: daily, weekly, monthly
• The Skill Leaderboard; skills.sh leaderboard, 24-hour trending and hot
• Project sites: aihero.dev/skills, ponytail.dev, anatomy, diagramdesign.dev, ecc.tools, hyperframes.dev, archify.si, impeccable.style, skills.addy.ie, openmontage.video, tasteskill.dev, claude-mem.ai, ppt-master examples, graphify.com, uupm.cc, docs.caveman.so, Cloudflare blog, Claude Code plugin docs, K-Dense arXiv paper
• Every repo is linked from the weekly tables, its profile or the Also tracked table; plus magentosh/superpowers
• Each repo's GitHub page, README and website, linked from its profile - checked 18 October 2026