The top 3 modular agent skills in every category โ ranked by community signal, real-world utility, and momentum over the last 7 days.
If you're new to the modular skills ecosystem, this is your weekly cheat sheet for what's worth installing. Every skill here follows the open SKILL.md standard โ a folder with a SKILL.md (YAML frontmatter + Markdown instructions), optionally bundled with scripts/, references/, and assets/. They run across Claude Code, Codex, Cursor, Gemini CLI, Copilot, and 40+ other agents.
This week covers 7 categories ร 3 ranks = 21 skills. We hunted these across GitHub trending, the skills.sh install leaderboard (133K+ skills from 8.8K publishers), Anthropic and Vercel collections, SkillsMP, LobeHub, Reddit/HN, and the SkillsBench quality benchmark. Notable: last week's Context Engineering podium (context-compression, memory-systems, data-structure-protocol) is still inside our 30-day dedup, so that category sits out this week โ exactly what the dedup rule is designed to catch. One exception is flagged below where a major version update clears it.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | Agent Skills (24-skill suite) | Addy Osmani | addyosmani/agent-skills | 86.1K โญ / 9.2K forks | SKILL.md + references | The full SDLC as 8 slash commands |
| ๐ฅ | Agent-Skills-Hunter | ZhanlinCui | zhanlincui/agent-skills-hunter | 180 โญ / 24 forks | SKILL.md + scripts | skillctl CLI, CI-validated, 51 real skills |
| ๐ฅ | coding-skills | swell-agents | swell-agents/coding-skills | 2 โญ | SKILL.md + agents + commands | Canonical, harness-agnostic TDD & review |
๐ฅ addyosmani/agent-skills โ The Google Chrome engineering lead's take on agentic SDLC. 24 skills map cleanly to DEFINE โ PLAN โ BUILD โ VERIFY โ REVIEW โ SHIP, exposed as 8 slash commands (/spec, /plan, /build, /test, /review, /webperf, /code-simplify, /ship). The killer detail: /build auto implements every task in one approved pass while keeping each task test-driven and individually committed. MIT-licensed, free, installable in one command via npx skills add addyosmani/agent-skills.
๐ฅ Agent-Skills-Hunter โ The first true skill manager. It's not a link list โ 51+ skills ship fully implemented in-repo, and the skillctl CLI handles search/enable/disable/update per IDE across 11 IDEs. GitHub Actions validate every SKILL.md's frontmatter. A small star count but a genuinely novel approach to the "which of my 500 skills actually work" problem.
๐ฅ swell-agents/coding-skills โ A canonical, portable engineering skill set (TDD cycles, code review, architecture design, commit hygiene, per-language conventions for Python/Go/Solidity/shell). It's harness-agnostic by design and ships parallel review agents (@code-reviewer, @security-auditor, @architect-review). Tiny community so far, but it's the cleanest reference implementation of composable workflow+rule skills we found this week.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | garden-skills | ConardLi | ConardLi/garden-skills | 10.2K โญ | SKILL.md + recipes | web-design-engineer w/ 25 style recipes |
| ๐ฅ | designer-skills | Owl-Listener | Owl-Listener/designer-skills | 1,983 โญ | SKILL.md (33 plugins) | 239 skills, researchโdelivery |
| ๐ฅ | awesome-design-md | ara.so | aradotso/trending-skills | 621 โญ / 19.7K | SKILL.md + DESIGN.md corpus | "Make it look like Stripe/Vercel/Linear" |
๐ฅ garden-skills โ Promoted straight off last week's watch list. Five production skills, headlined by web-design-engineer with 25 distinct style recipes โ a concrete answer to the "every AI UI looks the same" complaint. Instead of abstract taste advice, it gives agents named, loadable visual directions.
๐ฅ designer-skills โ 239 skills across 33 plugins spanning the entire design lifecycle, from research through delivery. It's the most comprehensive design-skill collection in the ecosystem, and it graduated from watch list to podium this week on the strength of its breadth and active maintenance.
๐ฅ awesome-design-md โ A clever twist: ship a curated corpus of DESIGN.md files reverse-engineered from popular sites so an agent can apply a specific design system on demand. Triggers like "make my UI look like Stripe/Vercel/Linear" map directly to a drop-in design token set. 19.7K installs on the marketplace and climbing.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | visual-skills | smixs | smixs/visual-skills | Growing | SKILL.md + reference files | AI film director (Murchโprompt syntax) |
| ๐ฅ | generative-media-skills | calesthio | calesthio/generative-media-skills | Research-backed | SKILL.md + EVAL.md | 153 packages, 18 media domains |
| ๐ฅ | crun-agent-skills | Crun AI | CrunTeam/crun-agent-skills | New (Aug 5) | SKILL.md + examples | Image/video/music/speech in one set |
๐ฅ smixs/visual-skills โ This is the "taste layer" media generation has been missing. The SKILL.md is a thin router; the craft lives in reference files the agent is forced to load in order: dramaturgy distilled from Walter Murch, Kurosawa, Fincher, Spielberg, and Bong Joon-ho, paired with verified prompt syntax for Seedance 2.5, Kling 3.0, Veo 3.1, Nano Banana 2, and GPT Image 2. Model references are checked against official vendor docs (July 2026).
๐ฅ generative-media-skills โ 153 independently-researched skill packages spanning 18 domains (3D, image, video, music, TTS, lip-sync, motion capture, world modelsโฆ). Every skill ships a repo-only EVAL.md with subject-specific scoring and critical-failure tests โ the most rigorous quality bar we've seen in a media skill collection. Free/open.
๐ฅ crun-agent-skills โ Announced Aug 5, 2026 as "build AI media generation agents in minutes." Modular, production-ready skills for image, video, music, and speech synthesis backed by Crun AI's unified API. Early days, but it's the freshest enterprise-style media skill launch this week.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | deer-flow deep-research | ByteDance | bytedance/deer-flow | 79.2K โญ / 10.8K forks | SKILL.md (public skills dir) | Systematic multi-angle methodology |
| ๐ฅ | Deep-Research-skills | Weizhena | Weizhena/Deep-Research-skills | 1,926 โญ / 156 forks | SKILL.md + agents | Two-phase, human-in-the-loop |
| ๐ฅ | last30days | (community) | nanoskill.ai social-listening | Active | SKILL.md | Engagement-weighted trend research |
๐ฅ deer-flow deep-research โ ByteDance's agent workflow ships a deep-research public skill that's the cleanest articulation of "research before you generate" we've seen. It forces multi-angle exploration, deep dives per dimension, and diversity/validation checks before any content generation. 79.2K stars on the parent repo give it the strongest community signal in this category by far.
๐ฅ Deep-Research-skills โ A structured two-phase workflow (extensible outline generation โ deep parallel investigation) with human-in-the-loop control at every stage. Commands like /research, /research-deep, and /research-report compose into a full paper/market/due-diligence loop across Claude Code, OpenCode, and Codex. Passed SkillsLLM's security scan.
๐ฅ last30days โ For "what is the internet saying right now" briefs. It searches Reddit, X, YouTube, Hacker News, and Polymarket in parallel, scores results by real engagement (upvotes, likes, prediction-market odds), and synthesizes a grounded brief via an AI judge. Zero-config for Reddit/HN/Polymarket/GitHub.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | agent-orchestration | yonatangross | yonatangross/orchestkit | 208 โญ / 20 forks | SKILL.md + rules/ | 10 rules, 4 coordination patterns |
| ๐ฅ | agent-orchestrator | SuperCorks | SuperCorks/agent-skills | Active | SKILL.md + Python script | Awaited worker sessions, run capture |
| ๐ฅ | agent-dispatch-harness | SUNRNEHUI | SUNRNEHUI/multi-agent-dispatcher | Active | SKILL.md | Direct/Lite/Full mode selection |
๐ฅ agent-orchestration โ Ten on-demand rule files across four categories (agent loops, multi-agent coordination, framework comparison, multi-scenario). It's grounded in concrete decision parameters โ max steps, temperature, memory window, supervisor routing โ rather than vibes. Notably updated for Claude Code 2.1.154's native /workflows, which it correctly positions as complementary rather than competitive.
๐ฅ agent-orchestrator โ A disciplined workflow plus a local helper script for launching awaited Codex/Claude/OpenCode worker sessions from the current thread, with sensible defaults (45-min timeouts, no worktrees unless asked, captured runs under .agent-orchestrator/runs/). It explicitly refuses to make background agents the default โ a refreshing discipline in a category prone to over-orchestration.
๐ฅ agent-dispatch-harness โ The routing authority for multi-agent requests. It picks the lightest mode that can finish: Direct Mode for small edits, Lite Orchestration for separable slices, Full Harness only for long/resumable/risky work with durable state and verification evidence. Anti-"coordination theater" by design.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | agent-skill-eval | Levente Csoke | tardigrde/agent-skill-eval | Active (PyPI v0.7.0) | CLI + config | Real-harness eval, pass@k, state-diff |
| ๐ฅ | agent-skills-eval | Rishabh Mehan | darkrishabh/agent-skills-eval | 2.5K weekly DL | TS CLI + HTML report | LLM-judge + baseline A/B |
| ๐ฅ | agent-eval | fsilavong | fsilavong/agent-eval | Active | SKILL.md + references | Component + E2E, reproducible |
๐ฅ agent-skill-eval โ The most rigorous answer to "does my SKILL.md actually make the agent better?" It runs your skill through the real agent CLIs (Claude Code, Codex, OpenCode) with and without the skill, grades with deterministic state-diff checks plus an LLM rubric, and reports pass rates, pass@k, token cost, and wall-clock. It even catches the failure mode that matters most: the agent never triggering your skill at all.
๐ฅ agent-skills-eval โ A TypeScript test runner with the same core idea (with-skill vs without-skill, LLM judge, baseline report) but a gentler on-ramp: npx agent-skills-eval ./skills --target gpt-4o-mini --judge gpt-4o-mini --baseline. Emits a static HTML report. 2.5K weekly npm downloads signals real adoption.
๐ฅ agent-eval โ Evaluates any agentic system (RAG pipelines, multi-agent, chains) at component and end-to-end level, producing a structured, reproducible report with metrics, datasets, and a regression strategy. Every run records model, prompt template, and seed โ the discipline most eval tooling skips.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | agent-governance-toolkit | Microsoft | microsoft/agent-governance-toolkit | Official | SKILL.md + policy engine | 10/10 OWASP Agentic Top 10 |
| ๐ฅ | awesome-agent-skills-security | LLMSecurity | LLMSecurity/awesome-agent-skills-security | Curated | Reference + models | Governance research atlas |
| ๐ฅ | agentic-skills-top-10 | kenhuangus | kenhuangus/agentic-skills-top-10 | Active | SKILL.md + checklist | AST09 inventory & approval |
๐ฅ agent-governance-toolkit โ Microsoft's official governance layer: policy enforcement, zero-trust identity, execution sandboxing, and reliability engineering for autonomous agents, covering all 10 OWASP Agentic Top 10 risks. This is the category that matters most right now โ Snyk's ToxicSkills research found prompt injection in 36% of tested skills, and an audit of 22,511 skills surfaced ~6.3 issues each. Governance skills are moving from "nice to have" to "table stakes."
๐ฅ awesome-agent-skills-security โ A curated atlas of security research and governance models (attested-evidence governance, adaptive runtime control, memory-poisoning defenses). Less a drop-in skill than a reference library for anyone building or auditing agent security โ and the single best starting point we found for the literature.
๐ฅ agentic-skills-top-10 โ A practical, field-oriented security skill built from real 2026 incidents (the ClawHavoc supply-chain attacks, Claude Code repo-config RCE CVEs, Meta's agent-email-deletion incident). Ships an inventory โ approval โ audit-logging โ identity-controls framework (AST09) plus a review checklist. Grounded in the Cisco finding that only 34% of enterprises have AI-specific security controls.
grill-me skill is still in our 30-day dedup; the version update is what makes it newsworthy.context-compression; budget policy, observation masking, and KV-cache strategy. Likely to lead a returning Token Reduction category once the 30-day window clears.Ranked by AgentSkillReview. Data sourced from GitHub, skills.sh leaderboard (133K+ skills), SkillsMP, LobeHub, community forums, and the SkillsBench quality benchmark. Install counts and stars verified against live sources Aug 12โ14, 2026.
Methodology: (1) community signal โ GitHub stars, skills.sh installs, weekly growth; (2) recency โ commits in past 30 days; (3) modularity โ clean SKILL.md + optional scripts/references; (4) real-world utility; (5) innovation. All skills follow the open SKILL.md standard. 30-day dedup enforced against reviewed_skills. No internal skills reviewed.