The 12 Best Agent Skills for Coding, Research, and Design (Tested)

"Best agent skills" listicles are easy to write and easy to fake — most of them link a repo nobody opening it has ever run. This one comes out of the agent skills hub I maintain: 90 indexed skills across eight categories, every entry linked back to its source.

For this list I didn't just skim descriptions. I cloned all 12 repos, opened every SKILL.md, and checked each against the same bar I described in how to write a SKILL.md: does the frontmatter parse, does the description actually trigger, is the body a workflow rather than vibes, and what ships alongside the file? The caveats you'll see below — long descriptions, missing triggers, oversized bodies — come from that inspection, not from vibes either.

Four for coding, four for research, four for design. Every one links to its source repo, and installation takes about two minutes at the end.

How I tested (so you can repeat it)

For each skill, four checks:

1. Structure — clone the repo, confirm SKILL.md exists where the docs say it does, and that the YAML frontmatter parses with a name and description.
2. Trigger — read the description as the agent would: does it name the artifacts a user would mention, and does it say when to fire ("Use when…")?
3. Body — read the instructions end to end. Is there a decision procedure in here, or just encouragement?
4. Bundle — list what ships next to it: scripts, references, data files, evals. That's the difference between a skill that tells and a skill that does.

Fair warning on scope: this is a file-level inspection of 12 skills, not a benchmark of agents running them. Where I'd want more run time before a strong recommendation, I say so.

Coding & agent building

1. agents-best-practices

Source: DenisSergeevitch/agents-best-practices

The closest thing I've found to a design doc for agent harnesses that doesn't belong to any one vendor. It covers agent loops, tool design, record provenance, permissions, system prompts, context compaction, memory lifecycles, speculative tool execution, and always-on agents — for OpenAI, Anthropic, and OpenAI-compatible APIs alike.

What impressed me in the clone: a 3,700-word body plus a references/ directory and assets — it's written like documentation, with the deep material filed away rather than dumped inline. The one weakness against my own checklist: the 94-word description never says "Use when…", so a harness matching strictly on triggers may not fire it as often as it should.

2. pydantic-ai-harness

Source: pydantic/skills

Straight from the Pydantic team: sub-agents, memory, context compaction, guardrails, and spend limits for Pydantic AI — plus a second skill for migrating deep-agents code into the harness.

Checks: frontmatter clean, description carries an explicit trigger, 2,200-word body with references/ for the deep dives. It's the rare vendor skill that's about your architecture decisions rather than a product tour.

3. deep-agents-core

Source: langchain-ai/langchain-skills

The entry skill for LangChain's Deep Agents framework: create_deep_agent(), harness architecture, the SKILL.md format itself, and configuration. The body walks through the built-in middleware — task planning, filesystem-backed context, sub-agent delegation, long-term memory.

The trigger is the most aggressive I saw: "INVOKE this skill when building ANY Deep Agents application" — 18 words of description, all signal. Shorter than most, and a good argument that you don't need a paragraph if you know exactly what task you own.

4. agent-skills-creator

Source: mblode/agent-skills

The meta-skill: it writes, audits, and merges skills — and ships the evidence to back that up: a scripts/ validator, evals/ with routing scenarios (would a real request activate this skill?), and keep/cut/merge/retire guidance for bloated skill folders.

This is the one skill on the list I'd install before writing any other skill. Its own quality bar is visible in the repo — which is exactly what you want from a skill that claims to judge skills. Note it wants Bash, Ruby with YAML, and Perl to run its validator.

Research & web

5. d-research

Source: d-init-d/d-research-skill

Browser-first deep research: source discovery, evidence ledgers, atomic fact verification, literature and market research, single-URL inspection, scoped OSINT, execution gates, and blocker reports.

The bundle is the story here — scripts/, templates/, schemas/, references/, even release-evidence/ and a tests directory. This isn't a prompt; it's a research pipeline with a config file (research.config.example.json). One honest ding: no literal "Use when…" in the description, though its 70-odd words list enough task phrases ("due diligence," "evidence ledgers") to match on.

6. last30days

Source: mvanhorn/last30days-skill

"What are people actually saying about X in the last 30 days?" — pulls posts and engagement from Reddit, X, YouTube, TikTok, Hacker News, Polymarket, GitHub and the web, then synthesizes. It versions itself (3.31.0), declares allowed-tools, ships an argument-hint so you can call it slash-command style, and includes something I'd like to see more often: a "doctor" health check that diagnoses broken or missing sources when a platform stops answering.

Bundle: scripts, references, agents, assets. Body is long at 4,200 words — the reference material probably belongs in references/ — but as a working research tool it's the most operationally mature skill I inspected.

7. agent-reach

Source: Panniantong/Agent-Reach

Read-only access to a spread of platforms most skills can't touch: Xiaohongshu, Bilibili, X, Reddit, YouTube subtitles, GitHub, V2EX, stock quotes, LinkedIn and job boards, Instagram, Facebook, and podcasts. One CLI, no API fees.

Amusing and instructive detail from the clone: its description is written in two languages — Chinese and English triggers side by side, because users phrase requests in both. If your users aren't monolingual, this is the pattern. The body is trim (330 words) with a references/ folder — about as close to textbook structure as anything on this list.

8. Scrapling (official)

Source: D4Vinci/Scrapling

The official skill for the Scrapling scraping framework: HTTP requests, stealth browsers, JavaScript rendering, browser actions, API capture, concurrent spiders, adaptive parsing.

The description is a masterclass in trigger-writing, so I'll quote its tell: "Use when asked to scrape, crawl, or extract data… web_fetch fails; the site has anti-bot protections." Naming your own failure mode as a trigger is how a skill gets invoked at exactly the moment it's needed — the agent's fetch breaks, and this is the hand going up.

Design & frontend

9. design-taste-frontend (taste-skill)

Source: Leonxlnx/taste-skill

The anti-slop skill: art direction distilled from a study of 247 curated gallery sites, dealt from a library of researched directions — idea, composition, type, palette, photographs — with a strict house style enforced by a real-browser inspector and a look-and-fix loop on screenshots.

Largest bundle on the list: library/, scripts/, references/, evals/, assets/. Also the longest description (145 words) and body (6,500 words) — I'd trim both. But nobody else ships a skill with a screenshot-driven fix loop and an eval directory, and that ambition is why it earned the slot.

10. ui-ux-pro-max

Source: nextlevelbuilder/ui-ux-pro-max-skill

A searchable local corpus: 79 styles, 192 product palettes with reasoning, 74 font pairings, 119 UX guidelines, 105 icons, 17 GSAP presets, 25 chart types, 22 stacks — served to the agent from bundled data files with scripts, so the guidance is lookup, not recall.

That's the design I like most in a skill: put the knowledge in data, let scripts query it. Description carries both the broad trigger ("designing, building, reviewing, or fixing interfaces") and the inventory, so the agent knows the depth before opening it.

11. animate (emilkowalski/skills)

Source: emilkowalski/skills

From Emil Kowalski (Sonner, Tweakcn): build an animation from scratch in the order that determines whether it feels right — should it animate at all, what purpose, which tool, which properties, which curve and duration, how it interrupts, how it exits — then write the implementation.

72 words of description, explicit triggers for both the build case ("animate something, add motion, make a component feel alive") and, neatly, a pointer to its sibling skill for the critiquing case. The body (1,800 words) is one tight procedure with no bundle at all — proof that a skill doesn't need scripts to be excellent. It needs an order of operations.

12. gsap-core (official GSAP skill)

Source: greensock/gsap-skills

The official skill from the GSAP team: .to(), .from(), .fromTo(), easing, duration, stagger, defaults, and gsap.matchMedia() for responsive and reduced-motion work. Eight skills ship in the same repo — timeline, ScrollTrigger, plugins, React, Vue, performance, utils — and gsap-core is the entry point.

Description does double duty: trigger words and a recommendation ("recommend GSAP when the user needs timelines, scroll-driven animation, or a framework-agnostic lib"). Vendor skills earn trust by telling you when not to use the vendor — this one does.

What the 12 have in common

Three patterns survived every inspection:

• The description is the product. The best ones (Scrapling, deep-agents-core, animate) read like trigger lists, not taglines. They name failure modes, artifacts, and the exact verbs users type.
• Bundles beat length. The skills I'd trust most ship scripts, data, and evals; the weakest signal on the list is a 4,000-word body with nothing beside it. Knowledge in data, procedures in steps, deep dives in references/.
• Every one of them encodes an order of operations. Do this first, check this, then that — the thing a senior person does automatically and a model won't invent on its own. That's the whole value of a skill.

Installing any of these

Each source link above goes to the repo. Copy the folder that contains SKILL.md into your agent's skills directory — ~/.claude/skills/ for Claude Code personal skills, .claude/skills/ in a project so it travels with the repo. Other harnesses use their own equivalent folder; the format is identical.

Looking for more? The agent skills hub indexes 90 skills across eight categories — browse the category tables on GitHub, no clone needed.

Further reading

• How to Write a SKILL.md — the anatomy these 12 were checked against, if you're writing your own.
• Agent Skills vs MCP vs Subagents — where skills fit relative to the other two ways you extend an agent.
• The agent skills hub — the full index itself, on GitHub.