best 14 claude code skills

seventeen skills including algorithmic-art, brand-guidelines, canvas-design, doc-coauthoring, frontend-design, mcp-builder, skill-creator, theme-factory, web-artifacts-builder and webapp-testing, plus the agent skills specification itself.

the four document skills — docx, pdf, pptx and xlsx — are not demos. they are the production skills powering claude's file-creation features, which makes them the most battle-tested things in the collection. `webapp-testing` is a real playwright toolkit and `mcp-builder` guides mcp server design.

install with `/plugin marketplace add anthropics/skills` then `/plugin install document-skills@anthropic-agent-skills` or `example-skills@anthropic-agent-skills`.

two caveats that get misreported constantly. there is no root licence file: most skills are apache-2.0 but the four document skills are source-available and not open source, so anyone redistributing needs to read the third-party notices. and anthropic's own disclaimer describes these as for demonstration and educational purposes — this is a reference collection, not a curated best-of.

this is matt pocock's actual working setup, published. the positioning is deliberate and correct: small, easy to adapt, composable — the opposite of the frameworks that want to own your entire development process.

it includes `/triage` with github, linear and local issue-tracker integration, and an interactive `/setup-matt-pocock-skills` configuration step. his `grill-me` skill alone reports over 629,000 installs.

the install design is unusually well thought out: `npx skills@latest add mattpocock/skills` to copy them, or install as a managed plugin to subscribe to updates rather than fork them. that choice — subscribe versus fork — is one most projects don't offer.

the honest limits: it is one person's personal workflow, tuned to typescript and web development, with newsletter-funnel branding through the readme and a bus factor of one.

`npx skills add owner/repo` is now effectively the standard way to install a skill, writing into `~/.agents/skills/` — the shared location around fifty agents now read. it accepts github shorthand, full urls, gitlab, any git url and local paths, and supports tag and commit-sha pinning plus `--global` and `--agent` targeting.

`npx skills use owner/repo@skill | claude` runs a skill without installing it at all, which is an elegant touch.

vercel also publishes its own collection, `vercel-labs/agent-skills`, with react and next.js best practices and web design guidelines. its `find-skills` skill reports 2.6 million installs, the highest in the directory.

two things to weigh. neither github repository carries a licence file, despite the npm package declaring mit — an ambiguity that will stop some legal reviews cold. and the discovery layer for an anthropic-originated open standard is now run by vercel, which is a concentration worth being aware of even if nothing has gone wrong.

twenty-four skills and eight slash commands mapped onto a lifecycle: `/spec` → `/plan` → `/build` → `/test` → `/review` → `/ship`. named skills include a five-axis code-review-and-quality, an `interview-me` that asks one question at a time, and a test-driven-development skill.

the modularity is what puts it above most collections. `--list` lets you browse first and `--skill <name>` installs one thing, so you aren't forced to take all twenty-four and pay the context cost for skills you'll never trigger.

the author is a credible name in web development, which matters in an ecosystem where anyone can publish anything.

one documented rough edge: a git ssh permission failure on windows and macos that requires a global git ssh-to-https rewrite. it also overlaps meaningfully with superpowers and mattpocock's collection — pick one lifecycle opinion, not three.

thirty-three documented patterns with before-and-after examples for stripping the tells out of ai-written prose. that's it. no scripts, no dependencies, no server.

it ranks this high partly on quality and partly as an argument: this is what a skill is for. one clearly-scoped job, expressed as instructions, portable to any harness, and auditable because you can read every pattern it applies.

install with `npx skills add blader/humanizer --global`, or `--agent '*'` for every harness on the machine.

two limits. it is prose-only, and it can flatten intentional voice along with the tells — if you have a deliberate style, check what it removed. worth noting too that 'humanising' has adjacent uses in academic and seo contexts that are less respectable than editing your own writing.

roughly thirty-five skills acting as ceo, designer, engineering manager, release manager, doc engineer and qa: `/office-hours`, `/plan-ceo-review`, `/design-shotgun`, `/review`, `/ship`, `/canary`, `/qa`, `/autoplan`, `/freeze`, `/guard`, `/retro`, `/learn` and more.

most collections in this category are engineering workflow. covering the roles around engineering is genuinely differentiated, and the founder credibility behind it explains the 124,000 stars.

the install is the weak point and worth being explicit about: it is not a plugin. you `git clone` it directly into `~/.claude/skills/gstack`, run a `./setup` script, and then manually add a section to your `CLAUDE.md`. no marketplace, no auto-update, no clean uninstall path.

installing thirty-five skills at once is also a large context footprint, and its `/browse` skill asks you to disable claude code's built-in browser tools, which is more invasive than most people will expect.

graphify turns a codebase — plus its documentation, sql schemas, configs and pdfs — into a queryable knowledge graph, using deterministic local ast parsing rather than embedding-based similarity. for questions where the right answer is structural, that is more reliable than vector search.

ingesting non-code artefacts alongside code is unusual and useful; most tools in this space see only source files.

install needs python 3.12 and `uv`, then `graphify install`. use `graphify install --platform agents` to target the cross-framework skills directory — the default is claude-code-only, so cross-agent users silently get less than they expect.

this is the most contested niche in the ecosystem and nobody has won it. `codegraph` has the best developer experience with pre-indexing and auto-sync on change, though its install wires the agent without indexing anything until you run `codegraph init` per project. `Understand-Anything` is the most learnable, aimed at graphs that teach rather than impress. all three are curl-pipe-shell style installs.

it makes the model write like a caveman, which cuts output tokens by a claimed sixty-five percent. that is the entire product and it is very funny, and it also works.

the reason it's ranked seriously is that output tokens are a real cost line, and for automated pipelines where no human reads the intermediate reasoning, terse output is strictly better. best value per line of code in the ecosystem.

install via the marketplace, `npx skills add JuliusBrussee/caveman`, or a shell installer.

the sixty-five percent figure is the author's own and we could not verify it. and it is obviously terrible anywhere you need to read and understand the reasoning — which is most interactive work. push cadence is the slowest in the top tier at around three weeks, though nowhere near abandoned.

an agent-led recency search, scored by upvotes, likes and real money rather than editorial picks. it exists to answer 'what is current' when the model's training data stops months ago — a problem anyone building with these tools hits constantly.

install with `/plugin marketplace add mvanhorn/last30days-skill` and `/plugin install last30days`, or cross-agent with `npx skills add mvanhorn/last30days-skill -g`.

the documentation deserves specific credit. it explicitly warns that claude code does not deduplicate across install methods, so installing both the plugin and the npx copy makes `/last30days` appear twice, and tells you to use one method per machine. almost nothing else in this ecosystem is that candid about its own rough edges.

it depends on external sources and apis, and its ranking methodology is its own opinion rather than a neutral measure. the readme also notes SKILL.md is the source of truth over the readme, so the readme can lag.

agent skills for working with obsidian vaults, written by obsidian's ceo. explicitly built to the agent skills specification, so it works across claude code, codex and opencode rather than just one.

for the largest personal-knowledge-management ecosystem there is, having a first-party-adjacent skill set from the person running the company is about as authoritative as this category gets.

install with `/plugin marketplace add kepano/obsidian-skills` then `/plugin install obsidian@obsidian-skills`, or via `npx skills add`.

the flag: at roughly six weeks since the last push it is the slowest-moving thing in the top tier here. that is not abandoned by any reasonable definition, and it's a stable domain, but everything ranked above it ships weekly.

skills for scientific and research workflows — a genuinely underserved audience in an ecosystem where almost everything assumes you're building a web app.

the install documentation is the best in this entire category. it explains the `~/.agents/skills/` convention rather than assuming you know it, names the compatible runtimes explicitly, supports interactive browsing and single-skill installs, and offers release-tag and commit-sha pinning. that last one is real supply-chain hygiene that most projects here don't bother with.

install with `npx skills add K-Dense-AI/scientific-agent-skills`, with a manual git clone path documented too.

it's narrow by design, and it's startup-backed, which carries the usual risk that priorities shift toward a commercial product.

browser automation that runs sandboxed javascript in a quickjs wasm sandbox with no host access, keeps pages persistent across scripts, auto-connects to a running chrome instance, and exposes the full playwright api.

the sandbox is the differentiator and it is genuine engineering, not marketing. compared with handing an agent raw playwright, having a real isolation boundary between the automation script and your machine is a meaningful difference. the project also publishes benchmarks, which almost nobody in this ecosystem does.

install with `npm install -g dev-browser` then `dev-browser install`. its readme labels the plugin marketplace route as legacy, so ignore older guides that point there.

at 6,500 stars it's the least adopted thing in this ranking, and it's vendor-affiliated. the quality is well ahead of the numbers.

skills for job searching: cv assessment against an a-to-f rubric, application tracking and interview preparation.

it's included partly on merit and partly to make a point. almost everything in this category assumes you're writing code, and the skill format is really just structured judgement expressed in markdown — which applies to plenty of work that isn't engineering. 61,000 stars suggests the audience agrees.

the graded rubric is the useful bit. 'improve my cv' produces mush; scoring it against defined criteria produces something actionable.

it's outside our own area of expertise to judge the career advice itself, so we're ranking the design and adoption rather than vouching for the guidance.

a wide collection installed with `npx agentic-awesome-skills`, writing into the shared `~/.agents/skills` location with per-host flags for cursor, gemini, codex and others.

it does the cross-agent thing properly, which a surprising number of collections still don't, and it's actively maintained with 44,000 stars.

there is nothing here you couldn't assemble yourself from the more focused entries above, and breadth collections carry the same problem throughout this category: every installed skill costs context whether or not it ever fires.

worth knowing about as a browsing surface. we'd install specific things from it rather than the whole set.