Skip to content

External Skill Collections

Individual Claude Code skills you can adopt à la carte — as distinct from the full marketplaces that ship an entire plugin.json-managed catalog. Collections on this page are typically installed by copying or symlinking a skill folder into .claude/skills/ (project scope) or ~/.claude/skills/ (user scope), sometimes via a third-party installer such as npx skills add owner/repo. None of this flows through Claude Code's plugin/marketplace mechanism, so there is no built-in version pinning, update channel, or review gate — treat installation the way you would treat adding any third-party dependency to a project.

Curated Collections

These three collections were independently confirmed to exist and match their description at the time of writing:

Collection Description Repository
Superpowers Process-layer skills — TDD, systematic debugging, brainstorming, code review, and worktree-based collaboration patterns that a coding agent is instructed to invoke before acting. obra/superpowers
mattpocock/skills Workflow-enforcement skills (planning, TDD, debugging, domain modeling, merge-conflict resolution) rather than raw capability extension. Ships its own script for linking selected skills instead of the whole repository. mattpocock/skills
gstack An opinionated multi-skill setup published under the garrytan GitHub account — CEO, design, eng-manager, QA, and security review personas exposed as slash commands. garrytan/gstack

Further Curated Lists

Beyond individual collections, third parties periodically publish broader roundups. One such list, reproduced below with its original categories, is Charlie Hills' "22 Claude Skills for Power Users" (MarTech AI Newsletter, July 2026).

Unverified third-party content

Only the three entries above have been independently confirmed. The table below is reproduced from the source as-is, with popularity metrics removed because they could not be verified and looked implausible for single-purpose skill repositories. Confirm that a repository exists, is actively maintained, and matches its description before installing anything from it.

Skill Description Repository
superpowers Plans and debugs before it writes obra/superpowers
gstack 23 skills in one setup garrytan/gstack
caveman No-fluff answers, no padding JuliusBrussee/caveman
ponytail Says nothing, writes one line (unrelated to this marketplace's own core:ponytail skill) DietrichGebert/ponytail
codex Runs Codex inside Claude Code openai/codex-plugin-cc
i-have-adhd Stops it burying the answer no repository given in the source
Skill Description Repository
ui-ux-pro-max Palettes and fonts that work nextlevelbuild/ui-ux-pro-max
taste-skill Makes a page look expensive leonxlnx/taste
impeccable The polish pass on any front end pbakaus/impeccable
hyperframes Turns a web page into a video heygen-com/hyperframes
emil Beautiful interface animations emilkowalski/emil
gsap Smooth web animations greensock/gsap
Skill Description Repository
skill-creator Build your own skills anthropics/skill-creator
graphify Turns notes into a map Graphify-Lab/graphify
last30days What the internet said this month mvanhorn/last30days
agent-browser Lets Claude use any website vercel-labs/agent-browser
find-skills Finds any skill for you vercel-labs/skills
claude-hud Shows what Claude is doing live jarrodwatts/claude-hud
Skill Description Repository
remotion Build full videos in React remotion-dev/remotion
marketingskills 40 marketing skills in one pack coreyhaines31/marketingskills
humanizer Strips the AI tells from your writing blader/humanizer
social-media-skills Charlie Hills' own content pack charlie947/social-media-skills

Using External Skills Without Bloating Your Context

Every installed skill's name and one-line description is loaded into the model's context on every turn, so that the model can decide whether to invoke it — this happens whether or not the skill is ever triggered. A large, indiscriminately installed collection therefore has a real, recurring token cost even on sessions that never touch it. Treat skill adoption as a curation problem, not a one-time install:

  • Prefer project scope over user scope. Put skills in a repository's .claude/skills/ rather than the global ~/.claude/skills/ whenever they are only relevant to that project. The token cost then only applies to sessions that actually work in that repository.
  • Cherry-pick, don't bulk-install. From a 20+ skill pack, symlink or copy only the two or three skills you actually invoke regularly, rather than pulling in the entire collection. Several collections (including mattpocock/skills) ship a helper script for exactly this purpose — check the repository's scripts/ or README.md before resorting to a manual copy.
  • Trial in isolation first. Symlink a new skill into a scratch project's .claude/skills/ inside a disposable worktree before adding it to your global (~/.claude/skills/) configuration, so a broken or overly broad trigger description does not affect every other session. (claude --plugin-dir <path> is a different mechanism — it loads a full plugin scaffold with .claude-plugin/plugin.json, not a bare skill folder.)
  • Write (or edit) trigger descriptions narrowly. A skill's description frontmatter is what the model matches against; an overly broad description causes false triggers, which then pull the skill's full body into context unnecessarily.
  • Audit periodically. Revisit .claude/skills/ and ~/.claude/skills/ every so often and remove anything you have not invoked in months. There is no built-in expiry — unused skills persist and keep costing tokens until removed.

A Note on Trust

A skill is not passive documentation — its instructions are read and acted on by the coding agent, and some collections bundle executable scripts invoked by hooks or slash commands. Review a skill's SKILL.md and any bundled scripts before installing it, exactly as you would review a third-party dependency's source before adding it to a project. This applies doubly to collections aggregated from many different, unrelated authors, such as the roundup above.

See Also