Which Claude Code skills to install
Grouped by the task, not the star count. Two were installed and run; ten
are judged from the repo page alone.
12 repositories · 5 scenarios 2 tested here / 10 not tested Data pulled 2026-10-19
Why this list isn't ranked by stars
Search for “best claude code skills” and every result is ordered by
stars. That ordering measures how far a repository travelled, not how
much it changes your work. A star is one click: someone saw a link,
thought “that looks useful”, and moved on. They did not install it. They
did not run it on a task. They did not find out that it fights their
existing setup or that its one useful skill is buried in forty others.
The highest-starred repository in GitHub's own search for claude skills
makes the problem concrete: ComposioHQ/awesome-claude-skills,
77,406 stars on 2026-10-19. It is a curated list. Install it and nothing
happens. No
instruction file lands in .claude/skills/, no trigger is
registered, no behaviour changes — you have installed a reading list and
told the agent about nothing. sickn33/agentic-awesome-skills
and VoltAgent/awesome-openclaw-skills are the same category:
catalogues of other people's skills, not skills.
The inverse is just as common. A skill that genuinely changes how an
agent works can sit at two thousand stars, because its value only appears
after you install it and run a real task — and that step produces no
stars. So the twelve below are grouped by scenario: dev workflow, design
and frontend, writing and docs, SEO and content, research. Each entry
says what it solves, who should skip it, and the numbers. Nothing else.
How to read the labels Tested here means the skill was installed, run against
the same baseline task with and without it, and written up. Two
entries carry that label. Not tested means exactly
what it says: the entry was judged from the repository page — the
description, the primary language, the last push date, the licence, the
open issue count. We did not install those ten, and we did not read
their READMEs in full. Where an entry says the trap is a data
connection or a stale template, that is a reading of the repository,
not an observation.
Dev workflow
This group changes how the agent behaves while it writes code, not which
framework API it knows. One sets the process, one enforces restraint, one
is about engineering habits. They do not conflict, but two weeks with one
of them will tell you more than installing all three at once.
Solves
Turns a one-line request into a staged workflow: brainstorm the
requirement, write a plan, write tests first, review the diff. The
agent stops typing on the first sentence you give it and starts asking
what the sentence means.
Skip it if
the change is a line of CSS or a typo. The planning overhead costs more
than the edit. Skip it too if your team already has a process and does
not want a skill setting the pace.
301,784 stars · last push 2026-10-10 · MIT · Shell · data 2026-10-19
Read the review →
Solves
Puts three questions in front of every implementation: does this code
need to exist, can the standard library do it, can the platform do it
natively. Only then does it let the agent write something new. The
repository's own line is “the best code is the code you never wrote”.
Skip it if
you work in security, concurrency or money math, where the shortest
path is not the safest one. It also argues with your requirements,
which is useful in week one and irritating the day before a release.
167,323 stars · last push 2026-10-18 · MIT · JavaScript · data 2026-10-19
Read the review →
Solves
A set of production-grade engineering skills, in the repository's
words, aimed at engineering habits rather than any one framework's API.
It is the layer above “does it run”: naming, boundaries, whether the
next person can maintain it.
Skip it if
you want the agent to follow a specific framework's documentation to
the letter. This is not an API reference and does not pretend to be.
108,327 stars · last push 2026-10-13 · MIT · JavaScript · data 2026-10-19
Third-highest star count on this page, and still only “depends”. We
have not run a benchmark on it, so the honest verdict is that we do not
know how much it changes behaviour. That changes when review eight
ships.
Judged from the repo page only.
Full review not published yet.
Design and frontend
The failure mode here is overlap, not quality. All three entries make the
agent produce something better looking than its default; install two and
you will use one. Pick the output you need — a page, an interactive
diagram, or a finished figure — and take one.
Solves
Gives the agent a direction before it writes markup: art direction,
reference images, then a browser check of the rendered page instead of
shipping the first generic AI landing page it can produce.
Skip it if
you already have a design system and only need execution. It argues for
a direction, and that argument will collide with brand rules you did
not write.
99,022 stars · last push 2026-10-19 · MIT · JavaScript · data 2026-10-19
Judged from the repo page only.
Solves
Turns an idea, a plan or a codebase into an interactive architecture
diagram, so the “how does this connect” conversation has a picture
instead of a paragraph.
Skip it if
you need a static image for Figma or a slide deck. The output is HTML
and the export is a step you build yourself. It also carries 252 open
issues and pull requests as of 2026-10-19, so read that list before
trusting it with a deadline.
85,560 stars · last push 2026-10-19 · MIT · JavaScript · data 2026-10-19
Judged from the repo page only.
Solves
Forty-four editorial diagram types as self-contained HTML and SVG. The
description is blunt about what it refuses: no shadows, no Mermaid
slop. It produces a figure you can publish, not a diagram you can
decode.
Skip it if
you need Mermaid source inside Markdown documentation. This outputs
finished HTML and SVG, which does not slot into that pipeline. It
overlaps with archify, so install one of the two.
57,972 stars · last push 2026-10-16 · MIT · HTML · data 2026-10-19
Judged from the repo page only.
Writing and docs
These two pull in opposite directions. One strips the habits that make
text read as machine-written; the other enforces the conventions of an
academic paper. Both are wrong outside their genre, and installing the
wrong one is worse than installing neither.
Solves
Removes the tells: the throat-clearing opener, the mechanical triads,
the summary at the end of every section. The repository describes it in
one line — an agent skill that removes signs of AI-generated writing.
Skip it if
the text is supposed to be structured. Runbooks, API references and
weekly reports need their scaffolding, and this removes scaffolding.
Do not point it at a paper either.
58,233 stars · last push 2026-10-16 · MIT · Python · data 2026-10-19
Judged from the repo page only.
Solves
Packages ML, CV and NLP paper-writing practice — adapted, per the
repository, from Peng Sida's public notes — into skills that give
structure and phrasing advice instead of free-form generation.
Skip it if
you do not write machine-learning papers. Humanities work and
non-English submissions do not map onto it. The last push is
2026-06-23, so check the template against your venue before trusting
it.
7,590 stars · last push 2026-06-23 · MIT · no primary language listed ·
data 2026-10-19
Judged from the repo page only.
SEO and content
The dividing line in this group is data. The same skill with Search
Console access and without it produces a bigger difference than the gap
between any two skills here. Neither entry is useful if it cannot see
your numbers.
Solves
Conversion research, copywriting, SEO, analytics and growth engineering
in one set. It turns “optimise this page” into a sequence of
steps — hypothesise, write, measure — that can be checked.
Skip it if
you have no traffic or conversion data. The advice leans on numbers you
supply, and brand work that is not chasing conversion does not fit the
frame.
55,433 stars · last push 2026-10-15 · MIT · JavaScript · data 2026-10-19
Judged from the repo page only.
Solves
SEO and generative-engine skills plus an MCP server, so keyword
research, rank tracking and site audits run against your real Search
Console, GA4 or ads data rather than guessed keywords.
Skip it if
you do not have Search Console or GA4 access, or you will not route
site data through a third-party MCP endpoint. The value lives in the
connection; without it you have templates. The repository is
vendor-maintained, so decide on the data flow before installing.
6,080 stars · last push 2026-10-13 · MIT · Shell · data 2026-10-19
Judged from the repo page only.
Research and reading
The value in this group is not generation, it is refusing to skip steps:
split retrieval, writing and checking apart, and stop where a person can
intervene. Anyone who wants one-click output will find both entries
annoying.
Solves
Splits academic work into five stages — research, write, review,
revise, finalise — so each one can be inspected instead of accepting
one long generation.
Skip it if
you want the whole paper written for you. This is process, not a
writing machine. The API reports the licence as NOASSERTION, so confirm
the terms before using it at work or in a submission.
52,865 stars · last push 2026-10-18 · NOASSERTION · Python · data 2026-10-19
Judged from the repo page only.
Solves
Structured deep research that stops and asks at key points, so the
agent cannot quietly fill a gap with plausible invention and hand you a
confident report.
Skip it if
you want a report in one command; it will keep asking. The last push is
2026-08-23, and retrieval tools break politely when the sites they read
change, so test it on a question whose answer you already know.
2,381 stars · last push 2026-08-23 · MIT · Python · data 2026-10-19
Judged from the repo page only.