Six GitHub repositories for Claude skills, ranked by stars

Which Claude skills have attracted a following? There are two different counters you can check. GitHub counts people who star a repository. The skills.sh directory counts reported installations of an individual skill across supported agents. Neither source measures how often somebody actually invokes a skill in Claude.

Six repositories by GitHub stars

On November 14, 2026, the first five entries in the Claude Skill Leaderboard were mattpocock/skills, obra/superpowers, multica-ai/andrej-karpathy-skills, anthropics/skills, and DietrichGebert/ponytail. We used that list to find candidates. GitHub supplied every star count in the table. Claude Skill Leaderboard · Repeat the GitHub search.

We made a second discovery pass for large projects the directory might miss. We rounded the fifth candidate’s roughly 190,000 stars down to a search floor of 180,000, then searched for non-forks with “claude” and “skills” in the name or description. GitHub returned two matches: multica-ai/andrej-karpathy-skills and affaan-m/ECC. Both had a SKILL.md with a skill name, description, and instructions on the default branch. Multica was already on the list, leaving six distinct repositories. ECC ranked third once we sorted the combined list by GitHub’s counts. Both discovery sources change over time; the names above record exactly which candidates we used that day.

For each project, we read GitHub’s stargazers_count, chose one skill to illustrate a distinct task, and inspected its SKILL.md at the commit linked below. The table is ordered by repository stars. Its last column shows reported installs of the selected skill across all supported agents, checked on the same date. To check a live star count yourself, open the GitHub API record for Superpowers and find stargazers_count.

¹ Install figures use the rounded values displayed on the linked skill pages. The directory collects opt-in, deduplicated telemetry from its CLI. Manual installations fall outside that count, and people can install a skill for multiple kinds of agents. How skills.sh measures installs.

We selected for variety of tasks, so a different choice from the same repository could change the order by installs. Among these six examples, grill-me led in reported installs at 1.5 million, followed by frontend-design at 1.1 million. Other skills in those repositories have their own counts. Measuring actual use in Claude would require invocation data from Claude sessions.

What each skill does

grill-me — question a plan before writing code

This entry point delegates to the companion grilling skill to interview you about a plan. Give Claude a real decision to explore. Suppose customers can edit a delivery address until a parcel ships. Ask it to interview you about what happens when a warehouse prints a label halfway through an edit. Look for a concrete decision about which address the warehouse should use. Read grill-me at the reviewed commit.

systematic-debugging — trace a failure to its cause

The skill asks Claude to gather evidence, compare a failing case with a working one, and test a hypothesis before changing code. When a checkout test passes alone and fails in the full suite, ask Claude to use the skill to investigate. Give it both test commands and the failing assertion. A useful first result is the smallest sequence that reproduces the failure and an observation that explains the difference. Read systematic-debugging at the reviewed commit.

article-writing — turn working notes into a tutorial

ECC’s writing instructions use your examples and evidence to shape the voice and structure of a draft. Give Claude a command you actually ran, its output, a short description of your reader, and a previously published example of your writing. Ask for a tutorial that opens with the result. Check each technical claim against your notes before publishing. Read article-writing at the reviewed commit.

karpathy-guidelines — keep a code change small enough to verify

The publisher is Multica. Its instructions credit Karpathy’s observations about models making assumptions and overcomplicating code as the inspiration; the skill file makes no claim that he maintains it. It asks Claude to surface assumptions, make a focused edit, and define a check for success. Ask Claude to fix a blank CSV row appearing in an export. Give it a sample with a trailing newline and ask what should happen to an intentionally empty field. Once you settle that distinction, the two examples become concrete checks for the change. Read karpathy-guidelines at the reviewed commit.

frontend-design — design around a real screen and task

Anthropic’s design skill guides choices about type, layout, and color around a specific audience and task. Try it on a film festival schedule. Supply the actual screening times, the existing visual identity, and a single task for a visitor on a phone: find two seats at the next available screening. Review the result with a sold-out show in the schedule so the design has to handle a real constraint. Read frontend-design at the reviewed commit.

ponytail-review — review the behavior behind a diff

The review follows callers, dependencies, and tests to assess the effects of a change. Give the skill a change that adds retries to an order import. Ask it to follow the caller and the code that writes the order. One question worth answering is whether a timeout after a successful write can create a second order on retry. A finding is easier to act on when it names the path and the test that would catch it. Read ponytail-review at the reviewed commit.

Try one skill in Claude Code

For a small first experiment, try Multica’s single-file guidelines. From a project directory, download the exact revision reviewed here into a temporary file and read it. This keeps your experiment reproducible as the repository changes.

curl -fsSL https://raw.githubusercontent.com/multica-ai/andrej-karpathy-skills/2c606141936f1eeef17fa3043a72095b4765b9c2/skills/karpathy-guidelines/SKILL.md -o /tmp/karpathy-guidelines.md
less /tmp/karpathy-guidelines.md

If the instructions suit your project, place the file in Claude Code’s project skill directory. The -i option asks before overwriting an existing copy.

mkdir -p .claude/skills/karpathy-guidelines
cp -i /tmp/karpathy-guidelines.md .claude/skills/karpathy-guidelines/SKILL.md

Start Claude Code from that project. Try this request: /karpathy-guidelines An exported CSV has a blank final row. A field can legitimately be empty. What examples would distinguish the bug from valid data? Look for two tests you can actually run, one for a trailing row and one for a valid empty field. Keep the skill if it helps you catch a distinction you would otherwise miss. Claude Code skills documentation.