FDF Labs
Cult of Claude

Every Claude Code skill, in one search box.

Cult of Claude is a directory of Claude Code skills and agents. A crawler goes through public GitHub every couple of hours, finds SKILL.md files, agent definitions and slash commands, leaves out what isn't worth listing, and publishes a page per entry with its source and the command to install it. FDF Labs builds and runs it. It is independent and not affiliated with Anthropic.

commit message

71 strong matches · 3,752 weaker further down

  1. Draft Commit Message

    27

    Analyze the staged and unstaged changes, then output a single conventional commit message line.

    Git community

  2. Pvc Git Commit Message Local

    26

    Claude Code skill that writes a git commit message in the PVC format (vX.Y.Z headline plus one bullet per real change, version bumped from the diff) and runs the commit locally on the named paths.

    Git community

  3. Gen Commit Msg

    51

    Generate intelligent commit messages based on staged changes. Invoke with /gen-commit-msg.

    Git community
A real search on cultofclaude.com on 24 October 2026: the first three of 3,823 skills that mention either word. The number on each card is its score, out of 100. Author and repository are left out here.

From 36,600 skills to 24 cards.

Search runs in a Cloudflare Worker, not in your browser. The index is split into small files, and each query reads only the ones its words touch.

q = commit message/api/search
  1. 36,600

    skills in the index

    meta.json, 1.8 MB: each one's score, stars, category and repository. Read once per Worker instance, not per visit.
  2. 2,125

    with “commit”

    Shard 252 of 256, 46 KB. Every hit already carries its weight: in the name it counts for more than in the description.
  3. 1,975

    with “message”

    Shard 228 of 256, 46 KB.
  4. 3,823

    with either word

    Those with both words come first, and a rare word weighs more than a common one.
  5. 71

    strong matches

    Those scoring at least a quarter of the best one. Each further result from the same repository keeps 65% of the previous one's weight, so one repository can't fill the page.
  6. 24

    on the first page

    Their cards come from 24 of the 2,048 card files: 139 KB.

11 KBis what your browser gets: one response. Before, every visitor downloaded the whole catalogue to search it, about 2.4 MB compressed when it held 20,993 entries.

The query from the top of the page, replayed on 24 October 2026 against the site's public files with the same ranking code the Worker runs. It matched the site: 3,823 and 71. The bars are on a log scale.

From a file on GitHub to an entry page.

The site hosts nothing. Every entry points to the repository it came from, and whatever fails the filter gets no page.

  1. Find

    The crawler searches public GitHub for SKILL.md files, agent definitions and .claude/commands/ files. Each file is one entry.

  2. Dedupe

    When the same skill turns up in two collections, the copy from the repository with more stars stays.

  3. Filter

    Every entry is scored from 0 to 100 before it is listed. Under 40, it is quarantined.

  4. Enrich

    Sparse entries get the README from their folder or their repository, and each one is filed under a category.

  5. Publish

    One page per entry, with its source, the tools it declares and the install command. The site rebuilds after every crawl.

The filter, point by pointunder 40: no page
A README over 100 characters
+15
A trigger condition
+10
A long description that differs from the short one
+10
… and 400 characters or more
+5
A description over 40 characters that isn't a template
+10
An official Anthropic origin
+10
A declared tool list
+5
An install command
+5
A skill, agent or command path
+5
Stars: over 0, over 50, over 500
+5 each
An author with fewer than 50 entries
+5
A bulk template description
−30
Describes an MCP server, not a skill
−25 / −20
A description identical to the name
−25
An author with more than 100 entries
−15
Short and long description the same, under 25 characters
−15
No README, body, trigger or tools
−10
A description under 20 characters
−10
An off-topic source, such as a list of links
−100

As written in the filter's code. It doesn't measure whether a skill works, and it isn't a security review: it measures documentation and adoption.

Before you install, it says where it writes.

A skill has no package manager: installing it means copying its author's file onto your machine. Each entry page gives the exact command for what it is and where it comes from.

~/.claude/
agents/
  <name>.md       ← curl
commands/
  <name>.md       ← curl
skills/
  <name>/
    SKILL.md      ← curl
  <name>/         ← git clone + cp -r
One file
Agents, slash commands and a lone SKILL.md: one curl straight into its folder. It asks for HEAD, so it works whether the repository uses main or master.
A folder
A skill with scripts and resources beside its SKILL.md: the repository is cloned and the whole folder copied.
No command
If the source names only a repository, with no path inside it, the page gives no command: there is no correct one to give.

Where each kind ends up. The site also asks you to read the file before you run the command: a skill can pre-approve tools and attach hooks.

The catalogue, in numbers.

It counts only what passed the filter, and it grows with every crawl. These are the site's figures on 24 October 2026.

skills
36,603
agents
20,168
source repositories
26,184
distinct GitHub owners
18,558

The 12 skill categories and how many each holds. Agents use 11 of their own. A repository that publishes twelve skills counts twelve times as skills and once as a repository.

What it won't make up.

When a number doesn't exist, the site says it's missing.

Find your next skill.

Browsing is free. If you publish skills, a public GitHub repository with a real SKILL.md is the whole submission.

Visit the site Write to us

← Back to home