Every Agent Skill, recompiled as a multi-agent Workflow.
A browsable, editorial catalog of the most-installed Claude Agent Skills from
skills.sh — each one automatically compiled into a
parallel, self-verifying Workflow by skill-to-workflow.
A Claude Agent Skill teaches one agent a procedure in a single context, one step at a time. A Workflow orchestrates many agents — fanning independent steps out in parallel, adversarially verifying every finding, and passing typed data between stages.
The parallelism is already sitting in a skill's prose ("check A, B and C, then verify each") — it's just flattened, because a single agent can't fan out. This catalog runs the most popular skills through skill-to-workflow and shows you the result: for each skill, exactly what was parallelised and verified, with the generated workflow source inline.
→ workflows-directory.pages.dev — hosted on Cloudflare Pages.
Search, filter by topic or status, and click any entry to read its conversion report and the generated workflow.
workflows-directory/
├── scripts/
│ ├── fetch-skills.py # 1. scrape the top-N skills from skills.sh (sitemap + GitHub)
│ ├── convert-batch.js # 2. a Workflow that compiles each skill → a workflow
│ └── build-index.py # 3. merge everything → web/public/directory.json
├── data/
│ ├── catalog.json # the ranked skill list (popularity order)
│ ├── skills/<id>.json # each fetched skill (SKILL.md + bundled manifest)
│ └── workflows/<id>.* # each generated workflow + a .meta.json
├── web/public/ # the static site (HTML/CSS/JS, zero build step)
└── wrangler.jsonc # Cloudflare Pages config
python3 scripts/fetch-skills.py --limit 1000 # 1. fetch top-1000 skills
# 2. convert: run scripts/convert-batch.js with the Claude Code Workflow tool
python3 scripts/build-index.py # 3. (re)build the dataset the site loads
npm run deploy # 4. push to Cloudflare PagesDeploy uses the claude-pages-deploy Cloudflare token via cfman:
export CLOUDFLARE_ACCOUNT_ID=41ebb4f61f901310c3acf958bbf08e4d
cfman wrangler --account claude-pages-deploy pages deploy web/public \
--project-name workflows-directory --branch main --commit-dirty=trueA deliberately non-AI-slop aesthetic: a technical-editorial reference catalog. Warm paper ground, ink type, Fraunces serif display, Hanken Grotesk text, JetBrains Mono for technical metadata, hairline rules, and a single cobalt accent — no dark-mode gradients, no purple glow.
- This indexes the top ~1000 skills by popularity, not all of skills.sh (which indexes hundreds of thousands). The popular head is what people actually use; converting the entire long tail is neither useful nor affordable.
- Generated workflows are machine-compiled and parse-validated, not hand-audited. Treat them as high-quality starting points.
- Independent project. Skill metadata via skills.sh; not affiliated with or endorsed by Anthropic or Vercel.
MIT © democra.ai
