best-of

A directory of directories: 22 “best-of” lists of open-source projects, grouped into 10 categories, 31K combined stars. 1,837 stars, 150 forks, CC0-1.0, repository created 2020-12-17 and last updated 2026-08-10 (all figures read 2026-08-10). The lists themselves live in other people’s repositories — ml-tooling/best-of-ml-python, lukasmasuch/best-of-react, fkromer/best-of-vue, lyz-code/best-of-digital-gardens and eighteen more — and this repository is the index over them.

The entries are not written by hand. Each list is generated from a projects.yaml file by best-of-generator, a separate MIT-licensed CLI tool (110★) that pulls metadata from GitHub and from PyPI, npm, Conda, Maven, Cargo, Go and Docker Hub, computes a score per project, and writes the markdown. A GitHub Action re-runs it on a schedule, so a list is a rebuilt artifact rather than an edited document.

Why this is here: the ranked answer to the star-ranked catalogue

openalternative ranks 500+ projects by GitHub stars and nothing else, and this spoke’s synthesis uses that as its example of a discovery frame that underweights load-bearing software. best-of is the same genre making the opposite bet. Its project-quality-score adds licence, release cadence, semantic versioning, project age, contributor and commit counts, closed issues, downloads, dependent packages and recent-commit activity, and stars enter only as log(stars)/2 − 1 — one logarithmic term among a dozen.

This is the first source in the corpus that treats maintenance as measurable rather than assumed. It also carries its own decay markers on the list entries: 🐣 new, 💤 six months without activity, 💀 twelve months without activity, 📈📉 trending on the score itself.

What the markers say about the index

Those markers are turned on the lists themselves, and the result is unflattering. Of the 22 entries, 12 carry 💀 and 5 carry 💤 — seventeen of twenty-two are, by this repository’s own definition, inactive or dead. The one list nearly everyone arrives through, best-of-ml-python, is among the live ones; most of the long tail is not.

So the machine built to detect abandonment documents its own. That is a point in the method’s favour and against the corpus it indexes: a hand-curated awesome list would simply have gone quiet, and this one prints the date it stopped. Automation makes the staleness legible without doing anything about it — a generator that runs on a schedule keeps producing a page whether or not anyone is still choosing what goes on it.

Provenance

T1 as a primary artifact — the repository is the thing. The counts above are snapshots, and the 💀/💤 markers are computed at generation time, so they age between runs like everything else here.

Connections

The methodological counterpart to openalternative in the open-source-alternative frame: same job, incompatible answers to what “best” means. Cross-spoke, the underlying tool is a piece of build-and-publish machinery closer to ../dev-tooling-wiki’s subject than to this spoke’s, and the lists it generates for curricula-adjacent topics sit beside ../engineering-education-wiki’s awesome-list corpus (awesome-system-design-resources) — noted rather than duplicated. The scoring question itself is paged as project-quality-score.