The piece was a round-up: three repositories you should have downloaded last week. Before installing anything we opened all three on GitHub. It takes about thirty seconds each.
VoltAgent/awesome-claude-design had 3,443 stars.
The write-up said 1.4k+. OpenCoworkAI/open-codesign had
7,757. The write-up said 2k+. Two and a half times out, and
nearly four times out.
sjsyrek/design-council was quoted no figure at all. It
had 177 stars and was last pushed on 28 May — in a
piece published in August, describing all three as if they had just
landed.
Why understated numbers are worth a post
Being wrong in your favour is the harmless direction. Nobody was talked into installing anything by a star count that was too low.
What it tells you is the useful part. A number that low is a snapshot nobody re-ran. The piece was written once, published, and left alone while the repositories carried on being real projects. Every other claim in it — what the tool does, what it costs, whether the install line still works — is exactly as old as the star count, and most of those you cannot check at a glance.
So the finding is not "the numbers were low". It is that the write-up is marketing copy with a shelf life, and you have just been handed a free way to measure how far past it you are.
What we checked, and it is three fields
Stars, last push date, and licence. All three sit on the repository's front page, and together they answer: is this alive, is anyone using it, and am I allowed to.
On these three, the licence was the reassuring half. All three are MIT, which we verified rather than assumed — a repository with no licence file at all is not permissively licensed, it is unlicensed, and using it in client work is a decision somebody should make on purpose rather than by clicking clone.
What the three actually are
awesome-claude-design is 68 templates that scaffold a
site's design language into your own project. Useful, with an obvious
trap: cloning a well-known product's look is how you end up with a site
indistinguishable from every other site built the same way.
open-codesign is the one we would actually reach for. It
runs locally against a model of your choosing, and the material stays
on your machine. That is the same reason we run what we run, and it
matters more than any feature on the list.
design-council runs up to twelve agents with different
roles, each holding its own context window so opinions do not bleed
into one another, with a lead agent making the final call. That is its
claim rather than our measurement — we did not run it.
The one we did not need
We already work that way. Specialists with separate contexts, a reviewer who is structurally unable to build, and one seat that decides. The part we would genuinely read that repository for is the bit we are weakest at: agents messaging each other directly. Ours still hands off through a person, by hand, every time.
The pattern underneath
This was one of several stale write-ups in the same batch. Another one, covered in its own post in this series, gave an install command for a package that does not exist on the index it names — and the repository's own README repeats it, so it was wrong at the source rather than in somebody's retelling.
None of that is dishonesty. It is what happens to a page nobody revisits.
What to do
Open the repository before you install from a round-up, and read three fields: stars, last push, licence. If any of them disagrees with the write-up, treat everything else it told you as unverified — including the parts you were going to take on trust.