We Checked Four Of These Star Counts. Three Were Too Low.

Thirty-two things to install, in four columns of eight: Plugins, Skills, MCP, Dev Tools. Each entry has a name, a six-word description, and – on some of them – a star count. The footer reads: “Star counts and features change often. Check each official page before you install.”

That star count is new. No other card in this series has attached a number to its entries, and a number changes how a list is read: it stops being a menu and starts being a ranking. If you are deciding what to install first, you will use it.

So we did what the footer asks. We opened four of them on GitHub.

We expected to find inflation – big numbers on small projects is the usual failure. We found the opposite, three times out of four.

superpowers. The card prints 192k. GitHub shows 297.0k today. Understated by about 105,000.

gstack. The card prints 82.7k. GitHub shows 136k. Understated by about 53,000.

humanizer. The card prints 2.9k. GitHub shows 55.3k. That is nearly nineteen times the printed figure.

The fourth, frontend-design, is printed at 277k and we could not find a matching repository at anything like that size – the closest Claude frontend-design skills we found sit in the hundreds or low thousands. We are not claiming the number is wrong; we are saying we could not identify what it refers to, which is its own kind of problem when the number is the thing you are deciding on.

The four we opened

Every figure in the third column was read directly off the repository page on 10 October 2026.

Diagram: printed versus actual

Look at row three and then at row one. The error on superpowers is bigger in absolute terms – a hundred thousand stars – and it costs you nothing, because superpowers is the headline item either way. The error on humanizer is smaller in absolute terms and changes the decision completely: the card ranks it as the least-starred thing in its column, and it is in fact one of the most-starred items on the whole card.

A stale number does not fail evenly. It fails hardest on the entries near the bottom, which are precisely the ones you were about to skip.

All thirty-two, and where the numbers are

Count the badges and something structural appears: ten of the thirty-two have a star count, and all ten are in the left two columns. The right two columns have none at all.

That asymmetry matters more than any individual figure. You cannot compare across the columns, because half the card is scored and half is not. A reader scanning for the biggest number will never choose anything from MCP or Dev Tools – not because those are worse, but because they were never given a score.

Why a stale number drifts one way

Here is the part worth generalising, because it applies to every card you will ever see with a metric on it.

Most of what sits on a tools card ages slowly. A name stays a name. A six-word description – “Reads and posts in channels” – will be true for years. Even a dead product takes months to die, as we found with a card recommending a vector editor that had closed over a year earlier.

A star count ages daily, and in a category this young it only moves in one direction. Claude Code add-ons are months old, not years. Repositories in this space are not drifting up by two per cent a quarter; they are doubling.

So when somebody photographs those numbers and the photograph circulates for a few weeks, the error is systematically downward. Not random. Not inflated. Too small, every time, by an amount proportional to how fast the project is growing – which means the fastest-growing items are the most understated.

That is exactly backwards from how people read an old number. The instinct is “this is probably hype, discount it”. The correct instinct here is “this is probably a floor, check the real one”.

And it produces a specific, nasty failure: you look at a card, see 2.9k next to something, decide it is a minor project, and move on. The thing you just skipped has fifty-five thousand stars. The number did not mislead you into installing something bad. It talked you out of something good, which nobody ever notices, because you cannot miss what you never tried.

What to check instead of the star

None of this means stars are useless – a repository with a hundred thousand of them is genuinely popular. It means the number in a screenshot is the wrong copy of it. Here is what to look at on the actual page, which takes about forty seconds.

Check one is the one to internalise. Stars only go up; commits can stop. A project with 80,000 stars and no commits since March is a monument, not a tool – and the star count is the one number that will never tell you that.

Four honest caveats

We expected the opposite and we were wrong. Going in, our assumption was that the big numbers would turn out to be inflated – that is the normal failure on a card like this. We checked because the claim was numerical, and the data said understated, three times out of four. Worth saying plainly, because the whole argument would have run the other way and it would have been false.

Our figures are already out of date. Every number in the third column was read on 10 October 2026 and is moving upward as you read this. That is not a weakness in the method – it is the entire finding. If you check tomorrow and get a different figure, that proves the point rather than contradicting it.

We checked four of ten, not all of them. Three understated and one we could not match is a small sample, and the direction could be coincidence. We would not have published a percentage off four data points; we are publishing the direction and the mechanism, because the mechanism is sound whatever the sample says.

And the card does not actually lean on the numbers. To be fair: the star counts are small badges in the corner, not the headline. The footer explicitly warns they change. The real content is the four columns and the six-word descriptions, which are accurate and genuinely useful – especially the Dev Tools column, which is the least hyped and most practical set on the card.

The one-line version

This card does something unusual and good: it puts a number next to its recommendations. And that number is the single most perishable thing on the picture – the names will be true next year, the descriptions will be true next year, and the stars were wrong within weeks.

Wrong in a specific direction. Stale star counts in a fast-growing category understate, and they understate most on the fastest-growing items. The card prints humanizer at 2.9k; GitHub says 55.3k. If you used the badges to decide what to install, you would have skipped it – and skipping is the mistake you never find out about.

So read the badges as a floor, not a figure. Then do the forty seconds that actually settles it: open the page, look at the last commit date, see who owns it, read what it asks permission for, and install one at a time.

And notice which columns have no numbers at all. The MCP and Dev Tools columns carry zero star counts and the highest stakes – every MCP entry connects Claude to an account of yours, and one Dev Tools entry hands it a browser.

Learning a lot does not make anyone good. Doing a lot does. Installing thirty-two things is not doing; it is a larger surface area and a worse night’s sleep. Installing one, using it for a week, and keeping it is.

If you want to be held to that, we run a free challenge at BANI Academy. There is no barrier to entry – you commit a small fee that is refunded when you finish, and the only requirement is that you actually do the work for the length of the challenge.

Pick one entry from the card, open its repository, and tell us what the real number is today. Comments are under the original post – and tell us the last commit date too, because that is the one that matters.

More on building a channel for a global audience at mmoyoutube.com.