There's a V2EX thread I keep coming back to, from a dev who spent years with a simple weekend ritual: open GitHub Trending, scan the overall board, then flip through the per-language cuts for Python, Rust, Go, and see what's new. His complaint now is blunt. The top of the board is wall-to-wall AI projects, mostly agent frameworks and "skill" bundles, and he doesn't believe anyone actually tries them all. His read is that AI repos farm stars easily, so they float to the top while real projects sink. Give it another year and the ranking is worthless.
Another commenter in the same thread nailed the deeper problem: you used to be able to judge a project by README quality. Now the READMEs and UI screenshots look polished, but a lot of the features are unimplemented or plain wrong. The surface signal you relied on has been inverted. What used to correlate with quality now correlates with marketing effort.
I'm not here to fix GitHub's ranking. I can't, and neither can you. What I can do is stop treating a public leaderboard as a filter and build a private one that reflects signals I actually trust. That's the debrief I want to share.
The insight worth acting on
The shared signal across the whole complaint is simple: star count is a popularity metric, and popularity has been captured. If you keep sorting by it, you're sorting by whoever games it best. The fix isn't a smarter scraper. It's picking signals that are harder to fake and cheaper to verify, then computing your own score.
Which signals?
The thread hands you a few directly. Commit recency, because a repo that hasn't moved in months is a different bet than one that shipped yesterday. Real issue engagement, meaning actual back-and-forth on problems rather than a wall of stars. Documentation depth that reflects a working thing, not a generated pitch. One commenter added a practical tell: PRs, commit history, and directory structure are still harder to fake than a README. Those are the inputs. None of them are perfect, but together they resist the exact gaming pattern people are angry about.
What I'd build with VicroCode
Three pieces, all inside one platform, no local server to babysit.
First, a scorer. I paste in the metadata I've pulled for a batch of repos (name, last commit date, open and closed issue counts, comment activity, whether docs exist beyond a landing README) and get back a score per my own weights. Because I can run Python online, the scoring logic lives as a small script I can rewrite whenever my thinking changes. Want to punish repos with zero closed issues?
Two-line change. Want to decay the recency weight over 90 days? Same. The point is that the formula is mine and it's editable, not a black box someone else tuned to make their own project look good.
A boundary worth stating plainly: VicroCode doesn't crawl GitHub for you. There's no confirmed capability here for live scraping or GitHub API ingestion, so the metadata comes in by paste or by a batch you assemble yourself. That's the honest trade-off. The value is in the scoring and the record-keeping, not in automated harvesting.
Second, a ledger. Scores are a starting point, not a verdict. I want to override them. Maybe a repo scores low on issue activity but I know the maintainer just does clean work quietly, or maybe a high scorer smells like slop the moment I read the code. A SQLite editor lets me keep the computed score in one column and my own annotated verdict in another, editing rows by hand as I learn more. Over weeks this becomes the thing GitHub Trending used to be for me: a memory of what I've already judged, so I never re-evaluate the same over-hyped repo twice.
Third, a board I can actually read. The scorer writes ranked results, the ledger holds my corrections, and a small hosted page shows the merged view sorted my way. With web app hosting for the HTML front end, the ranking lives at a URL I can open on any device on a Sunday morning, same ritual as before, different source of truth.
Why this holds up better than a leaderboard
The reason this survives the slop wave is that every input is one I chose and can inspect. If AI repos start gaming closed-issue counts next, I change the weight and re-run. If I notice my recency signal rewarding trivial commits, I annotate the offenders in the ledger and adjust. A public ranking can't do that for you because it has to serve everyone, which means it serves whoever optimizes hardest.
I want to be clear about what I haven't proven. I have no benchmark showing this catches good projects better than Trending does, no accuracy figure, no user count. That's all unverified, and honestly the value is personal anyway. The claim I'll stand behind is narrower: when a shared metric gets captured, the move is to compute your own signals, keep an editable record of your judgments, and stop outsourcing taste to a board that no longer earns your trust.
Start tiny. A dozen repos, three signals, a formula you can defend. Grow the weights as you catch the ways they fail. The ledger does the rest, quietly, one corrected verdict at a time.