The top 3 GitHub repos by stars aren't code — they're curricula

What happened

GitHub's all-time star leaderboard, as of this week, has a strange property: the three most-starred repositories aren't software. They're reading lists. codecrafters-io/build-your-own-x sits at 519,526 stars. sindresorhus/awesome — the original 'awesome list' — has 444,000. freeCodeCamp's curriculum repo is at 437,900. Combined, that's roughly 1.4 million stars of curated learning material, and zero lines of production-runtime code.

None of these are new. sindresorhus/awesome was created in 2014. freeCodeCamp's curriculum repo predates most of the JavaScript frameworks people learn from it. build-your-own-x has been the canonical 'recreate Redis, Git, Docker from scratch' index for years. Yet they keep accreting stars at a rate that puts them ahead of every framework, every language, every database, every AI tool. React (240k), Vue (210k), TensorFlow (190k) — not close.

The naive read is that this is leaderboard inertia: stars don't decay, and the oldest viral repos win by default. That's true but incomplete. The interesting question isn't why these repos have so many stars — it's why developers in 2026 are still actively starring them. build-your-own-x's star velocity has actually climbed, not flattened, since the GPT-4 era began.

Why it matters

There's a thesis you hear a lot in 2026: LLMs have devalued curated human knowledge. Why bookmark an awesome-list when you can ask Claude or Gemini to generate a tailored list of, say, 'production-grade Rust async runtimes ranked by community trust' in two seconds? The leaderboard suggests the opposite. The star data implies developers are increasingly willing to defer to a known human curator — sindresorhus, freeCodeCamp's editors, the build-your-own-x maintainers — precisely because LLM-generated lists are now everywhere and indistinguishable in quality from each other.

This is a classic trust-anchor problem. When the marginal cost of generating 'a list of N things' collapses to near zero, the value migrates upstream: which list do I trust? A human curator with a public reputation, a commit history, and skin in the game beats a model that hallucinates a plausible-sounding library that doesn't exist. The asymmetry isn't quality of synthesis; it's accountability.

build-your-own-x is the more interesting signal. It's not a list of tools — it's a list of *re-implementations*. 'Build your own Redis.' 'Build your own Git.' 'Build your own React.' The premise is that you cannot truly understand a system until you've written a toy version of it. In the pre-LLM era, this was a niche pedagogical stance. In the LLM era, it's becoming the defensive moat for actual engineering competence. When an LLM can produce a working Express endpoint in three seconds, the engineers who can debug, refactor, and operate that endpoint at 3am are the ones who've actually built one from a socket up.

The HN and Reddit comment culture around these repos reinforces this. Threads about build-your-own-x in the last year are dominated by variations of: 'I'm doing the Redis one because Copilot has made me realize I have no idea what's actually happening when I call SET.' That's not a learning preference. That's a calibration crisis.

What this means for your stack

Three practical takeaways. First, if you're hiring senior engineers in 2026, the 'have you built X from scratch' question has gotten sharper, not duller. The candidate who has written a toy raft implementation can reason about your distributed system. The one who has only consumed managed services and LLM-generated scaffolding often can't. The leaderboard suggests the market is internalizing this.

Second, if you maintain internal documentation, the awesome-list pattern still works — but only if you sign it. Anonymous internal wikis read like LLM output and get treated as such (ignored). A curated 'Here are the four databases we evaluated and why we picked Postgres, with the dissent footnoted' page authored by a named senior engineer is now disproportionately more valuable than the same content auto-generated. The signature is the value.

Third, if you're a developer relations or education team, freeCodeCamp's continued growth is a tell that structured, opinionated curricula still beat the 'just ask the AI' model for genuine onboarding. The fastest learners in 2026 aren't the ones with the most LLM access — they're the ones who use the LLM as a tutor against a structured curriculum that someone else designed.

Looking ahead

The interesting frontier isn't whether human-curated lists survive — the star data says they will. It's whether the next generation of curators understands that their job has changed. They're no longer summarizing what exists; they're providing the trust signal that the LLM cannot. Expect to see 'awesome-X' lists shift toward more opinionated, more dissent-tolerant, more explicitly-author-branded formats. The unsigned, scrape-and-aggregate awesome-list is dead. The signed, opinionated, occasionally wrong one just hit 444k stars.