The dashboard's analysis of public GitHub data shows roughly 90% of Claude Code-linked commits land in repos with fewer than 2 stars. By surfacing this data, the tracker provides empirical evidence that AI-assisted coding is overwhelmingly concentrated in low-visibility repositories rather than significant open-source or production projects.
Submitted the finding with the framing that 90% of Claude-linked output goes to repos with fewer than 2 stars, highlighting the gap between the narrative of AI-driven software transformation and the observable reality of where AI-generated code actually ends up.
The editorial argues this finding 'shouldn't be surprising if you think about the distribution of GitHub repositories generally.' The vast majority of all GitHub repos have zero stars — they're homework, scripts, weekend experiments — so Claude Code's usage distribution simply mirrors GitHub's existing long-tail shape rather than revealing something unique about AI coding tools.
The editorial acknowledges the methodology is 'an imperfect lens' since private repositories are entirely invisible and not every Claude-assisted commit carries the co-author tag. This means the most commercially significant usage — enterprise codebases, proprietary products — is systematically excluded from the analysis, potentially making the headline stat misleading about Claude Code's actual production impact.
A community-built analytics dashboard at claudescode.dev has been scraping and analyzing GitHub commits that bear Claude Code's fingerprints — the telltale `Co-Authored-By: Claude` trailers, Claude-specific commit message patterns, and other signals that indicate an AI coding assistant was involved. The dataset covers activity since Claude Code's launch, and the headline number caught the Hacker News crowd's attention: roughly 90% of all Claude Code-linked commits land in repositories with fewer than 2 GitHub stars.
The post hit 221 points on Hacker News, sparking a long thread that went well beyond the surface-level stat. The tracker itself offers a window into the shape of AI-assisted development — not the polished demos Anthropic puts in keynotes, but the messy reality of how developers actually use the tool.
The methodology relies on public GitHub data: scanning for commit attribution patterns, then cross-referencing the target repositories' star counts, languages, and activity levels. It's an imperfect lens — private repos are invisible, and not every Claude-assisted commit carries the co-author tag — but it's the most comprehensive public attempt to measure Claude Code's real-world footprint.
The 90% figure punctures a specific narrative: that AI coding tools are transforming how production software gets built at scale. They might be, eventually. But right now, the observable evidence says the overwhelming majority of AI-generated code lives in repos that nobody else is looking at.
This shouldn't be surprising if you think about the distribution of GitHub repositories generally. The vast majority of repos on GitHub have zero stars. They're personal projects, homework assignments, one-off scripts, job application take-homes, and weekend experiments. GitHub is as much a personal scratch pad as it is a collaboration platform. What this data tells us is that Claude Code's usage distribution mirrors GitHub's existing long tail — it's not disproportionately concentrated in high-visibility projects.
The Hacker News discussion surfaced several important nuances. Some commenters pointed out that the metric conflates "low quality" with "low visibility," which is a category error. A solo developer using Claude Code to build an internal tool for their employer might produce genuinely valuable software that simply lives in a private fork or a low-star repo. The absence of stars doesn't mean the absence of impact.
Others argued the opposite: that the data validates skepticism about AI coding productivity claims. If the tool were genuinely 10x-ing experienced developers on serious projects, you'd expect to see more signal in established codebases. The fact that Claude Code's visible footprint is overwhelmingly in throwaway-tier repos suggests the primary use case is experimentation, not production engineering.
A third camp — arguably the most interesting — noted that this is exactly what you'd expect from a technology in its adoption S-curve. Developers try new tools on low-stakes projects first. The weekend prototype today becomes the internal tool next quarter becomes the production service next year. Measuring a tool's impact by where early adopters first use it is like judging a programming language by what people build in their first week of learning it.
If you're a team lead or engineering manager evaluating AI coding tools, this data offers a useful calibration. The marketing pitch is "your developers will be 3-5x more productive." The observable reality is that most developers are using these tools to spin up new projects from scratch — not to accelerate work on existing, complex codebases.
That's not necessarily a knock on the tools. Greenfield development is where AI assistants genuinely shine: scaffolding a new service, generating boilerplate, wiring up APIs with well-documented interfaces. The hard part — navigating a 200k-line monolith with undocumented invariants and decade-old migration code — is exactly where current AI tools still struggle, and where developers are understandably reluctant to hand over the reins.
For individual developers, the takeaway is more practical: you're not behind. If your Claude Code usage is mostly side projects and prototypes, you're in the 90th percentile of normal. The developers posting polished AI-assisted PRs to popular open source projects are the visible minority, not the baseline.
It's also worth noting what this data can't tell us. Enterprise usage behind corporate firewalls is entirely invisible to this analysis. GitHub Copilot and Claude Code usage within private repositories at companies like Stripe, Shopify, or any Fortune 500 engineering org doesn't show up in public commit data. The 90% stat describes the visible public surface, which may or may not reflect the full picture.
The claudescode.dev dashboard is itself an interesting artifact — a community-built observatory for a phenomenon that the AI companies themselves have every incentive to frame selectively. As AI coding tools mature, independent measurement of their actual impact (not just their claimed impact) becomes increasingly valuable. The 90% number will likely shift over time as developers move from experimentation to integration. The question worth tracking isn't whether AI-assisted code reaches high-star repos — it will — but how long that transition takes and whether the productivity gains survive contact with real-world codebases.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.