A Skill Installs the Same Everywhere and Behaves Differently in Each Agent
Saturday edition · October 24, 2026 · 2002 words · 11 min read
Agent skills install with the same command in every harness and do not behave the same way across them, and only the first half has a counter. This week's biggest star climb in GatiFlow's GitHub data is a case study in both halves. According to GatiFlow intelligence, wheresryan22/anatomy, a folder with one instructions file that teaches a coding agent to draw interactive isometric illustrations of machines, rose from 8,938 stars at our first reading on October 14 to 18,460 on October 21, up 106.5 percent over eight tracked days, the largest climb among the repositories followed across that stretch. Growth has cooled from its launch pace without stalling: every reading on October 21 matched or exceeded the one before it. Per explainx.ai, the repository was created on October 8 and had about 2,176 stars on October 11. A spike like this invites a familiar conclusion, that skills are not packages, nothing counts their adoption, and stars are all anyone can see. That conclusion is out of date, and correcting it moves the real gap somewhere more useful.
Anatomy is an Agent Skill, the format Anthropic launched in October 2025 and published as an open standard on December 18, 2025, per The New Stack and AI Business. Vercel then released its open skills CLI and the skills.sh directory on January 20, 2026, per Vercel's changelog, with a leaderboard meant to track installs across the ecosystem. The anatomy README names five agents it runs in: Claude Code, Antigravity, OpenCode, Codex and MiniMax Code. The install surface is therefore not a development of recent weeks. It is nine months old, and what this cycle added is a demonstration of how fast a single repository can travel across it.
The star count is the familiar half of the story, and anatomy's 18,460 of them measure attention. The less familiar half is that installs made through the skills CLI are counted too. Per the skills.sh privacy page, each install event carries the skill, the agent it was installed for, a coarse timestamp and a fingerprint used for hourly deduplication. Anatomy's directory page, about two weeks after the skill first appeared there, shows install and star counts within a few percent of each other. The figure is a floor rather than a census, because users can disable the telemetry and git clones or manual copies, which the README also offers, never reach it. The skill is far smaller than the leaders: first-party skills from Vercel and Anthropic sit near the top of the all-time board with install counts far beyond anatomy's. Stars also turn over quickly: Gathin23/cleanyourtv and deepseek-ai/DeepInfer, both first collected on October 20, already stand at 18,497 and 17,658 stars according to GatiFlow intelligence, close to anatomy's total.
What the counter cannot see is use. Per the skills.sh privacy page, the directory collects no prompts, session content or tool calls, and per its FAQ no usage patterns, so an install is the agent-era equivalent of a package download: an acquisition event that says nothing about whether the skill is ever loaded. Even the early coverage of anatomy drew that line, with explainx.ai describing the first star count as evidence of interest rather than production use.
A public measurement of the underlying gap comes from the directory's own operator. In January evals of a skill that teaches Next.js 16 APIs, Vercel found that the agent never invoked the skill in 56 percent of cases, and the pass rate stayed at the 53 percent baseline. Explicit instructions lifted it to 79 percent, while a documentation index embedded in AGENTS.md reached 100 percent. Vercel adds that skills suit vertical workflows a user triggers on purpose. Anatomy leans that way, since its README leads with a slash command, though explainx.ai notes Claude can also load it when a request merely matches, which is the path where the invocation gap bites. Installed is not used.
Five questions can be asked of any skill, and the public counters answer three of them. Did anyone notice it: stars, which GitHub counts. Did anyone install it: CLI installs, which skills.sh counts and, through a documented API that requires a key, exposes as leaderboard, trending and audit data that can be charted over time. Is it safe to run: audits, which come from Gen Agent Trust Hub, Socket, Snyk and others, with skills that fail every partner audit excluded from the directory. Anatomy's page shows passes from Socket and Snyk and a warning from Gen Agent Trust Hub, and explainx.ai advises reading the code of any skill that executes scripts and drives a headless browser, so a popular skill is not automatically a trusted one.
The remaining two questions have no public counter in the field. Does the agent actually load the skill once it is installed: in production only the harness, whether Claude Code, Codex or Copilot, can see that. Does using it help: publishers answer for themselves, as NVIDIA does. Per its skills repository, every skill in its catalog ships with a signature, a skill card, an evaluation dataset and a benchmark report documenting uplift, and the catalog is syndicated to skills.sh.
Behavior is the half the counters cannot see, and benchmarks are where it shows. On the SkillsBench board that Vals AI runs with BenchFlow AI, last updated September 27, skills lifted the models tested by 14.52 points on average. Yet the same GPT 5.5 model references skill paths about 4.8 times per task on Codex and roughly 1.2 on OpenHands, and scores 62.55 percent on one and 62.21 percent on the other, while Claude Sonnet 4.6 references skills least and still posts the fifth-largest gain, 21.22 points. Differences in harness and model confound both comparisons, but together they say two things. The list of supported agents in a README describes where a skill installs, not how it behaves, and even where installs land is not yet visible: anatomy's skills.sh page shows no breakdown by agent, although each install event records one. And reference counts make a poor KPI, because how often an agent touches a skill is not how much the skill helps.
Polars on PyPI is the control case and shows what a mature counter looks like. GatiFlow's PyPI collector, which reads pypistats.org, logged 14,183,522 downloads for the latest full week, up 10.9 percent from 12,784,507 the week before. The pypistats page shows the same figure, but that agreement comes from a single feed, so it is not independent confirmation. One week at that pace already exceeds half of the more than 24 million monthly downloads Wikipedia reported for September 2025, the month Amsterdam-based Polars raised a Series A of €18 million led by Accel, per TechCrunch. Two cautions apply: a download counts CI runs and reinstalls as well as people, and pypistats lists 2.0.0 as the latest release, so part of one week's gain may be upgrade traffic. The difference with skills is maturity, not kind, and anatomy straddles the line: its finished figures also ship as the npm package anatomy-figures, which the registry shows was first published on October 9.
The research layer is also building outcome scoring for agent-built visual artifacts like anatomy's. GatiFlow's arXiv collector first logged three papers on October 20. BlenderWorld, from the Qwen team, scores agents that build 3D scenes in Blender against 2,651 binary rubric items across 223 tasks. OpenVLHarness reports average gains of 9.7 to 12.6 points over matched no-tool baselines on 22 benchmarks and eight backbones, which is the with-and-without test applied to a tool layer. EMMA organizes long video into events and retrieves evidence with several agents. A paper alone is not evidence of a market shift, but next to Vercel's invocation numbers, NVIDIA's shipped evaluation sets and the Vals board, these show the groups closest to agents investing in scoring outcomes, not just counting installs.
Two readings are tempting this week, and the data supports neither. One treats anatomy as a fad that stars exaggerate. The other treats its star count as proof of adoption. At the time of our check its counted installs matched its stars at roughly one to one, so the star count is not just applause, yet an install is an acquisition event and the skill remains small by install standards, so it is not yet infrastructure. GatiFlow's October 9 deep dive described distribution as scaling faster than any way of checking it. This cycle's evidence refines that: checking exists at the shallow end, with installs counted by Vercel's directory and security audited by outside partners, and it thins out exactly where it matters, in use and outcome, which only harness vendors and benchmark authors can see. Three observations would change this read: a harness vendor publishing aggregate invocation rates, skills.sh publishing installs by agent or adding a retention measure, or anatomy's installs falling well behind its stars as star growth fades.
If you build developer tools or run more than one coding agent, the conversation to have with your team in the next two to four weeks has three parts. First, test each skill you depend on the way SkillsBench does: the same tasks on the same agent with and without the skill, three trials each, in every harness you use, with uplift as the KPI and invocation count as a diagnostic. Second, treat a skill's skills.sh page as an acquisition record: when a skill spikes, ask whether installs are sustained, whether audits are clean and what uplift evidence the publisher can show. Third, if you control a harness, publish aggregate invocation and uplift rates by skill, because only you can see them. For investors, the open position is not another directory, since skills.sh already ranks installs and audits and BenchFlow and Vals already score skills in aggregate; as far as we found, none of them publishes a given skill's effect in the wild under a given agent against the same agent without it.
FORWARD CATALYSTS: GitHub Universe runs October 28 to 29 at Fort Mason Center in San Francisco, per the GitHub blog. The agenda includes a panel with Anthropic and OpenAI speakers on what comes next for AI-assisted development, plus a session on authorizing MCP servers, both close to the question of how agent harnesses load and trust external capabilities. GitHub Copilot is among the agents skills.sh lists as install targets, so any detail on how Copilot discovers, loads or reports on skills would bear directly on the invocation gap described above.
The folder is portable between agents. The evidence so far says the behavior is not.
Sources:
- anatomy: GitHub repository, wheresryan22 (https://github.com/wheresryan22/anatomy)
- anatomy on skills.sh (https://www.skills.sh/wheresryan22/anatomy/anatomy)
- skills.sh FAQ (https://skills.sh/docs/faq)
- skills.sh privacy (https://skills.sh/privacy)
- About skills.sh (https://skills.sh/about)
- skills.sh API documentation (https://www.skills.sh/docs/api)
- The Agent Skills Directory, skills.sh (https://www.skills.sh/)
- Anatomy Claude Code Skill: 2,100+ Stars in 3 Days, explainx.ai (https://explainx.ai/blog/anatomy-claude-code-skill-isometric-interactive-figures-wheresryan22-2026)
- anatomy-figures package metadata, npm registry (https://registry.npmjs.org/anatomy-figures)
- Introducing skills, the open agent skills ecosystem, Vercel changelog (https://vercel.com/changelog/introducing-skills-the-open-agent-skills-ecosystem)
- AGENTS.md outperforms skills in our agent evals, Vercel (https://vercel.com/blog/agents-md-outperforms-skills-in-our-agent-evals)
- Agent Skills: Anthropic's Next Bid to Define AI Standards, The New Stack (https://thenewstack.io/agent-skills-anthropics-next-bid-to-define-ai-standards/)
- Anthropic Launches Skills Open Standard for Claude, AI Business (https://aibusiness.com/foundation-models/anthropic-launches-skills-open-standard-claude)
- GitHub: NVIDIA/skills (https://github.com/nvidia/skills/)
- PyPI Download Stats, polars (https://pypistats.org/packages/polars)
- The startup behind open source tool Polars raises $21M from Accel, TechCrunch (https://techcrunch.com/2025/09/29/the-startup-behind-open-source-tool-polars-raises-21m-from-accel/)
- Polars (software), Wikipedia (https://en.wikipedia.org/wiki/Polars_(software))
- SkillsBench Leaderboard and Methodology, Vals AI (https://www.vals.ai/benchmarks/skillsbench)
- BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References, arXiv (https://arxiv.org/abs/2610.23986)
- OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory, arXiv (https://arxiv.org/abs/2610.26860)
- EMMA: Event-Driven Multimodal Memory System with Multi-Agent Retrieval, arXiv (https://arxiv.org/abs/2610.23892)
- Qwen BlenderWorld benchmark for visual agents in Blender, explainx.ai (https://explainx.ai/blog/qwen-blenderworld-benchmark-visual-agents-3d-blender-2026)
- Your guide to GitHub Universe 2026, GitHub Blog (https://github.blog/news-insights/company-news/your-guide-to-github-universe-2026-is-here-the-schedule-just-launched/)
Disclaimer: This article is generated by GatiFlow Intelligence for informational purposes only. It does not constitute investment advice, recruitment recommendations, or legal guidance. All data is derived from public sources and AI analysis — verify independently before making decisions. Past trends do not guarantee future results.
Where this came from
Every Deep Dive starts from GatiFlow's own pipeline: 13 public developer sources, collected every six hours, with a confidence score and the evidence behind each signal. The same signals, filtered to the topics you follow, are a JSON API.
No credit card required.
Get the next one by email
One article every Saturday morning in your time zone. No account needed, and nothing else is sent to the address.
Double opt-in: you confirm by email first. What we store, and for how long, is in the privacy policy.
Tell me I am wrong
Corrections, the version of this you have lived through, or what you would like covered next. It reaches me directly and is never published. It is kept for two years so it can be read and answered; the privacy policy has the details.
0/2000