The center of the cloud has not moved. “Skill” (1,215) leads again, ahead of “agent” (1,148) and “claude” (790). Those three are the standing grammar of repos built to be installed into, wrapped around, or remembered by coding agents. The interesting movement this week is one ring out, where the words stop describing what a repo does and start describing how well it claims to do it.
The standout there is “design” at 365, sitting eighth overall, above project tokens like “container” (339), “mimo” (330), and “headroom” (324). Around it are “engineering” (234), “taste” (212), “improve” (176), and “understand” (171). Last week the second ring was document plumbing: markitdown, ppt, notebook. This week it reads like a performance review. These are the qualities a repo wants to be associated with, attached to skills that promise good taste or a clearer plan rather than a new capability.
A smaller signal is more telling because it is human. “Karpathy” (144) and “andrej” (139) both made the cloud, and they come from the same repo, so discount them the way you discount any single project’s tokens, the way “cl4r1t4s” and “ponytail” get discounted here. Even discounted, the naming choice is new. “Hermes” and “odysseus” are myths borrowed for flavor; a living engineer’s name used as a skill label is a different reflex, closer to citing an authority than picking a codename. Set it beside “ponytail” (390) and “taste” (212), where the name points at an attitude, and the proper-noun layer has widened past project codenames and mythology.
“Claude” at 790 keeps doing something unusual for a vendor word. It outranks every other platform token by a wide margin: “codex” (215), “mcp” (199), “harness” (144), and “opencode” (126) trail well behind. The other vendors barely register, with “deepseek” down at 122. When one company’s product name behaves like a common noun in repo titles, it is because that product is the default thing these skills are written for. The platform vocabulary exists, but it points almost entirely at one host.
“Open” (499) is still large at fifth, and still worth a caveat. It is the oldest naming reflex on GitHub and it attaches to almost anything, so a high count says less about a trend than the agent and quality words do. It is doing the same job it did last week, pointing at open alternatives without telling you much about what those alternatives are.
The absence is the same shape as before, with a new wrinkle. Model and framework names still do not define the cloud: “python” sits at 137, there is no react, rust, or kubernetes in sight, and “llm” (299) is generic filler. The new wrinkle is below that. The verbs of building are quiet too, with “build” at 112, “ops” at 110, and “scratch” at 163, while the adjectives of quality sit higher. The vocabulary this week grades the work more than it describes it.
Two caveats hold, as always. This is tokenized from repository names, so it measures naming and not a curated taxonomy, and a single hot repo inflates its own words, which is why “cl4r1t4s” (524) and “ponytail” (390) appear at all. Even with that discount, the shift in naming is legible. For a few weeks the words named what a repo plugs into. This week they started naming how well it claims to do the job, reaching for the vocabulary of judgment instead of function.