Google Ships 138 Official Agent Skills: Inside the google/skills Repository

A YouTube short by Jack compresses the story into one line: Google just published its own skills on GitHub for free, and using them puts you ahead of 99 percent of people. The hook oversells it, the primary source behind it is more real and considerably larger than the teaser. At github.com/google/skills, Google maintains a repository of agent skills for its own products. As of today, September 14, 2026: 138 skills according to the generated index, 19,893 stars, 1,599 forks, 337 commits, with the latest push landing the same day. This post reads the repository not as a news blip but as an operating manual: what sits in those folders, how Google manufactures the skills, and what the format means for everyone who operates agents rather than merely using them.

A skill is a folder with a SKILL.md file at its core. The YAML frontmatter requires exactly two fields: name as the identifier and description stating what the skill does and when it applies. No fine-tuning, no API wiring: the agent reads the description, matches it against the task, and loads the rest of the folder on demand. The format originated at Anthropic, the open specification lives at agentskills.io, and Google has adopted it not as a reader but as a supplier. The repository's own description reads: Agent Skills for Google products and technologies. The repo has existed since March 31, 2026, ships under the Apache 2.0 license, and behaves like infrastructure rather than a marketing shelf: 337 commits, plus 21 open issues and pull requests reported through the GitHub API.

The index holds 138 entries, and the list reads like a product catalog. A large block belongs to the Agent Platform: alert-configuration, deploy, endpoint-management, eval-flywheel, inference, model-registry, prompt-management, troubleshooting, plus the skill-registry tooling. Another block covers the data products: BigQuery with a dedicated AI-ML variant and BigFrames, AlloyDB, Bigtable, Cloud SQL, and data-lineage tooling. Then the operations products: Cloud Run, Cloud Build, Cloud Logging and Cloud Monitoring, down to PromQL query generation. On the Gemini side sit the Gemini API skill for the Agent Platform, a managed agents API skill, and the Interactions API skill, alongside Genkit for Dart and Go. One skill exists purely for orientation: finding-google-skills fetches the current index and returns exact entry points so the agent does not hunt through an outdated blog post. The majority of the skills target Google Cloud, the rest spread across development and workflow.

The origin of the commits is the tell. Nearly all of them come from a Copybara account, the newest commit dated September 12, 2026 carries the number PiperOrigin-RevId 980133878, and several commit messages simply say: no public description. Piper is Google's internal version control system, Copybara is the export path to GitHub. The skills are therefore produced in internal pipelines and exported automatically, the same machinery that moves other Google projects into the open. Six contributors are publicly listed, with the pipeline doing the heavy lifting behind them. That tells you the repo is not a hobby folder but a production line. The pace follows the machinery: the cloud-run-alert-configuration skill was onboarded on September 11, 2026, three days before the latest push, and changes land in the catalog without an intermediate step.

The most instructive commit message in the repository is a lesson about lists. On August 28, 2026, Google reworked the hand-maintained plugin lists because both had drifted. A rules file claimed a plugin installs three skills; the plugin actually ships five. The count was correct when written and wrong at the first update, precisely the defect that recurs every time the plugin grows. Two marketplace manifests described the same marketplace, yet the Claude manifest listed seventeen plugins and the agents manifest sixteen; one entry had been maintained in only one of the two files. Google's response is smarter than a one-off correction: instead of fixing the wrong number, the file no longer states a count at all. It tells the agent to check what is actually available rather than trusting the file. A number that is never written down cannot go stale. Three conformance cases were added as automated checks: a rules file that enumerates two or more skills must name all of them; a file that states a count, as digit or word, must state the correct one; and both marketplace manifests must list the same set of plugins.

The second lesson sits in the September 12, 2026 commit. The retrieving-developer-knowledge skill documented a batch fetch as POST /v1/documents:batchGet with a JSON body. That method does not exist, the call returned a 404, and the documented operation had never worked. The v1 discovery document requires a GET with repeated names query parameters and no request body, plus an optional view parameter and a ceiling of 20 documents per call. The fix corrects SKILL.md and the api-fallback.md reference file, and adds the credential guidance both had been missing: they now link the public Developer Knowledge quickstart, which covers enabling the API and creating a key. The lesson for skill authors is sitting in plain sight: a skill is code written as prose, and prose rots just like code. Anyone describing an API should test the description against the live API, otherwise the agent will faithfully reproduce the documented 404 in its own workflow.

For anyone operating agents, the news is not the number 138 but the direction. First: the Anthropic format is no longer one company's product feature, it has become industry infrastructure. Two of the largest vendors ship their skills in the same folder scheme, and the standard at agentskills.io keeps the units portable: written once, loadable in any compatible harness. Second, skills move knowledge out of prompts and into versioned files. A prompt evaporates, a skill lives in the git history, can be shared across a team and reviewed through a pull request. That is exactly the joint where agent operations usually gets wobbly. Third, Google demonstrates what quality assurance looks like at this scale: not through reviews but through conformance cases that automatically test whether documentation and reality have drifted apart. If you write your own skills, you can copy the same chain: fewer claims, more checks.

The path into your own setup runs through the plugin interface. The repo ships two marketplace manifests, one for Claude-based harnesses and one for agent-agnostic tooling, each with one entry per plugin package. The cloud plugin bundles the Agent Platform and cloud skills, individual skills remain separately installable, and going through the package is optional. For searching the catalog there is the finding-google-skills skill: it fetches the generated index with 138 entries and returns the exact paths instead of leaving the agent to browse the repository by hand. Installation itself follows the standard: clone the folder, register the skill, done. If you prefer not to install anything, the SKILL.md files read perfectly well as documentation, which works too.

About the author: I run the HUNTER cyberdeck on a used Google Pixel, without root, with cloud models and terminal orchestration. I treat skills as an operations layer, not decoration. Last fact-checked: September 14, 2026.

Sources: github.com/google/skills (repository data via GitHub API and generated index.json, retrieved September 14, 2026), repository commit history with the commits of August 28, September 11 and September 12, 2026, the agent skills standard at agentskills.io, trigger: Jack on YouTube, short https://www.youtube.com/shorts/J0emMr26CgE (retrieved September 14, 2026).