Twenty-five skills turn engineering policy into agent instructions
Agent Skills packages 25 workflows for the parts of software work that coding agents often rush. The collection covers requirements interviews, specifications, task planning, incremental implementation, test-driven development, security, performance, code review, documentation, observability, and shipping. Each SKILL.md states when it should fire, lays out a process, names common shortcuts, and ends with evidence requirements. That makes the files useful as editable working agreements, especially when a team is tired of repeating the same guardrails in chat.
The pack also includes 9 lifecycle commands, 4 specialist personas, and 7 shared checklists. Claude Code gets commands and hooks, Gemini CLI gets command wrappers, and Codex reads the shared skills through plugin metadata. Cursor, OpenCode, Windsurf, GitHub Copilot, Antigravity, Kiro, and Command Code have documented routes too. The breadth is real, but each host discovers instructions differently. A successful copy into one folder does not prove that another agent will list, select, or obey the same skill.
Our 1 MB fixture passed one test, not the whole catalog
We cloned commit 2686b62 into an unprivileged Debian container with 3 CPUs and 8 GB of RAM. The checkout held 207 files, roughly 5,542 lines of source, and occupied 1 MB. Our test method followed the provided npm project under evals/fixtures/ci-cd-and-automation/, rather than treating Markdown files as an application. The result proves that a narrow fixture can install and run. It leaves the effect of all 25 skills on an agent's decisions unmeasured.
Our run completed installation in 6 seconds, added 0 packages, and kept disk use at 1 MB. The repository exposed no build script in that measured path, so there was nothing to compile. Node's test runner then passed 1 test in 2 seconds with 0 failures, while npm audit reported 0 known vulnerabilities. One CI workflow and a tests directory were present; no Dockerfile was found. These are clean mechanics for the exact path we ran, with very limited behavioral coverage.
A 6-second install still leaves host-specific setup
The 6-second install is the easy part. For the widest route, the README points to Vercel's npx skills add addyosmani/agent-skills, which can install all 25 skills or a named subset. Claude Code has a marketplace plugin, Gemini CLI accepts the repository's skills path, and Codex has a native plugin route. No separate server, project database, or repository-specific API credential is required to read and use the Markdown workflows.
Those adapters have practical conditions. Codex support calls for CLI v0.122 or newer. Claude Code's marketplace clone may use SSH, so a machine without GitHub keys needs the documented HTTPS URL or Git rewrite. Antigravity has a known wrapper-discovery limitation and directs users to namespaced skills. On Windows, open issue 214 says the Claude Code startup hook breaks when the plugin path contains a space. A team rollout needs one discovery check per supported host. A successful central install does not settle the others.
The 1 MB checkout exposes a single-skill portability gap
Our 1 MB checkout includes shared material outside individual skill folders, and that layout creates the clearest setup trap. The README says a per-skill npx install copies only skills/<name>/, leaving repo-level references/ behind. The selected skill still loads, but links to supplemental checklists can fail. Issue 597 describes exactly that after installing code-review-and-quality, including missing references and links to two other skills that were not present.
A whole-repository integration carries the 7 shared checklists and avoids that particular omission. Copying one skill by hand may require copying its dependencies or moving the needed checklist into a skill-local references directory. This weakens the appeal of selective installation, which is otherwise the safest adoption method. Before approving a small bundle, follow every relative link from each chosen SKILL.md and confirm the target arrived in the agent's actual discovery directory.
Claude Code 2.1.283 missed the owning skill in repeated direct requests
Open issue 620 reports a routing survey pinned to Claude Code 2.1.283 and Claude Opus 5.5. When code or a diff was already pasted into a direct request, the expected skill fired 0 of 6 times for simplification, debugging, commit messages, changelogs, SQL-injection review, performance work, unit tests, and pagination. Process-shaped prompts did better. The reporter labels the samples small, which is fair, but the repeated zeros expose the central adoption risk: installed policy can remain unused.
The proposed trigger wording moved two test-driven-development prompts from 0 of 6 to 6 of 6 in that report. This suggests skill descriptions can improve routing, while also showing why teams must test the descriptions against their own prompts and model versions. Package installation cannot guarantee automatic activation because the model still routes probabilistically. Explicit invocation is safer for release checks, security work, or any workflow where silently skipping the process would matter.
Release 0.6.11 fixed checks that had missed 7 of 14 cases
Release 0.6.11 shipped on September 26, 2026, the same date as the last repository push returned by GitHub. Its notes say a constraint floor guard had silently passed 7 of 14 cases it claimed to catch, including deleted tests and loosened budgets, and that the release fixed those cases. It also tightened validators, added operating-system coverage, and adjusted routing for reviews of pasted diffs and pull requests. That is useful maintenance, plus evidence that these workflows need regression tests like code does.
GitHub showed 99,855 stars and 118 combined open issues and pull requests on September 29, 2026. Current discussion includes routing behavior, model drift, portability, and integration fixes, so the queue is active and relevant. The MIT license is simple, and the documentation is detailed enough to audit before adoption. Still, our measured result was 1 passing test in a fixture. Start with two or three skills, record whether they fire, inspect the resulting work, and keep only the rules that measurably change it.