Vivswan/skills：面向 AI 编程智能体的技能合集

A collection of skills for AI coding agents. Skills are packaged instructions and resources that extend agent capabilities.

About This Repository

This repo keeps the collection-style catalog and install flow from vercel-labs/agent-skills, while also keeping each skill folder plugin-ready so MCP servers, hooks, or app integrations can be added later without changing the layout.

The root .claude-plugin/marketplace.json publishes the catalog as the vivswan-skills plugin for Claude Code marketplace installs, with .claude-plugin/plugin.json as the plugin manifest, plus a xeno plugin that groups the skills vendored from other repositories under their own heading, but the main experience stays centered on npx skills add ....

Available Skills

Automatic

The agent applies these on its own when the task matches:

/code-standards - House standards for maintainable code, folded into reviews
/codex-browser-fix - Codex only: keeps the bundled Chrome plugin driving Brave and other Chromium browsers through app updates and custom model providers
/craft-skills-and-memories - Create and repair skills and memories at their canonical source
/never-twice - Climb every repeated failure to the most durable fix reachable
/no-invalid-states - Invariants in the type system instead of runtime checks
/pr-and-issue-discipline - PRs and issues that show the change instead of describing it
/pr-landing-discipline - Draft flips, review rounds to convergence, line accounting, and a human-gated merge
/reply-and-review-discipline - Plain-first replies to issue reporters and one pending review on someone else's PR
/rubber-duck-review - Cross-model, read-only second-opinion code review
/verify-with-controls - Controls and evidence before a zero, alarm, success claim, or stillness becomes a conclusion
/watch-ci-after-push - Background CI watcher after every push or merge
/working-text - Working text a colleague scans to get work done, from a README or a comment to a PR or a diagram: shape rules, a paragraph-cap and path probe, and an architecture page that cannot rot
/worktree-hygiene - Safe worktree removal, gated retirement of landed branches, explicit handovers, shared-repo rules

Invoked by you

Load only when you invoke them (/skill-name in Claude Code, $skill-name in Codex):

/claude-remote-peers - Message a Claude Code session on another machine over ssh: two socket forwards, no copies, pushed replies
/natural-writing - Prose without AI writing tells
/orchestrator-mode - Parallel worktree subagents with gated landings, direct or PR-based

Xeno

Xeno, from the Greek for foreign: skills written in other repositories, vendored here from their upstream at a pinned commit and refreshed nightly (how):

/unslop - Cut AI tells from any writing (from cursor/plugins)
/frontend-design - Distinctive visual design for new or reshaped UI, invoked by you in Claude Code (from anthropics/skills)
/design-taste-frontend - Anti-slop frontend for landing pages, portfolios, and redesigns, invoked by you in Claude Code (from Leonxlnx/taste-skill)

How the skills reference each other (an arrow means "mentions and hands off to, where installed"):

---
config:
  layout: elk
  elk:
    nodePlacementStrategy: NETWORK_SIMPLEX
---
graph LR
  wt["/working-text"]
  cs["/code-standards"]
  nw["/natural-writing"]
  om["/orchestrator-mode"]
  pid["/pr-and-issue-discipline"]
  rrd["/reply-and-review-discipline"]
  rdr["/rubber-duck-review"]
  pld["/pr-landing-discipline"]
  nis["/no-invalid-states"]
  nt["/never-twice"]
  vwc["/verify-with-controls"]
  csm["/craft-skills-and-memories"]
  wca["/watch-ci-after-push"]
  wh["/worktree-hygiene"]
  us["/unslop"]
  cbf["/codex-browser-fix"]
  crp["/claude-remote-peers"]
  fd["/frontend-design"]
  dtf["/design-taste-frontend"]

  cs <--> wt
  cs <--> rdr
  csm <--> nt
  nis <--> rdr
  nt <--> rdr
  pid <--> pld
  pid <--> rrd
  rdr <--> rrd
  rdr <--> vwc

  wt --> rdr
  wt --> pid
  nw --> wt
  om --> rdr
  om --> wca
  om --> cs
  om --> pid
  om --> pld
  om --> wh
  rrd --> wt
  pld --> wca
  pld --> rdr
  cs --> nt
  cs --> nis
  nt --> nis
  om --> vwc
  csm --> nw
  csm --> wt
  om --> wt
  wt --> us
  pid --> us
  rrd --> us

  click wt "./skills/working-text/SKILL.md"
  click cs "./skills/code-standards/SKILL.md"
  click nw "./skills/natural-writing/SKILL.md"
  click om "./skills/orchestrator-mode/SKILL.md"
  click pid "./skills/pr-and-issue-discipline/SKILL.md"
  click rrd "./skills/reply-and-review-discipline/SKILL.md"
  click rdr "./skills/rubber-duck-review/SKILL.md"
  click pld "./skills/pr-landing-discipline/SKILL.md"
  click nis "./skills/no-invalid-states/SKILL.md"
  click nt "./skills/never-twice/SKILL.md"
  click vwc "./skills/verify-with-controls/SKILL.md"
  click csm "./skills/craft-skills-and-memories/SKILL.md"
  click wca "./skills/watch-ci-after-push/SKILL.md"
  click wh "./skills/worktree-hygiene/SKILL.md"
  click us "./xeno/unslop/SKILL.md"
  click cbf "./skills/codex-browser-fix/SKILL.md"
  click crp "./skills/claude-remote-peers/SKILL.md"
  click fd "./xeno/frontend-design/SKILL.md"
  click dtf "./xeno/design-taste-frontend/SKILL.md"

Installation

npx skills add Vivswan/skills -g

Install a specific skill:

npx skills add Vivswan/skills -g --skill rubber-duck-review

Install to a specific agent:

npx skills add Vivswan/skills -g --skill rubber-duck-review -a codex
npx skills add Vivswan/skills -g --skill rubber-duck-review -a claude-code
npx skills add Vivswan/skills -g --skill rubber-duck-review -a github-copilot

Install directly from the skill folder:

npx skills add https://github.com/Vivswan/skills/tree/main/skills/rubber-duck-review -g

Or install everything as a Claude Code plugin:

/plugin marketplace add Vivswan/skills
/plugin install vivswan-skills@vivswan-skills
/plugin install xeno@vivswan-skills

Usage

Once installed, a skill fires on its own when its trigger matches the task.

Skills marked explicit-invocation-only (/claude-remote-peers, /natural-writing, /orchestrator-mode, /frontend-design, /design-taste-frontend) load only when you invoke them. For the two xeno copies that holds in Claude Code; Codex reads the upstream policy, which lets the model invoke them.

Examples:

Rubber duck this patch before we merge it

Get a second opinion on this refactor

Compatibility

The repo is designed around a shared skill core plus optional agent-specific plugin metadata.

Capability	Claude Code	Codex	GitHub Copilot
Install with `npx skills add Vivswan/skills`	Yes	Yes	Yes
Plain `SKILL.md` workflow	Yes	Yes	Yes
Root `.claude-plugin/marketplace.json`	Yes, optional enhancement	Not primary path	Not primary path
Per-skill `.codex-plugin/plugin.json`	Not primary path	Yes, optional enhancement	Not primary path
MCP via per-skill config	Yes, if the plugin host supports it	Yes, if the plugin host supports it	Keep a skill-only fallback

Compatibility rules for every skill in this repo:

npx skills add ... is the primary install path across agents.
SKILL.md must contain a complete fallback workflow even if plugin or MCP integrations are added.
Plugin manifests are additive enhancements, not a requirement for using the skill.
GitHub Copilot compatibility should assume skill-first installation rather than plugin-specific behavior.

Validation

The pre-commit hook runs the targeted checks on every commit; run them by hand with:

bun run check:staged

That runs the static checks listed in scripts/check-staged.mts, then only the tests a staged file reaches, by import or by name in the test's text.

CI runs the full bun run check (its steps are the check script in package.json) through .github/workflows/checks.yml, plus an end-to-end test that the real npx skills CLI discovers and groups every skill (scripts/cli-discovery-test.ts).

License

Individual and Small Organization License; the version and terms are in LICENSE.md.