convez.ai
SkillDesenvolvimento WebIniciante
RayDataCo

Ray Plugins — Marketplace de Exemplo Claude Code

RayDataCo/ray-plugins

Template educacional para marketplaces de plugins Claude Code

Ver no GitHub
InicialPluginScore · de 0 a 100

Template mínimo e funcional para criar um marketplace de plugins Claude Code. Inclui exemplos de skills básicas (/hello, greeting) e estrutura completa para integração com o ecossistema Ray Data Co, servindo como ponto de partida para desenvolvedores que querem construir plugins próprios.

Instalação

$npx convez-ai RayDataCo/ray-plugins

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
37%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
3

Casos de uso

  • Aprender a estrutura de um plugin Claude Code
  • Criar um marketplace de plugins corporativo
  • Desenvolver skills reutilizáveis em projetos Ray Data

Veredito

Repositório útil como referência educacional para entender a arquitetura de marketplaces Claude Code, mas com conteúdo limitado em produção real. Melhor aproveitado por desenvolvedores que estudam a estrutura antes de construir plugins complexos.

README

A tradução para o português deste README ainda está na fila.

ray-plugins — Ray Data Co's example Claude Code plugin marketplace

A minimal, working example of a Claude Code plugin marketplace. Add it and install the example plugin in two commands.

Use it

/plugin marketplace add RayDataCo/ray-plugins
/plugin install rdco-example@ray-plugins

Then say hello to it, or ask for a Ray Data Co greeting — both are skills.

What's inside

.claude-plugin/marketplace.json      # marketplace manifest (lists plugins)
plugins/rdco-example/                 # minimal copy-me template (skills-only)
  .claude-plugin/plugin.json          # plugin manifest
  skills/rdco-greeting/SKILL.md       # a skill (explains plugin structure)
  skills/rdco-hello/SKILL.md          # a skill (say hello, by name)
plugins/copilot-studio-agent-builder/ # real plugin: build Microsoft Copilot Studio agents (skills-only)
  .claude-plugin/plugin.json
  skills/copilot-studio-agent-builder/  # lifecycle orchestrator (plan->build->test->publish->maintain)
    SKILL.md
    reference/                        # generic runbook (living) + verified facts + doc-endpoint index
    templates/                        # the 4 per-agent deliverable templates (#2-#5)
  skills/copilot-studio-pull-runbook/   # skill: pull the runbook into the current project
  skills/copilot-studio-ground-runbook/ # skill: re-ground the runbook in the latest MS docs
plugins/ab-skill-factory/          # real plugin: the agent brigade that manufactures skills
  .claude-plugin/plugin.json
  README.md + DESIGN.md + BUNDLE-SPEC.md + RAIL-SPEC.md  # pattern docs + interface specs
  skills/expo/                        # the deciding role at the pass (routing + exit set)
  skills/execution-eval-station/      # lift-over-base-model value gate (per-fixture, per-tier)
  workflow/*.run.js                   # reference workflow runs (as-executed)
  examples/                           # worked examples w/ real run reports (variance-analysis, generate-tests)
plugins/ab-managerial-accounting/     # discipline brigade: mise + expo + service over finance stations
  .claude-plugin/plugin.json
  MENU.md                             # the station roster (situation -> station / base-model-covered)
  skills/mise/                        # readiness gate (mise.py engine + mise.toml)
  skills/expo/                        # the composing coordinator (decompose -> select -> compose)
  skills/service/                     # the mise-gated on/off switch
  skills/variance-analysis/           # + 4 more eval-proven finance stations
  base-model-covered/                 # tasks the base model covers (exemplar prompts + evidence)
plugins/ab-data-engineering/          # discipline brigade: mise + expo + service over DE stations
  .claude-plugin/plugin.json
  MENU.md
  skills/mise/ + skills/expo/ + skills/service/   # the standard brigade surface
  skills/pipeline-failure-triage/     # first eval-proven DE station
plugins/ab-legal/                     # discipline brigade: structural (never advisory) legal review
  MENU.md + skills/mise|expo|service/ # the standard brigade surface
  skills/regulatory-notice-coverage/ + skills/contract-clause-coverage/   # eval-proven stations
  evals/                              # two-arm execution-eval evidence
plugins/ab-domain-research/           # FILL brigade: stocks a cellar with domain competency + gold exemplars
  MENU.md + skills/mise|expo|service/ # the standard surface (mise incl. CELLAR_ROOT check)
  skills/cert-body-sourcing/ + standards-regulatory-sourcing/ + academic-ocw-sourcing/ + public-filings-exemplars/ + award-case-study-exemplars/
  evals/                              # two-arm execution-eval evidence (both rounds)
plugins/ab-marketing/                 # discipline brigade: first exemplar-graded generative station
  MENU.md + skills/mise|expo|service/ + skills/effectiveness-narrative-draft/ + evals/

Plugins

  • rdco-example — minimal template: two example skills (rdco-greeting + rdco-hello), skills-only. Copy it to start your own.
  • copilot-studio-agent-builder — guides the full lifecycle of building a Microsoft Copilot Studio agent and emits five standard deliverables (generic runbook + per-agent implementation / evaluation / documentation / maintenance docs). Generative-orchestration baseline; Work-IQ-off + grounded-internal defaults; built from a verified Microsoft Learn walkthrough. Works in Claude Code (deliverables as markdown files) and Claude Web/Desktop (deliverables as project assets). Skills-only (no slash commands) — ships two companion skills: copilot-studio-pull-runbook (pull the runbook into your project) and copilot-studio-ground-runbook (refresh it against the latest Microsoft docs — the field moves fast).
    /plugin install copilot-studio-agent-builder@ray-plugins
    
  • ab-skill-factory — the agent brigade that manufactures Claude skills: stations (spec → tests → author → critic) coordinated at the pass by an expo that decides routing via a closed exit set (advance · refire-to-author · reroute-to-spec · kill), pulling mutable tickets off a rail. Includes the execution-eval-station (measures a produced skill's lift over the base model, per-fixture and per-tier), a deterministic skill-lint gate, full pattern docs (README/DESIGN/BUNDLE-SPEC/RAIL-SPEC), reference workflow runs, and two worked examples with real measured-lift reports.
  • ab-registrar — the first house role (a coordinator/record-keeper, not a kitchen brigade): ships mise + steward (intake, context curation, close-out sweep — moved here from the factory), registry (menu catalog), and orders (open-orders report). No service/fire/stations — a house role never works a ticket.
  • ab-managerial-accounting — managerial-accounting discipline brigade (né discipline-skills): the standard brigade surface — mise (readiness gate) + expo (the composing coordinator that decomposes a request, fires the finance stations it needs, and synthesizes one answer, incl. compound asks like a full due-diligence picture) + service (mise-gated on/off) — over 5 eval-proven finance stations (variance-analysis, annual-budget-build, close-management, treasury-liquidity-analysis, debt-schedule), plus MENU.md (the station roster) and the base-model-covered/ registry (5 tasks with eval-verified base-model coverage). Only eval-passers ship.
  • ab-data-engineering — data-engineering discipline brigade: the same mise + expo + service surface over eval-proven DE stations. First station pipeline-failure-triage (+1.00 lift on silent zero-row failure classification); MENU.md carries honest per-task status for the rest (held-for-refire / weak-evidence / base-model-covered-pending). Only eval-passers ship.
  • ab-legal — legal-review discipline brigade, structural not advisory: the mise + expo + service surface over eval-proven document-checking stations. Two live — regulatory-notice-coverage (GDPR/CCPA disclosure-element coverage) and contract-clause-coverage (standard clause-set coverage by contract type, +1.00 lift) — plus redline-against-playbook held-for-refire. Every station checks a document against an explicit enumerated standard and reports gaps; it never gives legal advice.
  • ab-domain-research — the fill brigade (kitchen kind, domain-centric analog of company research): stocks a knowledge cellar with domain competency and gold exemplars — the two things a generative-skill build needs, since generative skills have no computable oracle and get graded against exemplars. Standard surface (mise incl. a CELLAR_ROOT landing-target check, composing expo, service) over 6 eval-proven stations: cert-body-sourcing (+0.60), standards-regulatory-sourcing (+1.00), academic-ocw-sourcing (+0.60, CC-licensed-only gate), public-filings-exemplars (+0.40–0.60 — EDGAR exhibits/MD&A, SAM.gov RFP/SOWs, court briefs, patents, GAO/CRS, each exemplar with a why_gold), award-case-study-exemplars (+0.60 — creative-domain gold exemplars from juried award libraries), and freshness-watch (99% per-element — re-verifies landed grounding against PRIMARY sources on a cadence and routes factory refire tickets: EVAL-SPEC's world-change regression trigger, live). Every station enforces license-gate-before-content, authority tiering, per-source dispositions with reason classes, provenance frontmatter, and untrusted-data discipline (embedded instructions ⇒ injection-suspect). Curation, never bulk scraping; eval evidence in evals/.
  • ab-marketing — marketing discipline brigade: the mise + expo + service surface over eval-proven marketing stations. Three live generative stations — the marketplace's first exemplar-graded generative skills (eval rubrics derived from real GAO campaign-evaluation + S-1 positioning gold sourced by ab-domain-research the same morning): effectiveness-narrative-draft (+1.00), positioning-narrative-draft (+0.80/+1.00 across rounds; round-1's critic-caught license violation fixed and re-verified), marketing-brief-draft (+0.80 both rounds). Plus advertising-claims-endorsement-review (structural FTC 255 coverage review, base 33%→93% per-element under the founder-ruled measure-skill-quality bar). MENU.md holds an honest 4-round evidence trail for the 2 still-held structural stations. Structural stations report elements, never legal advice. Only eval-passers ship.

Copy it

plugins/rdco-example/ is a template. Copy the folder, rename it, edit plugin.json, add your own skills/, agents/, hooks, or MCP servers, then add an entry to .claude-plugin/marketplace.json. (RDCO standardizes on skills over slash commands — build capabilities as skills.)

Other sources

  • Private GitHub repos work — Claude Code uses your existing gh/git credentials (anyone with repo access can add it). For background auto-updates, set GITHUB_TOKEN.
  • Bitbucket / GitLab / self-hosted work via a full git-URL source, e.g. /plugin marketplace add https://bitbucket.org/team/plugins.git.

Built by Ray Data Co.

Ver o original no GitHub

Bom para

Relacionados