Skills Workflow
semi-sentient/skills-workflow
Orquestração de skills para agentes com suporte cross-platform
Ver no GitHubConjunto de skills para orquestração de agentes durante a execução em múltiplos harnesses (Claude Code, Cursor, Roo). Fornece pipeline completo de ideia até implementação com skills para grill, PRD, planejamento e execução encadeados.
Instalação
npx convez-ai semi-sentient/skills-workflow1 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
Casos de uso
- Orquestrar fluxos de trabalho entre diferentes agentes de IA
- Executar pipeline idea-to-implementation com PRD e planejamento
- Integrar skills customizadas em projetos com Claude Code, Cursor ou Roo
- Validar planos e designs através de questionamento sistemático (grill)
Veredito

Ferramenta sólida para quem trabalha com múltiplos agentes e precisa de orquestração robusta. Modular via Vercel Skills CLI, mas exige familiaridade com arquitetura de agentes para máximo proveito.
README
A tradução para o português deste README ainda está na fila.
skills-workflow
Agent skills workflow with an emphasis on orchestration during execution that works across Claude Code, Cursor, Roo, and other agent harnesses.
Recommended Workflow
These skills are designed to chain together. See WORKFLOW.md for a guided walkthrough of the full idea-to-implementation workflow using grill-with-docs (or grill-me) → write-a-prd → prd-to-plan → run-plan.
Installation
Requires Node.js. Install specific skills into any project using the Vercel Skills CLI:
npx skills@latest add semi-sentient/skills-workflow --skill <skill-name>
Install multiple skills at once:
npx skills@latest add semi-sentient/skills-workflow \
--skill commit \
--skill grill-me \
--skill grill-with-docs \
--skill prd-to-plan \
--skill run-plan \
--skill tdd \
--skill write-a-prd \
-y
By default, skills are installed to all detected agent directories in the project. Use --agent to target a specific agent:
# Claude Code only (.claude/skills/)
npx skills@latest add semi-sentient/skills-workflow --agent claude-code --skill <skill-name>
# Agent-agnostic (.agents/skills/)
npx skills@latest add semi-sentient/skills-workflow --agent agents --skill <skill-name>
Updating
Update all installed skills to the latest version:
npx skills@latest update
Update a specific skill:
npx skills@latest update <skill-name>
Claude Code users: The
updatecommand does not support--agentand may create an.agents/skills/directory even in projects that only use.claude/. To update without this side effect, re-runaddwith--agent claude-code:npx skills@latest add semi-sentient/skills-workflow --agent claude-code --skill <skill-name> -y
Available Skills
Universal
Skills that apply to all projects regardless of language or framework.
| Skill | Description |
|---|---|
grill-me | Stress-test a plan or design through relentless questioning, resolving each branch of the decision tree one by one. Pass --light for a faster pass that skips counter-challenges. |
grill-with-docs | Stress-test a plan against the project's existing domain model and documented decisions, sharpening terminology and updating CONTEXT.md / ADRs inline as decisions crystallise. Pass --light for a faster pass that skips counter-challenges. |
tdd | Guide test-driven development using red-green-refactor with vertical slices (one test, one implementation, repeat). Includes the prove-it pattern for bug fixes. |
write-a-prd | Create a PRD through user interview, codebase exploration, and deep-module design. Outputs a Markdown PRD file to the project's plans directory, with an optional prompt to publish it as a GitHub issue labeled epic. Pass --no-github to stay local-only. |
prd-to-plan | Break a PRD into a phased implementation plan using tracer-bullet vertical slices. Outputs a Markdown plan file to the project's plans directory, with an optional prompt to publish it as a sub-issue of the PRD-epic. Accepts a #N / issue URL argument to link an existing PRD, or --no-github for local-only. |
run-plan | Orchestrate a multi-phase implementation plan by delegating phases to specialized sub-agents (Research, Code, Architect, Debug) with fresh context windows. Accepts a plan file path, a #N plan-sub-issue reference, or a full issue URL. Owns the full git/GH ceremony (work branch, per-phase commits, sub-issue sync, PR submission) with opt-out flags (--no-branch, --no-commits, --no-github, --no-pr). |
commit | Generate a Conventional Commits message from staged changes and commit. Accepts an optional ticket identifier. |
Skill dependencies
Most skills are standalone, but a few read or invoke others at runtime. If you install the skill on the left, you should also install everything on the right to get full functionality:
| Skill | Depends on | Why |
|---|---|---|
run-plan | commit | Invokes /commit after each phase for per-phase commits. Without it, the commit step fails (unless --no-commits is passed). |
run-plan | tdd | Every Code agent brief instructs the agent to read the tdd skill. Without it, agents silently skip the TDD workflow. |
write-a-prd | tdd (soft) | Reads the tdd skill while authoring to shape the PRD's Testing Decisions section. Without it, the PRD still gets written with generic guidance. |
write-a-prd | grill-with-docs or grill-me (soft) | Invokes a grilling skill at step 3 to stress-test the design before writing the PRD. grill-with-docs is preferred when available (adds inline CONTEXT.md / ADR maintenance); grill-me is the fallback. Skipped if a grilling session already ran in the same conversation, or if neither skill is installed. |
grill-me, grill-with-docs, prd-to-plan, tdd, and commit have no skill dependencies and work standalone. The install example at the top of this README lists all six workflow skills — use it as the default for any project that will run the full pipeline.
Frontend
Skills for client-side projects (React, TypeScript, etc.).
| Skill | Description |
|---|---|
| coming soon |
Backend
Skills for server-side projects. Stack-agnostic skills apply to any backend; stack-specific skills note their target in the name.
| Skill | Description |
|---|---|
| coming soon |
ML
Skills for machine learning and data science projects (Python, PyTorch, etc.).
| Skill | Description |
|---|---|
| coming soon |
Infra
Skills for DevOps, CI/CD, and infrastructure tasks.
| Skill | Description |
|---|---|
| coming soon |
Repository Structure
skills-workflow/
├── README.md
├── universal/ # Language/framework agnostic
│ ├── commit/
│ │ └── SKILL.md
│ └── tdd/
│ └── SKILL.md
├── frontend/ # Client-side (React, TypeScript, etc.)
├── backend/ # Server-side (Kotlin, Node, etc.)
├── ml/ # Machine learning (Python, PyTorch, etc.)
└── infra/ # DevOps / CI/CD
Each skill is a directory containing a SKILL.md file with YAML frontmatter (name, description) and markdown instructions. Supporting files (templates, scripts, reference docs) can live alongside SKILL.md in the same directory.
The domain folders (universal/, frontend/, backend/, ml/, etc.) are organizational — the CLI discovers skills by their name frontmatter, not their directory path. Stack-specific skills should make their target obvious in the skill name (e.g. ktor-conventions, hono-conventions) rather than relying on directory nesting.
Contributing
See CONTRIBUTING.md for guidelines on adding new skills and submitting changes.
Acknowledgments
The following skills were adapted from Matt Pocock's skills repo:
grill-megrill-with-docsprd-to-plantddwrite-a-prd
Many thanks to Matt for sharing his work. He gave an excellent talk about these and other skills at the AI Engineer - Europe conference which you should definitely check out if you're reading this. See docs/UPSTREAM.md for a per-skill summary of how our versions have diverged from upstream.