System Design Skill
ftvision/system-design-skill
Prática de design de sistemas com entrevistador IA e múltiplos modos de aprendizado
Ver no GitHubSkill para treinar entrevistas de design de sistemas com Claude Code e Codex. Oferece quatro modos: mock (entrevistador rigoroso), learn (você entrevista Claude ou observa dois agentes), postmortem (análise de entrevistas reais) e generate (criar novas questões com rubrica).
Instalação
npx convez-ai ftvision/system-design-skill1 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
- Preparação para entrevistas técnicas de design de sistemas
- Aprendizado observando respostas de qualidade de um agente IA
- Geração automática de questões e rubricas de avaliação
- Análise pós-entrevista com feedback estruturado
Veredito

Ferramenta bem executada para quem estuda design de sistemas. Mock rigoroso, modo observatório inteligente e geração de rubrica são pontos fortes. Sem interfaces gráficas, mas integração com Claude Code e Codex é sólida.
README
A tradução para o português deste README ainda está na fila.
system-design
A skill for practicing system design interviews. Four modes:
mock— Claude plays a strict staff-level interviewer. Phases, time-boxing, deep-dive, and a scored debrief at the end.learn— Reverse mock. You interview Claude. Watch what good answers look like. With--auto, Claude orchestrates two sub-agents (interviewer + candidate) and you observe a full session.postmortem— Diagnose a real interview you took. Pass--filewith notes, or answer structured questions.generate— Author a fresh question + rubric. Writes four files (question.md,assumptions.md,description.md,rubric.md) to./system-design-questions/<slug>/.
Works in Claude Code (raw skill, or as a plugin via the marketplace) and Codex CLI.
Demo
learn --auto on design X.com's timeline feed (staff level). Two sub-agents — one interviewer, one candidate — run a full 5-round interview while you watch, and the session closes with a written retrospective grading both sides.
https://github.com/user-attachments/assets/24673305-6dbf-4b39-a04a-6262a61d961c
A shorter clip showing first-run setup (examples/init/demo-init.mp4) is also checked into the repo.
Install
Pick the path that matches your harness. The repo ships three pre-built install layouts so you can copy the right one.
Claude Code — plugin (recommended)
Install via the built-in marketplace:
/plugin marketplace add ftvision/system-design-skill
/plugin install system-design@system-design
After install, /system-design is available in any Claude Code session. Updates flow through /plugin update.
Claude Code — raw skill copy
git clone https://github.com/ftvision/system-design-skill.git /tmp/system-design-skill
cp -R /tmp/system-design-skill/.claude/skills/system-design ~/.claude/skills/system-design
Verify: ls ~/.claude/skills/system-design/SKILL.md. Then /system-design is available.
Codex CLI
User-wide:
git clone https://github.com/ftvision/system-design-skill.git /tmp/system-design-skill
mkdir -p ~/.agents/skills
cp -R /tmp/system-design-skill/.agents/skills/system-design ~/.agents/skills/system-design
Project-local: copy .agents/skills/system-design/ into your project's .agents/skills/.
After install, open /skills in Codex (or invoke $system-design ...). Restart Codex if the skill doesn't appear immediately.
Usage
Replace the prefix with whatever your harness uses (/ for Claude Code, $ for Codex).
system-design mock # generate a question, run a strict mock
system-design mock "design a URL shortener"
system-design mock --level=senior
system-design postmortem # structured Q&A about a past interview
system-design postmortem --file=./notes.md
system-design generate # generate question + rubric (4 files)
system-design generate "rate limiter"
system-design generate "chat system" --level=staff
system-design generate --direction=ml-infra # bias topic to ML infra subdomain
system-design generate --direction=llm # LLM inference / RAG / agents
system-design learn # you interview Claude
system-design learn "design Cursor's autocomplete backend"
system-design learn --auto # two sub-agents, you watch
system-design learn --auto --exchanges=15 # cheaper run (default 30 exchanges)
State (persisted across sessions)
The skill writes to ~/.system-design/state/:
| File | Purpose |
|---|---|
runs.md | One row per scored session: date, slug, mode, level, direction, the five dimension scores, and a one-sentence action item for the next session (may be blank). The primary tracker — read at the start of each mock/postmortem to surface a preamble (total sessions, recurring weak dimensions, last 3 slugs, plus the previous session's action item if set). Slug column also serves as the "already practiced" list for mock and generate. |
weaknesses.md | One row per weak dimension (score ≤3) from past mock debriefs and postmortem diagnoses, with a one-line context quote. mock biases its deep-dive picks toward recurring weak dimensions. |
level.md | Your target level (default staff if absent). |
State is per-user, harness-neutral, and not in this repo. It's created on first use. Same state directory whether you installed in Claude Code or Codex — practice carries over.
What gets generated where
generatewrites question packages to the current working directory, under./system-design-questions/<slug>/. Run it from wherever you want the files to live (a notes repo, a dotfiles directory, etc.).mockandpostmortemupdate state files in~/.system-design/state/.learnandlearn --autodon't write any files unless you ask.
Scoring dimensions
Both mock debriefs and postmortem diagnoses score across:
- Requirements scoping
- High-level structure
- Deep-dive depth
- Tradeoff reasoning
- Communication
Each gets a 1–5 with one-line justification grounded in something you actually said.
Cost note on learn --auto
Each exchange in --auto mode = 2 sub-agent calls. Default 30 exchanges = ~60 calls. The skill will tell you the count before starting and let you override with --exchanges=N. Use a smaller number to sanity-check the format before committing to a full run.
Repo layout
system-design-skill/
├── README.md # this file
├── skill/ # canonical source (edit here)
│ ├── SKILL.md
│ └── reference/
│ ├── mock.md
│ ├── learn.md
│ ├── postmortem.md
│ └── generate.md
├── scripts/sync.sh # syncs skill/ -> the three install layouts
│
├── .claude-plugin/marketplace.json # Claude Code marketplace manifest
├── plugin/ # Claude Code plugin layout
│ ├── .claude-plugin/plugin.json
│ └── skills/system-design/ # synced from skill/
│
├── .claude/skills/system-design/ # Claude Code raw-install layout (synced)
└── .agents/skills/system-design/ # Codex CLI raw-install layout (synced)
Edit skill/ only. The three install layouts are derived copies — never edit them by hand. After changing skill/, run ./scripts/sync.sh to regenerate them, then commit all four (skill/ + the three derived dirs) so users can cp directly without running the script.
Drift is guarded: ./scripts/sync.sh --check verifies the derived copies match skill/ (no writes), and CI runs it on every PR — so a hand-edited or un-synced copy fails the build.