Claude Code MCP Optimizer
ieteerapat/ClaudeSkills1.0
Otimiza tokens, Figma MCP e memória persistente para Claude Code
Ver no GitHubSkill package para Claude Code que integra otimização de tokens, acurácia em Figma MCP, persistência de memória via Obsidian e criação de vídeos com HyperFrames. Inclui ferramentas de auditoria de contexto, compactação inteligente e workflows procedurais validados contra padrões agentskills.io.
Instalação
npx convez-ai ieteerapat/ClaudeSkills1.01 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
- Reduzir consumo de tokens em projetos Claude Code com Figma
- Manter contexto persistente entre sessões via Obsidian
- Converter workflows HTML em vídeos com HyperFrames
- Auditar e compactar contexto inteligentemente
- Migrar sites stack-agnostic com precisão
Veredito

Skill maduro e bem estruturado, com filosofia clara: carregar apenas workflows que o agente não faria sem eles. Excelente para desenvolvedores que usam Claude Code + Figma + Obsidian e precisam gerenciar contexto agressivamente. Instaladores multiplataforma são um diferencial real.
README
A tradução para o português deste README ainda está na fila.
Claude Code MCP Optimizer
A skills package for Claude Code focused on Figma MCP accuracy, token efficiency, persistent memory (Obsidian vault), and video authoring (HyperFrames). Includes companion-tool installers (RTK, Obsidian MCP). Validated against the agentskills.io open standard.
Philosophy
Per agentskills.io best practices:
- Skills should add what the agent wouldn't know without them
- Rules the agent already follows (be terse, use diffs) belong in CLAUDE.md, not skills
- Skills should be procedural workflows, not declarative style guides
This package keeps only skills that pass the test: "Would the agent get this wrong without the skill?"
What's Included
claude-code-mcp-optimizer/
├── CLAUDE.md # Project rules (response style, token budget, MCP discipline)
├── .claude/
│ ├── settings.json # Token caps + context statusline
│ ├── commands/
│ │ ├── compact-now.md # /compact-now — smart context compaction
│ │ ├── figma-extract.md # /figma-extract — token-efficient Figma data pull
│ │ ├── token-audit.md # /token-audit — check session token health
│ │ ├── save-state.md # /save-state — persist context before compaction
│ │ └── restore-state.md # /restore-state — resume from saved state
│ ├── hooks/
│ │ ├── prevent-large-reads.json # Warns before cat-ing files >100KB
│ │ ├── context-budget-alert.json # Alerts at 500K+ tokens
│ │ └── read-before-edit.json # Ensures files are read before editing
│ └── skills/
│ ├── figma-accuracy/ # Pixel-perfect Figma MCP workflow
│ │ ├── SKILL.md
│ │ └── references/
│ │ ├── mcp-servers.md # All Figma MCP server options + configs
│ │ ├── property-mapping.md # Figma → CSS exact translation tables
│ │ └── tool-reference.md # All 16 official tools documented
│ ├── api-token-optimization/ # Direct Anthropic API cost optimization
│ │ └── SKILL.md
│ ├── session-memory/ # Cross-session context preservation
│ │ └── SKILL.md
│ └── hyperframes-video/ # HTML → MP4 video (self-contained)
│ ├── SKILL.md
│ └── references/
│ ├── animation-patterns.md # GSAP + adapter registration recipes
│ ├── cli-reference.md # Full CLI commands + prerequisites
│ └── pipeline.md # 7-step multi-beat video pipeline
│ └── site-migrate/ # Stack-agnostic site migration engine (vendored)
│ ├── SKILL.md
│ ├── config/ # default mask-rules + thresholds
│ ├── references/ # intake, seo, sources, targets, ci, etc.
│ └── scripts/ # intake/seed/capture/compare/extract/manifest (.mjs)
├── installers/
│ ├── install-rtk.sh # RTK token-saver installer (macOS/Linux/WSL/Git Bash)
│ ├── install-rtk.ps1 # RTK installer (native Windows)
│ ├── setup-obsidian-vault.sh # Install Obsidian + register the vault locally
│ ├── setup-obsidian-mcp.sh # Connect Claude Code to the vault via MCP
│ ├── obsidian-mcp.md # Obsidian MCP setup guide
│ └── README.md # RTK setup + usage guide
├── vault/ # Obsidian vault — persistent memory store
│ ├── .obsidian/ # Obsidian config + Local REST API plugin
│ ├── Index.md # vault home / map of content
│ ├── Sessions/ # one note per work session
│ ├── Knowledge/ # durable facts and decisions
│ ├── Daily/ # daily notes
│ ├── _templates/ # session / knowledge / daily templates
│ └── README.md # vault guide
├── .mcp.json.example # Obsidian MCP config template (real .mcp.json gitignored)
├── CHANGELOG.md # Per-skill version history
├── README.md
└── SOURCES.md # All research links
Skills (5 — validated against agentskills.io)
| Skill | Version | Why It Passes | What It Does |
|---|---|---|---|
figma-accuracy | 1.3.0 | Agent gets Figma implementation wrong 65-80% of the time without structured workflow | Enforces correct tool sequence, handles large designs, validates against screenshots, uses Code Connect, Code to Canvas |
api-token-optimization | 1.0.0 | Agent gets prompt-cache invalidation wrong (images break cache, volatile-before-breakpoint) without it | Direct Anthropic API cost cuts: prompt caching (90%), Batch API (50%), effort tuning, prefill — fills the API-side gap |
session-memory | 1.1.0 | Agent has no built-in way to persist state across compactions | Saves/restores work state to files so you never re-discover the same codebase |
hyperframes-video | 2.0.0 | Agent gets HyperFrames composition rules wrong (clip markers, paused timelines, registration) without them | Self-contained: create deterministic MP4 videos from HTML/CSS + GSAP. Reference files for animation, CLI, and pipeline |
site-migrate | 1.8.0 | Agent has no reliable parity-gated migration loop; without it "done" means "built" not "matches source" | Stack-agnostic site migration engine (WordPress/crawl → Next.js+Tailwind/Astro). Intake → golden-fixture capture (Playwright) → manifest-driven per-page loop → parity gate. Bundled scripts + references |
See CHANGELOG.md for version history.
What Was Removed (and why)
| Removed Skill | Reason | Where It Went |
|---|---|---|
token-saver | Response style rules — agent follows these from CLAUDE.md | Absorbed into CLAUDE.md |
output-filter | Same — formatting rules, not a procedure | Absorbed into CLAUDE.md |
model-routing | Declarative info agent already knows. Claude Code has /fast built-in | Absorbed into CLAUDE.md |
mcp-token-optimizer | Overlapped with figma-accuracy. General MCP rules are in CLAUDE.md | Split between CLAUDE.md and figma-accuracy |
figma-design-to-code | Fully overlapped with figma-accuracy which is more thorough | Merged into figma-accuracy |
Installation
For Claude Code
# Clone
git clone https://github.com/ieteerapat/ClaudeSkills1.0.git
# Copy into your project
cp -r ClaudeSkills1.0/.claude/ /path/to/your/project/
cp ClaudeSkills1.0/CLAUDE.md /path/to/your/project/
Global (all projects)
cp -r ClaudeSkills1.0/.claude/skills/* ~/.claude/skills/
Figma MCP Setup (Quick Start)
# 1. Install official Figma plugin (includes MCP server + skills)
claude plugin install figma@claude-plugins-official
# 2. Set token limit for large designs
export MAX_MCP_OUTPUT_TOKENS=100000
# 3. Generate project-specific rules (one-time)
# In Claude Code, prompt: "create design system rules for my project"
Companion Tools
| Tool | What It Does | Install |
|---|---|---|
| RTK | Compresses bash output 60-90% | sh installers/install-rtk.sh (see installers/) |
| Context Mode | Sandboxes MCP output, 98% reduction | /plugin marketplace add mksglu/context-mode |
| CodeGraph | Knowledge graph, 92% fewer tool calls | npx @colbymchenry/codegraph |
| Claude-Mem | Persistent memory across sessions | npx claude-mem install |
| Caveman | Extreme output compression 75% | curl -fsSL .../install.sh | bash |
Validated External Skills (install separately)
These skills pass the agentskills.io test ("would the agent get this wrong without it?") and complement this package:
| Skill | Source | Install | Why It Helps |
|---|---|---|---|
| Figma Official Plugin | Anthropic marketplace | claude plugin install figma@claude-plugins-official | Includes figma-implement-design, figma-use, figma-code-connect, figma-create-design-system-rules |
| skill-creator | anthropics/skills (67K ⭐) | /plugin marketplace add anthropics/skills | Create and evaluate your own skills with structured evals |
| Figma MCP Server Guide | figma/mcp-server-guide | npx skills add https://github.com/figma/mcp-server-guide --skill figma-implement-design | Official pixel-perfect implementation workflow |
Skills That FAILED Validation (don't install)
Evaluated against agentskills.io criteria — these don't add value over CLAUDE.md rules:
| Skill | Why It Fails |
|---|---|
| Most "terse output" skills | Agent follows these from CLAUDE.md — not procedural |
| Most "code review" skills | Declarative checklists — agent already knows good practices |
| Most "commit message" skills | Simple formatting — one line in CLAUDE.md is enough |
| Most "SOLID principles" skills | Agent already knows SOLID — this is general knowledge |
| Generic "best practices" skills | Vague declarations, not procedures with domain expertise |
Research Sources
See SOURCES.md for 50+ GitHub repos, articles, and official documentation.
Validation
Skills in this package follow the Agent Skills specification:
name: lowercase, hyphens only, matches folder namedescription: describes what + when to use- Body: procedural instructions, not declarative rules
- Under 500 lines, references/ for detailed content
- Evaluated per agentskills.io/skill-creation/evaluating-skills