convez.ai
SkillDesenvolvimento WebAvançado
ieteerapat

Claude Code MCP Optimizer

ieteerapat/ClaudeSkills1.0

Otimiza tokens, Figma MCP e memória persistente para Claude Code

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill 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/ClaudeSkills1.0

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
14%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
2

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)

SkillVersionWhy It PassesWhat It Does
figma-accuracy1.3.0Agent gets Figma implementation wrong 65-80% of the time without structured workflowEnforces correct tool sequence, handles large designs, validates against screenshots, uses Code Connect, Code to Canvas
api-token-optimization1.0.0Agent gets prompt-cache invalidation wrong (images break cache, volatile-before-breakpoint) without itDirect Anthropic API cost cuts: prompt caching (90%), Batch API (50%), effort tuning, prefill — fills the API-side gap
session-memory1.1.0Agent has no built-in way to persist state across compactionsSaves/restores work state to files so you never re-discover the same codebase
hyperframes-video2.0.0Agent gets HyperFrames composition rules wrong (clip markers, paused timelines, registration) without themSelf-contained: create deterministic MP4 videos from HTML/CSS + GSAP. Reference files for animation, CLI, and pipeline
site-migrate1.8.0Agent 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 SkillReasonWhere It Went
token-saverResponse style rules — agent follows these from CLAUDE.mdAbsorbed into CLAUDE.md
output-filterSame — formatting rules, not a procedureAbsorbed into CLAUDE.md
model-routingDeclarative info agent already knows. Claude Code has /fast built-inAbsorbed into CLAUDE.md
mcp-token-optimizerOverlapped with figma-accuracy. General MCP rules are in CLAUDE.mdSplit between CLAUDE.md and figma-accuracy
figma-design-to-codeFully overlapped with figma-accuracy which is more thoroughMerged 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

ToolWhat It DoesInstall
RTKCompresses bash output 60-90%sh installers/install-rtk.sh (see installers/)
Context ModeSandboxes MCP output, 98% reduction/plugin marketplace add mksglu/context-mode
CodeGraphKnowledge graph, 92% fewer tool callsnpx @colbymchenry/codegraph
Claude-MemPersistent memory across sessionsnpx claude-mem install
CavemanExtreme 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:

SkillSourceInstallWhy It Helps
Figma Official PluginAnthropic marketplaceclaude plugin install figma@claude-plugins-officialIncludes figma-implement-design, figma-use, figma-code-connect, figma-create-design-system-rules
skill-creatoranthropics/skills (67K ⭐)/plugin marketplace add anthropics/skillsCreate and evaluate your own skills with structured evals
Figma MCP Server Guidefigma/mcp-server-guidenpx skills add https://github.com/figma/mcp-server-guide --skill figma-implement-designOfficial 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:

SkillWhy It Fails
Most "terse output" skillsAgent follows these from CLAUDE.md — not procedural
Most "code review" skillsDeclarative checklists — agent already knows good practices
Most "commit message" skillsSimple formatting — one line in CLAUDE.md is enough
Most "SOLID principles" skillsAgent already knows SOLID — this is general knowledge
Generic "best practices" skillsVague 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 name
  • description: 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
Ver o original no GitHub

Bom para

Relacionados