Claude Skills Vault
Bradhardwicke/Claude-Skills-Vault
Cérebro de referência otimizado para Claude encontrar e invocar a skill certa
Ver no GitHubSistema de referência token-eficiente que organiza e indexa skills especializadas para Claude, permitindo descoberta rápida sem pré-carregar todas as definições em cada conversa. Usa um sistema de 3 níveis (hardcoded, índice YAML compacto, notas sob demanda) para minimizar consumo de tokens enquanto mantém 128+ skills descobríveis.
Instalação
npx convez-ai Bradhardwicke/Claude-Skills-Vault1 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 custo de tokens em conversas Claude mantendo acesso a 128+ skills
- Organizar skills por categoria e trigger para descoberta inteligente
- Servir como single source of truth para definições de skills em Obsidian
Veredito

Vault bem estruturado que resolve o problema real de discovery de skills sem inflacionar custos de contexto. Útil para quem gerencia múltiplas skills no fluxo Claude-Obsidian, mas requer manutenção manual de índices e é mais relevante para power users com dezenas de plugins customizados.
README
A tradução para o português deste README ainda está na fila.
Claude Skills Vault
Agents Read, Humans Write. A token-efficient skill reference system so Claude can identify and invoke the right skill for any task.
What This Is
Claude has access to hundreds of skills specialised tools for writing documents, analysing data, running SQL, managing projects, building agents, and more. The problem is discovery: how does Claude know which skill to reach for, without pre-loading every skill definition into every conversation?
This vault solves that problem. It is Claude's reference brain authored by humans, optimised for Claude to read quickly at conversation start. It is not a general-purpose personal knowledge base. Every note in this vault exists to help Claude do its job better.
The Problem It Addresses
Pre-loading all 128+ skills into every Claude conversation costs thousands of tokens before the first useful word is written. It also creates ambiguity —when everything is loaded, nothing is prioritised.
Most community patterns for AI-assisted PKM follow the rule: human thinking goes in the vault, Claude outputs go in ~/.claude/. This vault consciously inverts that pattern. Skill definitions are the primary content here. Claude's session memory, working plans, and task outputs live elsewhere.
How It Works
The vault uses a three-tier lookup system to minimise token cost whilst keeping all skills discoverable.
| Tier | What Claude reads | Token cost |
|---|---|---|
| 1 | Top 25 skills hardcoded in CLAUDE.md | Free (already in context) |
| 2 | Index/skills-lookup.yaml — compact intent-to-skill-id map | ~2,800 tokens |
| 3 | Individual skill notes in Skills/ | On demand only |
At session start, Claude checks Tier 1 first. If the required skill is in the Top 25, it invokes it immediately. If not, it reads the compact YAML index. Only when a skill is actually invoked does Claude read the full skill note. Total lookup cost across all three tiers: approximately 5,000 tokens.
Skill Notes
Each skill note has two frontmatter blocks — Obsidian metadata followed by the verbatim SKILL.md content from the plugin:
---
# Obsidian frontmatter
name: data-analyze
skill-id: "data:analyze"
plugin: Data
category: Data
tags: [skill, data, analysis]
aliases: [Data Analysis, Analyze Data]
triggers:
- data question
- metric lookup
- what drove
- analytics report
---
---
# SKILL.md frontmatter (copied verbatim from plugin)
name: analyze
description: Answer data questions...
---
# Full SKILL.md content follows...
The triggers field is the single source of truth. The build script reads every trigger across every skill note and generates both index files automatically.
Vault Structure
Claude Skills Vault/
├── Skills/ # 128 skill notes across 12 categories
│ ├── Documents/ # docx, xlsx, pptx, pdf, view-pdf
│ ├── Data/ # 10 data:* skills
│ ├── Design/ # 7 design:* skills
│ ├── Operations/ # 9 operations:* skills
│ ├── Sales/ # 9 sales:* skills
│ ├── Productivity/ # 4 productivity:* skills
│ ├── Obsidian/ # 5 obsidian:* skills
│ ├── Superpowers/ # 14 advanced dev workflow skills
│ ├── PM Skills/ # 8 project management skills
│ ├── Executive Mentor/ # 6 executive decision skills
│ ├── Finance/ # 3 finance skills
│ ├── Markdown HTML/ # 11 markdown-to-HTML skills
│ └── Engineering/ # 15 engineering and tooling skills
├── Index/
│ ├── skills-lookup.yaml # AUTO-GENERATED — compact intent index
│ ├── Skill Quick Reference.md# AUTO-GENERATED — curated intent guide
│ ├── Skills Index.md # Human-readable map of categories
│ └── Skills Database.base # Obsidian Base view (for human use)
├── _scripts/
│ └── build_skill_brain.py # Regenerates both index files from skill notes
├── .claude/commands/ # /preserve, /compress, /resume commands
├── CLAUDE.md # Session memory and Top 25 skill list
├── Project Context.md # Design decisions and project history
└── Welcome.md # Vault overview
Skill Categories
| Category | Skills | Purpose |
|---|---|---|
| Documents | 5 | Word, Excel, PowerPoint, PDF creation and manipulation |
| Data | 10 | SQL, analysis, visualisation, dashboards, statistical testing |
| Design | 7 | UX copy, critique, accessibility, user research |
| Operations | 9 | Process docs, status reports, runbooks, risk assessment |
| Sales | 9 | Research, outreach, call prep, competitive intel, forecasting |
| Productivity | 4 | Task management, memory systems, working context |
| Obsidian | 5 | Vault interaction, markdown, bases, CLI, JSON canvas |
| Superpowers | 14 | Advanced dev workflows: brainstorming, TDD, debugging, code review |
| PM Skills | 8 | Project management, Jira, Confluence, scrum, communications |
| Executive Mentor | 6 | Board prep, stress testing, decisions, postmortems |
| Finance | 3 | Financial analysis, valuation, SaaS metrics |
| Engineering | 15 | Workflows, skill creation, MCP building, debugging tooling |
Current Status
128 skills documented. All have full SKILL.md content and triggers frontmatter. The Index contains 1,105 intent triggers across all skills.
Recently Added Categories
The following six categories were added in the most recent build session, bringing the vault from 71 to 128 documented skills:
| Category | Skills Added |
|---|---|
| Superpowers | 14 — brainstorming, writing-plans, executing-plans, TDD, debugging, code review, subagent-dev, git-worktrees, and more |
| PM Skills | 8 — senior-pm, scrum-master, meeting-analyzer, jira-expert, confluence-expert, atlassian-admin, atlassian-templates, team-comms |
| Executive Mentor | 6 — executive-mentor, board-prep, challenge, hard-call, postmortem, stress-test |
| Finance | 3 — financial-analyst, saas-metrics-coach, finance-skills |
| Markdown HTML | 11 — markdown-html-orchestrator, md-document, md-review, md-slides, design-system, and command variants |
| Engineering | 15 — workflow-builder, write-a-skill, caveman, slopmop and sub-commands, mcp-builder, ai-agents-architect, and more |
Skills Still to Add (Optional)
| Skill | Notes |
|---|---|
helm-chart-builder | Engineering category |
frontend-design | Claude API plugin variant |
webapp-testing | Claude API plugin variant |
slack-gif-creator | Claude API plugin variant |
2slides-ppt-generator | Claude API plugin variant |
Known Issue: Index Regeneration
The build_skill_brain.py script was updated with all new TRIGGERS and INTENT_GROUPS entries covering all 128 skills, but the Write tool silently truncates large files on the Windows filesystem mount. The script currently contains only its file header (18 lines).
Fix: Write the full script via bash heredoc in chunks, then run it to regenerate both index files. The full script content is documented in the session transcript.
Both skills-lookup.yaml and Skill Quick Reference.md currently reflect only the original 71 skills until the script is rebuilt and run.
Adding a New Skill
- Create a note in the appropriate
Skills/subfolder. - Add Obsidian frontmatter with
skill-id,plugin,category,tags,aliases, andtriggers. - Paste the full
SKILL.mdcontent below (including its own frontmatter block from the plugin). - Run the build script from the vault root:
python "_scripts/build_skill_brain.py"
The script is idempotent — it skips notes already processed and regenerates both index files from scratch on every run.
Design Principles
Agents Read, Humans Write. The dominant community pattern is: human thinking goes in the vault, Claude outputs go in ~/.claude/. This vault inverts that pattern. It is Claude's reference brain. Skill definitions are the primary content, not personal notes or session memory.
Single source of truth. Triggers live in frontmatter only. The build script reads them to generate all lookup files. Adding a skill means editing one note, then running one script.
Lazy loading. Claude reads the minimum it needs. Top 25 are free. The YAML index is compact. Full skill notes load only on invocation.
Boundary rule. Session memory, working plans, task outputs, and Claude-generated content belong in ~/.claude/ or Cowork's memory system — not in this vault.
Future Direction
QMD Semantic Search
The current dual-index approach costs approximately 5,000 tokens per session. QMD (npm install -g @tobilu/qmd) could replace this with a single semantic query returning only the relevant skill. Windows-compatible. Trial recommended before committing.
Custom Slash Command
A /skills command (D:\Claude Repo\Claude Skills Vault\.claude\commands\skills.md) provides a structured alternative in the meantime — loading the right index file on demand instead of pre-loading both.
Repository
Local vault path: D:\Claude Repo\Claude Skills Vault
MCP integrations:
- Obsidian CLI — enabled
- Desktop Commander — active, used for file read/write on this vault
Last updated: 7 June 2026 | 128 skills | 12 categories | 1,105 triggers