Definition
The external skills ecosystem refers to the set of installable agent skills that provide LLM Wiki and second brain functionality. Nine skills across four bundles were evaluated alongside this vault’s approach.
Key Bundles
nicholasspisak/second-brain (4 skills)
An onboarding wizard and operational toolkit for LLM Wiki vaults. Includes:
- second-brain — 5-step interactive wizard: vault name, location, domain, agent config, tooling
- second-brain-ingest — Read source → discuss takeaways with user → create source summary → create wiki pages → update index/log
- second-brain-lint — Broken wikilinks, orphan pages, contradictions, stale claims, missing cross-references
- second-brain-query — Index-based search, then qmd for large wikis, than read pages with wikilink traversal
Key difference from this vault: the second-brain bundle uses a wiki/sources/ directory with source summary pages, and explicitly discusses takeaways with the user before writing.
akillness/oh-my-skills@llm-wiki (1 skill)
A self-contained implementation of the LLM Wiki pattern as an agent skill. Covers bootstrap, ingest, query, lint, and optional tooling (Scrapling for URL ingestion, qmd for search, Obsidian CLI). Uses a raw/ → wiki/ → index/log structure nearly identical to this vault’s.
anthropics/knowledge-work-plugins (4 skills)
Enterprise-oriented knowledge work skills for MCP-connected environments:
- knowledge-synthesis — Cross-source deduplication, confidence scoring by freshness/authority, result summarization
- memory-management — Two-tier memory: CLAUDE.md (hot cache, ~30 entries) + memory/ directory (full knowledge base)
- source-management — MCP source connection management (chat, email, cloud storage, project tracker, CRM, knowledge base)
- search — Multi-source search with query decomposition and result synthesis
These skills assume MCP-connected enterprise sources (Slack, email, Google Drive) — a fundamentally different operational context from this vault’s file-based approach.
Comparison to This Vault
| Dimension | This Vault | second-brain | llm-wiki skill | Anthropic plugins |
|---|---|---|---|---|
| Architecture | raw/ → wiki/ → governance | raw/ → wiki/sources/ → wiki/ | raw/ → wiki/ | MCP sources → synthesis |
| User interaction | Autonomous (LLM decides) | Consultative (discusses with user) | Mixed | Autonomous |
| Search | index.md + wikilinks | index.md + qmd | index.md + qmd | Multi-MCP search |
| Confidence | frontmatter confidence field | Not observed | Not observed | Freshness/authority scoring |
| Source type | Files only | Files only | Files + URLs (Scrapling) | MCP-connected enterprise |
| Target user | Individual researcher | Individual | Individual | Enterprise teams |
Beyond LLM Wiki: the coding-agent skill ecosystem
Este vault também foi exposto ao ecossistema mais amplo de agent skills para coding — não só para second brain. Superpowers e Matt Pocock’s Skills representam os dois pólos filosóficos que definem o desenho de quase toda skill framework:
- Enforced discipline: superpowers impõe um pipeline obrigatório ao agente (evita “vibe coding” acidental, garante processo). 14 skills, ~268k★, 10+ harnesses. Custo: token burn, bootstrap frágil em WSL.
- Composable toolset: matt-pocock-skills dá skills invocáveis manualmente, flexíveis. 38 skills, ~208k★, flagship Claude Code. Custo: carga de operador mais alta; colisão de comandos se ambos rodarem como rotas ativas.
A tese que emerge (exportada de uma pesquisa comparativa) é que não são mutuamente exclusivos — a comunidade compõe os dois (superpowers-vs-matt-pocock-skills) em pipelines próprios. Projetos documentados: devflow, omni-skills, dmi_superpowers, lean-skills. Para este vault, a lição é que o desenho de skills deve ser consciente de qual pólo prioriza: garantia de processo vs. agilidade sob controle humano. Compare com as skills deste repo, que já misturam ambos.
Related Concepts
- llm-wiki — The pattern these skills implement
- ingest — All skills have an ingest operation
- lint — Lint is universal across LLM Wiki implementations
- query — How different implementations approach retrieval
- agent-memory-systems — The broader field that overlaps with these skills
- is-firecrawl-useful-to-this-wiki-purpose — Filed verdict on a hosted web-capture/research skill (firecrawl-deep-research) vs this vault’s pipeline
- superpowers — coding-agent skill framework (enforced discipline)
- matt-pocock-skills — coding-agent skill toolkit (composable toolset)
- superpowers-vs-matt-pocock-skills — the head-to-head and the case for composition
Implications
The external skills ecosystem reveals three design dimensions this vault could evolve:
- Consultative mode: The second-brain bundle’s “discuss with user before writing” step would reduce incorrect page creation. This vault currently trusts the LLM entirely.
- Enterprise MCP integration: The anthropic plugins assume MCP-connected sources. If this vault ever needs to search across Slack, email, or cloud storage, that architecture is a model.
- qmd for the search: Both second-brain and llm-wiki skills recommend qmd for large wikis. This vault should consider qmd when index.md exceeds ~200 entries.
Terceiro paradigma: role-based orchestration (omni-skills, 2026-08-07)
Além do embate Superpowers (disciplina imposta) × Matt Pocock (skills composáveis), o ecossistema já tem um terceiro paradigma: o omni-skills (devos-ing) não organiza por filosofia técnica, mas por papéis e playbooks. Instala-se um “team” (Startup Team, Finance Team), cada papel (CEO, CTO, PM, Founding Engineer, QA Lead) com seu workflow e evidence ledger (claims em Verified/Inferred/Assumed), gate de plano antes de implementar, verificação e replay de expectativas do usuário na entrega. Execução por subagentes internos com stage packets limitados; fallback “Prepared, not executed” quando não há host capaz.
Relevância para este conceito: é um eixo de design onde o trade-off não é mais auto vs manual, mas orquestração por papéis — um passo além da composição Superpowers vs Pocock. A evidência (README/demos do omni-skills) é raw/external/github-com-omni-skills-d0c3d6d1.md; métricas de impacto ainda não são públicas (ver Open Questions). Também surgem sub-conjuntos enxutos: lean-skills (zhoukaichaoaa) reduz Matt+Superpowers a 9 skills (5 residentes) com foco em budget de contexto — evidência de que o modelo “pega o que precisa, sem o resto” é uma estratégia valorizada pela comunidade.
Cauda do ecossistema: tooling de segunda camada (2026-08-07)
Além do embate de frameworks, aparecem ferramentas que administram o acervo em vez de criar skills:
- agentic-code-stack (makafeli) — stack de referência mínimo por declaração: um
AGENTS.mddrop-in na raiz do projeto dá ao agente o mapa de skills/plugins/MCP; base = Matt Pocock + extras. “The constraint is the feature” — a interseção enxuta é o produto. - skill-manager (Skillet, jnyross) — TUI para inventar, arquivar e reverter skills em Claude Code + Codex; scriptável para agente/CI. A resposta ao “skill sprawl”.
- opencode-ship (Viktorxyz) — pacote de entrega tech-neutral para OpenCode: issue/PR/worktree lifecycle + reviewer/verifier. Não é skill, é delivery layer.
- Token-Effort (HeadlessTarry) — skills minimalistas pro OpenCode no eixo de custo/contador (“low-stakes intelligence”); reforça o trade-off token-budget.
Evidência imutável: raw/external/github-com-agentic-code-stack-6d1a1ef4.md, raw/external/github-com-skill-manager-b28afc7a.md, raw/external/github-com-opencode-ship-7ff38c33.md, raw/external/github-com-token-effort-d6464fd9.md.
Implications
The external skills ecosystem reveals design dimensions this vault could evolve on top of the existing three:
- Stack declarado: a forma “novo agent-visible em 1 arquivo” (agentic-code-stack) é um modelo interessante para este vault (AGENTS.md já é o nosso drop-in).
- Segunda camada de gestão: com o grande acervo deste repo, um inventário/auditoria que “vê” todas as skills ref-lds ao mesmo tempo reduz o skill sprawl (analogo ao Skillet).
- Entrega como camada separada: um delivery layer (issue→worktree→PR→verificar) é o que queremos ao escalar o vibe-coding além do protótipo.
Open Questions
- Does consultative mode (discussing with user) produce better pages than autonomous ingest? No empirical comparison exists.
- Can the anthropic plugin architecture be adapted for file-based vaults, or is it fundamentally MCP-dependent?
- Which skill’s lint implementation is most effective? Each covers a different subset of checks — no comprehensive benchmark exists.