convez.ai
SkillAutomaçãoIntermediário
ndjordjevic

pin-llm-wiki

ndjordjevic/pin-llm-wiki

Converta URLs em wiki estruturada, citável e consultável por agentes

Ver no GitHub
BomPluginScore · de 0 a 100

Skill que automatiza a criação de uma base de conhecimento local e estruturada a partir de URLs (web, GitHub, YouTube). Ingere fontes externas em pastas organizadas (raw/ e wiki/), gera referências cruzadas e citações, e disponibiliza tudo para agentes IA consultarem antes de responder.

Instalação

$npx convez-ai ndjordjevic/pin-llm-wiki

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
76
Forks
14
Acompanhando
1
Saúde
42%
Último commit
há 4 meses
Versão
—
Issues abertas
1
Contribuidores
2

Casos de uso

  • Criar bases de conhecimento locais para domínios específicos (ex: LLMs, arquitetura)
  • Manter memória durável e versionada em git para contexto de agentes IA
  • Organizar pesquisas e documentação externa em estrutura wikilink para reutilização

Veredito

Skill bem pensada que resolve um problema real: transformar fontes externas dispersas em knowledge base consultável e durável para agentes IA. Funciona nativamente em Claude Code, Cursor e Copilot, com fluxo claro (init → queue → ingest → lint). Recomendado para quem trabalha com pesquisa intensiva ou precisa manter contexto estruturado entre múltiplas sessões.

README

A tradução para o português deste README ainda está na fila.

pin-llm-wiki

A multi-editor skill that automates the Karpathy LLM Wiki pattern: drop in URLs, get a local, citable, cross-referenced wiki that agents can read before answering.

Where it runs: Claude Code (slash commands), GitHub Copilot and Cursor (install the skill + follow the same workflows; see below).

Why use it

pin-llm-wiki turns external sources into a durable knowledge base:

  • raw/ keeps immutable captures of the original sources.
  • wiki/ holds summarized, wikilinked pages with citations back to raw files.
  • AGENTS.md tells AI agents to consult the wiki before answering domain questions.
  • inbox.md gives humans and agents a simple queue for future sources.

The result is a repo-local memory layer: reviewable in git, queryable by agents, and less dependent on whatever context happens to fit in one chat.

Install

Use the skills CLI: npx skills@latest add ndjordjevic/pin-llm-wiki (project) or -g (global, e.g. ~/.cursor/skills). Project installs land in .agents/skills/ (Cursor, Copilot, …) and .claude/skills/ (Claude Code)—not in ./skills/ (that path is only this repo’s package layout).

For Claude Code + Cursor + Copilot in one repo, create .claude first so the CLI can install one real copy under .agents/skills/ and symlink Claude Code to it.

mkdir -p .claude
npx skills@latest add ndjordjevic/pin-llm-wiki \
  --agent claude-code --agent cursor --agent github-copilot -y

npx skills@latest add ndjordjevic/pin-llm-wiki -g   # global

Use --agent … for a predictable layout in agent terminals. --list, skills.sh, npx skills@latest add --help.

Update

Refresh the skill files from GitHub. Scope should match how you installed:

# Project install: from that repo’s root, or force project scope
npx skills@latest update pin-llm-wiki -p

# Global install
npx skills@latest update pin-llm-wiki -g

Use -y to skip interactive scope prompts (the CLI can auto-pick project vs global when only one applies). skills update may recreate agent directories you do not use; delete those folders if you want a minimal tree. See npx skills@latest update --help.

Quickstart

pin-llm-wiki CLI workflow: install via npx skills, then init, queue, ingest, and lint across Claude Code, Copilot, and Cursor

Inside the repo that should become a wiki:

/pin-llm-wiki init
/pin-llm-wiki ingest https://github.com/org/repo
/pin-llm-wiki queue https://example.com
/pin-llm-wiki ingest
/pin-llm-wiki lint

After init, follow the generated wiki's AGENTS.md; it is the operating manual for agents working in that knowledge base.

Commands

Use /pin-llm-wiki in the agent. Claude Code, Cursor, and GitHub Copilot all use the same SKILL.md.

SubcommandWhat it does
initScaffold inbox.md, .pin-llm-wiki.yml, AGENTS.md, wiki/, and raw/
ingest [url]Ingest one URL, or omit url to process every pending item in inbox.md
queue <url> ...Add URLs to inbox.md without fetching or ingesting
lintValidate wiki health and apply light non-destructive fixes
remove <slug>Soft-delete a source into wiki/.archive/

Ingest rules, inbox HTML tags, companion repos, and multi-product deep mode live in skills/pin-llm-wiki/SKILL.md and its sibling workflow files.

What gets created

inbox.md              source queue; drop URLs under ## Pending
.pin-llm-wiki.yml     config: domain, detail level, source types, lint cadence
AGENTS.md             canonical instructions for agents in the generated wiki
wiki/
  index.md            start here; full source list
  overview.md         rolling cross-source synthesis
  log.md              append-only ingest, refresh, and removal history
  sources/            one page per ingested source
  .archive/           soft-deleted sources
raw/
  github/             immutable GitHub repo captures
  youtube/            immutable YouTube transcripts + metadata
  web/                immutable web page captures

The wiki is Markdown with [[wikilinks]]; opening the repo as an Obsidian vault is a comfortable way to read and navigate it (start from wiki/index.md).

Demo wiki

For a maintained wiki built with this skill, see ndjordjevic/agentic-ai-wiki.

Source types

Fetches are detail-aware: broader at standard / deep than at brief. Ingest then turns each raw capture into cited wiki pages (ingest.md).

TypeRaw outputSmarts
GitHubraw/github/<org>-<repo>.mdgh: metadata, README, repo layout, and more docs/ at higher detail; <!-- branch --> / optional <!-- clone --> (deep).
YouTuberaw/youtube/<video-id>-<slug>.mdyt-dlp: description, chapters, cleaned transcript (or a no-transcript flag).
Webraw/web/<slug>.mdCrawls landing + docs (plus llms.txt / sitemap where useful); often pulls a companion GitHub repo into one unified page unless <!-- no-companion -->. deep can split multi-product sites: one umbrella + one sub-page per product, still backed by one raw file—similar to how langchain.com becomes a hub plus pages like LangGraph / LangSmith.

GitHub URLs with a path after the repo (/tree/…, /blob/…, …) are single-page web only: no full-repo ingest, companion discovery, or deep product split on that URL.

Agent behavior

Generated wikis include AGENTS.md, which tells AI agents to:

  • Read wiki/index.md before answering domain questions.
  • Follow [[wikilinks]] into relevant source pages.
  • Cite wiki page names in answers.
  • Say when the wiki does not contain an answer, then fetch current information online.

Agents are also instructed not to run git commit or git push unless the human explicitly asks.

Limits

This is a reviewable knowledge workflow, not an unattended publishing system. Generated pages should be inspected in git diffs. Large fetches have token guards, and Phase 1 lint defers contradiction and terminology-collision checks.

Ver o original no GitHub

Bom para

Relacionados