llm-wiki
Phluxxed/llm-wiki
Wiki automático mantido por IA com scaffolding, templates e controle de qualidade
Ver no GitHubSkill que cria um scaffold de wiki persistente mantido por IA, seguindo o padrão de wiki pessoal de Karpathy. Inclui estrutura de diretórios pronta, templates, scripts de linting e avaliação com LLM-as-judge, e histórico de mudanças append-only.
Instalação
npx convez-ai Phluxxed/llm-wiki1 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
- Documentar padrões e convenções de projetos com IA
- Manter histórico estruturado de experimentos e decisões arquiteturais
- Criar bases de conhecimento colaborativas com validação automática
- Organizar runbooks, ADRs e papers em formato único
Veredito

Implementação sólida do padrão Karpathy com tooling robusto (linting, eval, render). Ideal para equipes que querem documentação estruturada mantida por agentes, mas exige Python e familiaridade com scaffolding de projetos.
README
A tradução para o português deste README ainda está na fila.
llm-wiki
A skill that scaffolds a persistent, LLM-maintained wiki — inspired by Karpathy's personal wiki pattern.
Drop it into your agent's skills directory, then run /wikime in any project to get a fully operational wiki scaffold in seconds.
What it sets up
your-wiki/
├── wiki-agent.md ← agent operating manual (all instructions live here)
├── CLAUDE.md ← @wiki-agent.md (or natural-language pointer for other agents)
├── CONVENTIONS.md ← human-readable naming and structure reference
├── README.md ← quick start and scripts reference
├── index.md ← one-liner catalog of all wiki pages
├── log.md ← append-only change history
├── {page-type}/ ← primary wiki pages (e.g. papers/, use-cases/, experiments/)
│ └── *.md
├── entities/ ← entity and concept pages
│ └── *.md
├── _templates/
│ ├── {page-type}.md ← template for new pages
│ └── entity.md ← template for entity/concept pages
├── sources/ ← immutable raw inputs (never edited after saving)
└── scripts/
├── render.py ← generates wiki.html — single-file reader with eight views
├── lint.py ← structural health check
├── query.py ← frontmatter queries (filter by status, category, tag, etc.)
└── eval.py ← LLM-as-judge quality eval (grounding, contradictions, redundancy) with regression gating
Installation
# Clone the repo
git clone https://github.com/Phluxxed/llm-wiki ~/llm-wiki
# Symlink into your agent's skills directory (recommended — keeps it version-controlled)
# Claude Code example:
ln -s ~/llm-wiki ~/.claude/skills/wikime
Each agent has its own skills directory — your agent will know where to look.
Requires pyyaml and markdown for the scripts:
pip3 install pyyaml markdown
scripts/eval.py's claude judge additionally needs claude-agent-sdk (pip3 install claude-agent-sdk); skip it if you run the eval with --judge codex or --judge none.
Usage
In any directory (empty or existing project), run:
/wikime
The skill asks two questions — what the wiki is for, and what the primary page type is — then scaffolds everything in one pass.
Supported page types: use cases, papers, experiments, runbooks, ADRs, or anything else you name.
Safe on existing projects: if a CLAUDE.md, AGENTS.md, or GEMINI.md already exists, the skill appends a single pointer line rather than overwriting it. If wiki-agent.md already exists, the skill stops — you already have a wiki.
Scripts
| Script | Command | Output |
|---|---|---|
scripts/lint.py | python3 scripts/lint.py | Structural health check — missing sections, broken refs, open risks, index consistency |
scripts/query.py | python3 scripts/query.py --help | Frontmatter queries — filter by --status, --category, --type, --tag, --stale, --risks |
scripts/render.py | python3 scripts/render.py | Generates wiki.html — single-file reader (Home, Page, Search, Graph, Risks, Recent changes, Open questions, Entities). Open in a browser or view as a Claude artifact |
scripts/eval.py | python3 scripts/eval.py --gate | LLM-as-judge quality eval — grounding, cross-page contradictions, redundancy, near-duplicate disambiguation; per-metric thresholds + regression gating (exit code). Auto-detects your agent CLI (claude/codex) as a keyless judge; run records in .eval/ |
Agent compatibility
Works with any agent that supports skills. The running agent creates its own schema file — a one-line pointer to wiki-agent.md where all instructions live.
| Agent | Schema file | Pointer format |
|---|---|---|
| Claude Code | CLAUDE.md | @wiki-agent.md (native include) |
| Codex | AGENTS.md | Natural-language pointer |
| Gemini CLI | GEMINI.md | Natural-language pointer |