Skills - Agent Skills para Claude
emerose/skills
Skills modulares para pesquisa acadêmica e experimentos científicos rastreáveis
Ver no GitHubColeção de skills (competências) modulares para Claude e outros agentes LLM, empacotados como plugin do Claude Code. Inclui o bibliographer, um gerenciador de biblioteca acadêmica integrado com busca semântica e full-text, e o scientist, um rastreador de proveniência para pipelines de experimentos científicos.
Instalação
npx convez-ai emerose/skills1 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
- Gerenciar e pesquisar biblioteca pessoal de artigos acadêmicos com metadados automáticos
- Rastrear pipelines de experimentos científicos com proveniência completa
- Exportar coleções acadêmicas em BibTeX e formatos padronizados
Veredito

Solução robusta e bem-arquitetada para pesquisadores acadêmicos, com integração nativa ao Claude Code e suporte a múltiplas fontes de metadados. Requer configuração de embedding backend (DeepInfra ou local), mas oferece rastreabilidade completa e busca semântica dentro de PDFs.
README
A tradução para o português deste README ainda está na fila.
Skills
A collection of Agent Skills
for use with Claude and other LLM agents, packaged as a
Claude Code plugin marketplace.
Each skill is a self-contained folder under skills/ with a SKILL.md and any
bundled scripts, references, and assets.
Skills
bibliographer
Manage a personal collection of academic articles: add papers from a DOI, arXiv
ID, PMID/PMCID, Semantic Scholar ID, or a PDF (auto-fetching metadata from
Crossref / arXiv / PubMed / Semantic Scholar / Unpaywall), file the PDFs into a
human-readable author tree, bulk-import a folder, recover metadata for unverified
scans (enrich), run semantic + full-text search inside the papers, export
BibTeX, browse the collection through a generated self-contained HTML viewer
(index.html), and audit the library for correctness. Built on libkit as its
single store (no separate database) via a bundled bib CLI.
Because libkit embeds every document, an embedding backend is required: set
DEEPINFRA_API_KEY (remote, no local model) or install libkit[fancychunk-*] for
local embeddings. See skills/bibliographer/SKILL.md
and .env.example for the available keys.
scientist
Manage a tree of scientific experiments — one folder per experiment — end to end as
a single provenance-tracked pipeline raw → data → analysis → claims. Extract raw
measurements out of CRO files (Excel, GraphPad Prism, Word/PDF/PowerPoint) into tidy
deterministic data/ CSVs; re-derive analysis (EC50/Hill fits, stats, summaries,
figures) from that data; assert grounded scientific claims (each a re-runnable
pytest spec linking a statement to sha-pinned evidence with a strength); index
everything — claims and internal summaries first — into a libkit store for
semantic + full-text search; and trace any result back to the original measurements.
One sci CLI for the deterministic ops, plus a pytest plugin for claims. See
skills/scientist/SKILL.md; for the per-phase detail, the
files under skills/scientist/references/.
Install
As a Claude Code plugin (recommended)
/plugin marketplace add emerose/skills
/plugin install bibliographer@emerose-skills
/plugin install scientist@emerose-skills
Claude Code clones this repo, discovers the skills, and invokes them automatically
when relevant (or manually via /bibliographer:bibliographer / /scientist:scientist).
The bib CLI, as a standalone tool
Install the bundled command with uv so it's on your
PATH everywhere:
uv tool install "git+https://github.com/emerose/skills#subdirectory=skills/bibliographer"
bib --help
Or run it once, without installing, via uvx:
uvx --from "git+https://github.com/emerose/skills#subdirectory=skills/bibliographer" bib --help
No install at all
Every script here is a PEP 723 uv script that
declares its own dependencies, so you can run it straight from a checkout — no
virtualenv, no install:
uv run skills/bibliographer/scripts/bib.py init
uv run skills/bibliographer/scripts/bib.py add arXiv:1706.03762
uv run skills/bibliographer/scripts/bib.py import ~/papers --dry-run
uv run skills/bibliographer/scripts/bib.py query "why do transformers scale"
The scientist skill works the same way — the sci CLI for the deterministic ops
(zero-install via PEP 723), and the claims harness via an ephemeral editable install:
uv run skills/scientist/scripts/sci.py extract "K1-000000 - Potency" # raw → tidy data/
uv run skills/scientist/scripts/sci.py query "dose-dependent gait effect" # semantic search
uv run skills/scientist/scripts/sci.py trace "K1-000000 - Potency" # claim → … → raw
uv run --with-editable skills/scientist pytest "K1-000000 - Potency/analysis/claims"
Layout
.claude-plugin/
marketplace.json # marketplace catalog (lists the plugins below)
plugin.json # this repo, exposed as the `bibliographer` plugin
skills/
<skill-name>/
SKILL.md # name + description frontmatter, then instructions
scripts/ # bundled executable helpers (PEP 723 uv scripts)
references/ # docs loaded on demand
evals/ # test prompts for the skill
tests/ # unit tests for the bundled scripts
pyproject.toml # optional: lets the CLI be `uv tool install`-ed
The repository is simultaneously the marketplace and its single plugin: the
plugin source points at the repo root, and Claude Code discovers the skills
under skills/ automatically.
Using these skills
- Claude Code: install via the marketplace (above). The
descriptionin eachSKILL.mdcontrols when the skill triggers. - Claude.ai / other agents: point your harness at the
SKILL.mdand let the agent run the bundled scripts (oruv tool installthe CLI).
Extending these skills
If you're an agent working on a skill here, read AGENTS.md first: capture lessons as you go, push rote work into code, contribute generic fixes upstream by PR, and keep stateful skills audited.
License
MIT © Sam Quigley