convez.ai
SkillEscrita & ConteúdoIntermediário
sjmoran

Claude Academic Prose

sjmoran/claude-academic-prose

Enforce prosa acadêmica clara, empirista e verificável em papers

Ver no GitHub
InicialPluginScore · de 0 a 100

Uma skill que impõe regras rígidas para escrita acadêmica clara e verificável. Enforça voz empírica plana, remove construções de hype retórico e estrutura argumentos para que cada afirmação seja testável em segundos, incluindo um checklist de pré-submissão automático.

Instalação

$npx convez-ai sjmoran/claude-academic-prose

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
2
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Escrita e revisão de papers acadêmicos e artigos científicos
  • Eliminar linguagem hype e construções retóricas em textos técnicos
  • Estruturar claims com rastreabilidade direta para evidências

Veredito

Ferramenta especializada e opinionada para pesquisadores que priorizam clareza empírica e verificabilidade sobre estilo. Excelente para papers em IR/NLP, mas exige adesão às convenções muito específicas que impõe — não é configurável para outros gêneros acadêmicos.

README

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

claude-academic-prose

A Claude Code skill that enforces clear, evaluable academic prose.

Most paper-writing tools help you produce text. This one constrains how the text is allowed to be written, on the theory that the first judgement of a paper rests on two signals: whether the prose is clear and plain, and whether the claims can be checked against the evidence quickly. The skill encodes both as hard rules.

Install

/plugin marketplace add sjmoran/claude-academic-prose
/plugin install academic-prose@claude-academic-prose

The skill then triggers automatically on any paper-prose request ("tighten the abstract", "write the intro", edits to .tex files), and runs a pass/fail pre-submission self-check before any draft is declared done.

What it enforces

1. Voice — plain-empirical. A register distilled from a corpus of published empirical-IR papers: one claim per sentence, flat "We X" declarative abstracts, problem-first openings, deflated definitions ("A topic model is simply a probability distribution over the vocabulary"), intuition stated beside every piece of formalism, limitations raised mid-argument rather than quarantined. Synthetic exemplar sentences ship with the skill.

2. No hype. Named constructions that put rhetoric ahead of substance are banned and rewritten on sight: em-dash chains, rhetorical inversion, the not-X-but-Y pivot, triadic drama ("It unifies. It predicts. It transfers."), aphoristic titles, intensifier adverbs (crucially / notably / remarkably / surprisingly), stacked parenthetical hedges. The test: any sentence that sounds like a keynote slide or ad copy gets rewritten flat.

3. Evaluability. Structural rules that make the claim→evidence path walkable in thirty seconds: a numbered claims box closing the introduction, each claim pointing to the section and table that tests it; one protocol per headline number (variants live in ablations; the significance-tested number is always the headline); abstracts ≤160 words with at most one number per claim and zero parentheticals; caveats consolidated into two structural homes rather than interleaved — relocation, never deletion.

4. Pre-submission self-check. Eight items, reported pass/fail explicitly before any draft is called finished.

Example

Before (violates the rules six ways — hype constructions, two numbers for one benchmark, footnote-dependent comparison):

It's not just smoothing — it is a fundamentally different estimator. Crucially, our remarkably simple approach unifies. It predicts. It transfers. We evaluate on two collections and, surprisingly, find improvements of 4.2% MAP (p<0.05, paired t-test, but also 5.1% under a different split…).

After:

We study a smoothed estimator of the relevance model for retrieval with sparse queries. The estimator interpolates the normalised query distribution with a corpus background model, and we show that it reduces to Dirichlet smoothing under a particular choice of the interpolation weight. We analyse the behaviour of the estimator on the TREC-8 ad-hoc retrieval task. Our experiments show an improvement of 4.2% MAP over the baseline, significant under a paired t-test at the 0.05 level. The main contribution of this work is an empirical characterisation of when smoothed relevance estimates help sparse queries.

What it deliberately does not do

No citation fetching, math checking, LaTeX hygiene, venue templates, or reviewer simulation — pair it with a mechanics stack for those (e.g. claude-scholar). The skill declares precedence over such tools on anything prose-related: mechanics skills verify and format; this skill owns the words.

License

MIT

Ver o original no GitHub

Bom para

Relacionados