Scientific Voice: Submission-Ready Prose, English & Spanish

What it does

Reviewers read your draft once. This skill makes your agent edit scientific prose so that once is enough: the finding comes first, each claim carries one hedge sized to its evidence, sentences about the text itself disappear, and every vague word gives way to the number it stands for. It works in English and Spanish, including a paper with an English Results section and a Spanish "Resumen" in the same file.

• Three modes. A full rewrite with a change log, a light edit shown before → after, or a review that only lists findings, ordered by impact.
• 16 voice rules, applied in order. Claim before caveat, one point per paragraph, numbers with comparators and inline tests, plain reporting of what did not work, one hedge per claim, precise words instead of hype, verbs instead of nominalisations, sentence length, tense, and one term per concept.
• Spanish done in Spanish. Spanish hedges (including the softening conditional), "se realizó un análisis de" turned back into "analizamos", "cabe destacar que" cut, false friends such as "eventualmente", "asumir que", "consistente con" and "billón" fixed, and one decimal separator throughout.
• Meaning is fixed. Numbers, units, citations, math and your technical terms are never changed, and a claim is never strengthened. What it cannot fix without guessing gets a visible marker such as [test?] and a question for you.
• Your own recurring errors. A plain-text personal patterns file the linter checks on every run; when you accept the same correction twice, the agent offers to add it.

The bundled linter

prose_lint.py (Python 3.10+, standard library, no network, writes nothing) checks Markdown, LaTeX, reStructuredText and plain text, and decides each paragraph's language on its own. It flags hedge stacks, meta-commentary, vague intensifiers, promotional words, "significant" with no test, buried verbs, wordy phrases, Spanish false friends, mixed US/UK spelling, decimal commas in English and mixed separators in Spanish, long sentences and paragraphs, doubled words and leftover markers. House preferences (spelling variant, decimal separator, em-dash ban, banned words, limits) live in a small JSON file. In the demo it found 31 tics in one English paragraph and one Spanish abstract; after the rewrite, only one long sentence kept on purpose and three gap markers remained.

What's included

• SKILL.md: modes, house preferences, a 12-pass rewrite order, hard constraints and the report format
• references/voice-rules.md: the 16 rules with before/after examples
• references/spanish.md: hedges, impersonal and passive forms, wordy phrases, false friends, numbers and units, bilingual abstracts
• references/lint-rules.md: every rule, configuration, personal patterns and what the linter cannot judge
• scripts/prose_lint.py, assets/prose-lint.example.json and assets/personal-patterns.txt

Part of the Aurora Lab research pipeline

Works entirely on its own on any text you point it at. Pairs well with Aurora Lab's Claim Check (does each sentence stay inside its evidence?), Journal House Style (does the paper fit the venue?) and Test Readers (how a cold reader takes it), but needs none of them.