---
name: academic-style
description: Style rules for academic prose in ML / computational biology venues. Loaded by draft-writer and prose-polisher to ensure consistent voice and avoid LLM-typical patterns.
---

# Academic Style Rules

This skill encodes the writing style preferred for ML / computational biology papers. It is loaded by `draft-writer` and `prose-polisher`. Rules here SUPPLEMENT those in CLAUDE.md, they do not replace them.

## Voice

<!-- TODO: 主人填具体偏好。建议：
- We vs. Authors vs. passive: Use "we" for actions taken by you/your team. Use passive only when the actor is genuinely irrelevant ("Sequences were aligned using MAFFT").
- Tense: Past for methods/results ("We trained..."). Present for established facts ("BGCs encode..."). Present for figure references ("Figure 3 shows...").
- Hedging: Use "suggests" only when evidence is genuinely indirect. Avoid "may potentially", "could possibly", stacked hedges.
-->

## Sentence Construction

- Prefer 15-25 word sentences. Break anything over 35 words.
- Lead with the subject of interest, not with throat-clearing.
  - BAD: "It is interesting to note that bag-of-Pfam baselines outperform learned embeddings."
  - GOOD: "Bag-of-Pfam baselines outperform learned embeddings."
- One idea per sentence. If you find yourself writing "; furthermore," consider splitting.

## Paragraph Construction

- Do not start every paragraph with a topic sentence. Sometimes lead with the finding, then unpack.
- Paragraphs of 4-7 sentences are typical. Single-sentence paragraphs are acceptable for emphasis, not by default.
- Avoid the rigid "First, ... Second, ... Third, ... Finally," scaffolding more than once per section.

## Forbidden Phrases (Replace These)

| Forbidden | Why | Replace with |
|-----------|-----|--------------|
| "It is important to note that" | Throat-clearing | (delete entirely) |
| "delve into" | LLM signature | "examine", "study" |
| "leverage" (as verb) | Corporate-LLM signature | "use", "apply" |
| "in the realm of" | Pretentious | "in" |
| "shed light on" | Cliché | "reveal", "show" |
| "comprehensive" (as filler adjective) | Unsupported claim | (delete or be specific) |
| "robust" (as filler adjective) | Same | (delete or specify what's robust) |
| "It should be noted that" | Same as #1 | (delete) |
| "play a [pivotal/crucial/important] role" | Vague | Be specific about the role |
| "tapestry", "intricate", "myriad" | LLM tells | (rephrase concretely) |
| "In conclusion," (mid-paper) | Premature closure | (delete) |

## Numerical & Citation Format

<!-- TODO: 主人按目标 venue 调。默认建议：
- Numbers: spell out one through nine, use digits for 10+
- Statistics: "p < 0.001" with space, italicized p
- Effect sizes: report alongside p-values (Cohen's d, η², AUC, etc.)
- Citations: (Author et al., 2024) for parenthetical, Author et al. (2024) for narrative
- Multiple citations: alphabetical order within parens (Smith, 2020; Wang, 2024) — verify with venue
-->

## Figure & Table References

- "Figure 3 shows X" — strong, direct
- "(Fig. 3)" — parenthetical, when X is the focus and figure is supporting evidence
- AVOID: "As can be seen in Figure 3", "It can be observed in Figure 3 that..."

## Domain Terminology Discipline

- Define a domain term at first use unless target venue is specialist.
- For BGC / natural product papers: Pfam, antiSMASH, MIBiG, BiG-SLiCE, CORASON, etc. — the venue determines whether you need to define.
- Do not use jargon as decoration. If a simple word works, use it.

## Things to Preserve (Do NOT "Polish Away")

- The user's specific phrasings that carry argumentative weight
- Hedging that genuinely reflects uncertainty (not all hedges are bad)
- Repetition that serves emphasis ("not just X — X *and* Y")
- Sentence fragments used for emphasis (rare, but valid)