Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› eacl-writing-style
eacl-writing-style
GitHub用于修订EACL论文，确保首页清晰呈现任务、差距、贡献及量化结果。规范LLM声明边界与错误分析，检查匿名性，优化多语言表述，并指导在篇幅限制下保留核心论证，避免将关键内容推入附录。
触发场景
需要修改EACL论文以符合首页叙事结构
需要审查论文中的过度主张或匿名性泄露风险
需要在有限篇幅内压缩内容并强化量化证据
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill eacl-writing-style -g -y
SKILL.md
Frontmatter
{
"name": "eacl-writing-style",
"description": "Use when revising an EACL paper's prose for a task-first first page, a concrete page-one example, scoped LLM-era claims, quantified error analysis, an anonymity-safe voice, honest Limitations, and compression into the 8-page long or 4-page short content budget without pushing the argument into appendices."
}
EACL Writing Style
Use this to revise an EACL paper so its contribution is legible fast and its claims are scoped.
EACL reviewers read many papers in a short window; the ones that land put the task and the
result on the first page, quantify rather than assert, and name their limits. Pair this with
the worked example in ../../resources/worked-examples/01-introduction.md.
The EACL first-page arc
- Task — the specific problem, in the first breath, not "great progress in NLP."
- Gap — why current methods fall short, each reason nameable.
- What we do — the contribution, stated plainly.
- Measured result — a number tied to a table, with variance.
- Honest scope — what the result does and does not cover.
Habits to cut, habits to keep
| Cut | Keep |
|---|---|
| "Achieves strong performance" | "Improves F1 by X (95% CI ...) over baseline B" |
| Generic "prior work is limited" | A specific failure per cited approach |
| A concrete example only on page 5 | A worked example on page 1 |
| Unscoped "our method generalizes" | "On the six languages tested; see Limitations" |
| Roadmap standing in for an argument | A one-line roadmap after the argument |
Scope the LLM-era claim
- If the paper uses or evaluates LLMs, bound the claim to the models, prompts, and settings tested, and disclose contamination risk. An unscoped "LLMs can/cannot do X" invites the reviewer to name the counterexample.
- Report prompts and decoding as part of the method, not as trivia (see
eacl-reproducibility).
Quantify the error analysis
- A page-one or early-section error analysis with counts ("40% of errors are agreement errors; examples in Table 3") is worth more than adjectives. EACL rewards papers that show where and why a system fails, especially across languages.
Anonymity-safe voice
Anonymity check before submission:
- no author names, affiliations, or acknowledgements
- no "as we showed in our EMNLP 2025 paper" -> use third-person citation
- no links that identify authors (personal repos, named grant pages)
- self-citations phrased neutrally
Multilingual clarity
- Name languages and scripts explicitly; render diacritics correctly in the PDF and later in the
Anthology metadata (
eacl-camera-ready). - Do not let an aggregate multilingual score stand in for per-language honesty — a table beats an average.
Compression discipline
- The content pages carry the argument; appendices carry detail. If cutting for length pushes
a core claim into an appendix, cut something else instead (see
eacl-supplementary). - The Limitations section is free space and read — use it to state scope, not to hide results.
Output format
[First-page arc] Present / Missing elements: <task/gap/what/result/scope>
[Overclaims] <phrases to scope, with fix>
[Evidence pairing] <claims lacking a table/number>
[Anonymity] <any leak>
[Multilingual honesty] <aggregate-hiding issues>
[Compression] <what to cut so the body carries the claim>
版本历史
- 9f86f09 当前 2026-07-19 15:15
