如何一句话让Claude Scientific Writer写出你要的论文?20个实战Prompt模板(附完整示例)
【免费下载链接】claude-scientific-writerA general purpose scientific writer项目地址: https://gitcode.com/gh_mirrors/cl/claude-scientific-writer
Claude Scientific Writer是一个开源的 AI 科研写作工具(deep research and writing tool):你只需一句话,它先自动检索文献、逐条核实引用,再为你生成可发表的科研论文、基金申请书、会议海报、文献综述和临床报告,默认输出 LaTeX + PDF。本文整理了20 个可直接复制的实战 Prompt 模板,帮你零基础写出带真实引用的论文初稿。
1分钟上手:Claude Scientific Writer 安装三步
写 Prompt 前,先确保工具可用。三种安装方式里推荐 Claude Code 插件(README.md 的 Quick Start 一节有完整说明):
插件方式(推荐):在 Claude Code 中添加插件市场并安装:
/plugin marketplace add https://gitcode.com/gh_mirrors/cl/claude-scientific-writer /plugin install claude-scientific-writerpip 安装(CLI/API):
pip install scientific-writer源码安装:
git clone https://gitcode.com/gh_mirrors/cl/claude-scientific-writer后运行uv sync
然后在你的项目目录执行初始化命令(它会写入 CLAUDE.scientific-writer.md 等指令文件并激活全部 26 个技能):
/claude-scientific-writer:scientific-writer-init初始化完成后,你的项目里会出现CLAUDE.md和AGENTS.md,之后所有 Prompt 都直接"说人话"即可。
万能公式:一条好 Prompt 的 5 个要素
对比 README.md 里的所有官方示例,你会发现高效果 Prompt 都包含这 5 个要素。记住它,20 个模板都是它的应用:
| 要素 | 作用 | 示例措辞 |
|---|---|---|
| 📄 文档类型 | 决定模板与结构 | "Create aNature paper/NSF proposal/conference poster" |
| 🎯 主题 | 研究领域 | "on CRISPR gene editing" |
| 📊 数据引用 | 让结果有真实支撑 | "Present knockout_efficiency.csv (5 cell lines)" |
| 🖼️ 图片指定 | 明确插图位置 | "Include Western blot (figure1.png) and flow cytometry (figure2.png)" |
| ⚖️ 对比/基准 | 凸显创新性 | "Compare with published Cas9 benchmarks (70-75%)" |
💡 小技巧:把数据文件和图片提前放进项目的
data/目录,工具会自动把图片分拣到figures/、数据保留在data/(见 docs/FEATURES.md 的 Data & File Integration 一节)。
模板组 1:科研论文写作(模板 1-4)
对应技能:skills/scientific-writing/,自动遵循 IMRaD 结构(引言-方法-结果-讨论,摘要最后写)。
模板 1|带数据图的完整论文(最常用)
Create a Nature paper on CRISPR gene editing. Present editing_efficiency.csv (5 cell lines, n=200 each), include Western blot (protein_knockout.png) showing target depletion, and flow cytometry (editing_percentages.png) with 87% efficiency in HEK293. Compare results to published Cas9 benchmarks (70-75%).模板 2|指定会议的 ML 论文
Write a NeurIPS paper on transformer architectures, following NeurIPS formatting, with about 30-45 verified citations.模板 3|结合最新文献的论文(自动触发实时检索)
Create a paper on recent advances in quantum computing (2024). Compare published values with our gate_fidelity_results.csv (99.2% for 2-qubit gates) and cite papers achieving >98% fidelity.模板 4|指定图表插入位置
Create a NeurIPS paper on deep learning optimization. Include training_curves.csv showing convergence after 50 epochs, accuracy_comparison.png (our method 94.2% vs baseline 89.1%) in the results, and architecture_diagram.svg in the methods.📖 论文写作原则可查阅 skills/scientific-writing/references/writing_principles.md。
模板组 2:基金申请书写(模板 5-7)
对应技能:skills/research-grants/,内置 NSF / NIH / DOE / DARPA 四大机构的专用格式与评审标准。
模板 5|NSF 基金 + 初步数据
Write an NSF proposal for quantum computing research. Present preliminary data from gate_fidelity.csv showing 99.2% fidelity (vs 95% state-of-art). Include circuit_topology.png, error_rates_comparison.png, and scalability_projections.csv for 100-qubit systems.模板 6|NIH R01 Specific Aims(一页纸目标)
Generate NIH R01 Specific Aims (1 page) for cancer immunotherapy research, with preliminary data from pilot_study_results.csv.模板 7|带时间线与预算的完整申请
Generate an NSF grant proposal presenting preliminary data from quantum_results.csv (99.2% gate fidelity), circuit_topology.png, and error_rates.csv. Include a 5-year timeline with milestones_budget.xlsx.📖 机构指南:skills/research-grants/references/nsf_guidelines.md、nih_guidelines.md。
模板组 3:海报与学术演示(模板 8-10)
对应技能:skills/latex-posters/(默认海报格式)与 skills/scientific-slides/。
模板 8|A0 会议海报
Generate an A0 conference poster. Highlight findings from efficacy_study.csv (n=150 patients, 40% response rate). Feature mechanism_diagram.png, survival_curves.png, biomarker_heatmap.png, and statistical_forest_plot.png (p<0.001 primary endpoint).模板 9|从已有论文生成海报
Generate a conference poster from my paper. Feature dose_response_graph.png as the central figure and emphasize the p<0.001 significance for the primary outcome.模板 10|限时演讲幻灯片
Create a 15-minute conference presentation on my machine learning research, gathering 8-12 verified citations for the background and discussion.模板组 4:文献综述与实时检索(模板 11-13)
对应技能:skills/literature-review/ 与 skills/research-lookup/。
模板 11|带元分析的系统综述
Create a systematic review on AI in drug discovery. Reference studies_database.csv (127 papers, 2020-2024). Include success_rates_meta.png (pooled OR=2.3, 95% CI 1.8-2.9) and show oncology dominance (45% of studies) in the synthesis.模板 12|@ 唤起技能做文献调研
@research-lookup Find papers on mRNA vaccine efficacy (2022-2024). Compare with our trial_outcomes.csv (n=500, 94% efficacy) and antibody_titers.png.模板 13|事实核查式提问
What are the current success rates for CAR-T therapy in B-cell lymphoma? Compare with our clinical_trial_outcomes.csv (n=45, 62% complete response) and published JULIET and ZUMA trials.模板组 5:临床与医学文档(模板 14-16)
对应技能:skills/clinical-reports/(12 套专业模板 + 8 个自动校验脚本,内置 HIPAA 去标识化检查)。
模板 14|罕见病病例报告(CARE 规范)
Create a clinical case report for a rare mitochondrial disease. Include patient_timeline.csv (6-month progression), diagnostic_scans/ folder (MRI, CT), and lab_values.xlsx showing elevated lactate (8.2 mmol/L) and creatine kinase (450 U/L). Describe treatment response.模板 15|治疗方案
Create a treatment plan for a 55-year-old patient with newly diagnosed type 2 diabetes and hypertension, with SMART goals, monitoring schedule, and follow-up plan.模板 16|队列分析(按生物标志物分层)
Analyze a cohort of 45 NSCLC patients stratified by PD-L1 expression (<1%, 1-49%, >=50%), including ORR, PFS, and OS outcomes with statistical comparisons.模板组 6:同行评审与质量把关(模板 17-18)
对应技能:skills/peer-review/ 与 skills/scholar-evaluation/(ScholarEval 8 维度量化评分)。
模板 17|ScholarEval 框架评审
Evaluate this manuscript using ScholarEval. Reference power_analysis.png (n=30, underpowered), assess methodology against CONSORT standards, and verify citations match claims.模板 18|投稿准备度评估
Assess publication readiness for Nature Machine Intelligence: score all 8 dimensions, list weaknesses first, and give prioritized revision recommendations.✅ 其实不用特意写:CLAUDE.md 规定每次生成后会自动进行一轮同行评审,并输出
PEER_REVIEW.md。
模板组 7:文件操作与持续修改(模板 19-20)
工具支持"智能论文识别"——你用口语化关键词即可定位已有文档,无需给路径(机制见 docs/FEATURES.md 的 Intelligent Paper Detection)。
模板 19|继续编辑已有论文
Add a methods section describing the protocols used to generate binding_assay.csv data, including statistical tests (t-tests in stats_summary.csv) and sample size justification from power_calculation.xlsx.模板 20|查找并补充指定论文
Find the CRISPR paper and add discussion of limitations shown in off_target_analysis.csv and efficiency_variation.png across different cell types.常用"续写/检索"关键词:continue、update the paper、find the <主题> paper、new paper on <主题>。
写 Prompt 的 4 条进阶技巧
- 数字要具体:写 "87% efficiency in HEK293, p<0.001" 而不是 "效果很好"——系统会严格区分观察值、估计值与推测(writing_principles.md 的准确性原则)。
- 引用零容忍造假:系统对每一篇引用都会走 research-lookup 核实,找不到真实论文就整条丢弃,所以放心让它"多引一点"(Nature 类目标 35-50+ 篇)。
- 图越提越明确:指定文件名 + 应出现的章节,工具还会自动补充图形摘要(graphical abstract)等插图。
- 长文档不中断:系统指令明确"绝不中途询问是否继续",50+ 页市场报告也会一次写完整(CLAUDE.md 的 CRITICAL COMPLETION POLICY)。
产出在哪里?目录结构与常见问题
每次任务的输出都在独立目录,结构固定(CLAUDE.md 的 File Organization):
writing_outputs/<时间戳>_<主题>/ ├── drafts/ # v1_draft.tex、v2_draft.tex(版本号自动递增,从不覆盖) ├── references/ # references.bib(全部引用经核实的 BibTeX) ├── figures/ # 论文插图 ├── final/ # manuscript.pdf + manuscript.tex ├── sources/ # 所有文献检索结果(可追溯的审计留痕) └── progress.md / SUMMARY.md / PEER_REVIEW.md新手最常问的 3 个问题
- 引用是编造的吗?不会。每条 BibTeX 元数据(作者、卷期页码、DOI)都要通过在线核实,验证失败的引用会被直接丢弃。
- 输出格式能改吗?默认 LaTeX,你可以要求 Word、PPTX 海报等格式(PPTX 海报需显式说 "PPTX",见 skills/pptx-posters/)。
- 遇到问题怎么办?先查 docs/TROUBLESHOOTING.md,全部能力清单见 docs/SKILLS.md。
写在最后:Claude Scientific Writer 把"先检索、后写作、逐条核实引用"做成了默认流程。你只需要像给同事交代任务一样说一句话——文档类型、主题、数据文件、对比基准,剩下的 LaTeX 排版、图表生成和同行评审它会全包。现在就从模板 1 开始试试吧!🔬
【免费下载链接】claude-scientific-writerA general purpose scientific writer项目地址: https://gitcode.com/gh_mirrors/cl/claude-scientific-writer
创作声明:本文部分内容由AI辅助生成(AIGC),仅供参考