从去年年底开始就在思考着如何构建一个辅助日常阅读与写作的 Skills 集合,然后逐渐地开始想要搭建一些 MCP 本地服务出来。但一直没有一个全貌,这样架构究竟应具备怎样的形态,这和之前的提示词有什么区别,和 Skills 又有哪些不同。
截至本文撰写时,写了两年多提示词的我,认为提示词是当下交给模型的简单任务,让模型回答某一个指定的问题。而 Skills 呢,则是设置了一系列的流程和任务,并且通过细致的指导(领域知识)让模型在细分任务上能够更专业地解决问题。
今天在 github 上看到 K-Dense 团队打造的138 项即用型科学技能 的全面集合。粗略扫了一眼就觉得可以用来作为一个初步的分析样本,以此优化目前对 Skills 定义中那些依然晦涩、混沌的认知。。
项目 README 中表述得十分清晰:有 138 项技能,另外还有 28+ 数据库,55+ Python 包,这已不再仅仅是 Skills 的集合,而是一个完整的项目体系。。
- MCP 可以托管在 K-Dense 团队服务器上
- 也可以在自建的 MCP 服务器上
如果这样,那必定也可以运行在部署的 Docker 服务器上了。
接下来开始介绍一个科研数据分析全流程:
- 明确目标与交付物:到底要解决什么问题,最后要交付什么东西(图表、结论、报告、代码)。
- 收集数据(内部 + 外部):内部:实验或临床的原始文件;外部:公共库和文献。形成一个数据清单。
- 质控与清洗:ETL,产出一套干净数据,还有 QC 记录。
- 注释与整合:形成能互相对得上的合并数据表。
- 核心分析与建模:排序打分、差异表达、分组预测、找临床试验,检验结论。
- 验证与交叉核查:测试、复现、文献对比、校验等。
- 可视化与解释:让人一眼看懂。
- 报告与归档:摘要、方法简述、主要发现、局限性、下一步等。
这个示例很重要的是不仅给出了详细流程,还给出了内外部数据与工具:怎么找、怎么用。譬如外部数据分为了化学、结构、基因/蛋白等等,那么分别对应的数据如何访问如何查询等都已实现。
并且按用途给出了常用工具,如统计与机器学习领域使用的 statsmodels、scikit-learn…… 文档与图表使用matplotlib、seaborn、ReportLab。
更关键的是,还给出了场景化通用方法,以如何寻找肺癌抑制剂为例:
- 找出哪些细胞类型、哪些基因在不同条件下有变化
- 预测患者预后或分层,找稳定的组合特征
- 找到能与蛋白发生相互作用的小分子
- 临床评估遗传性癌症风险
- 然后用系统生物学网络分析
- 最后形成报告让信息好看又好懂
总结为一张表:流程、动作、工具、产出
| 流程阶段 | 在做什么 | 常用工具/数据库 | 典型产出 |
|---|---|---|---|
| 1. 明确目标 | 写清问题与交付物 | — | 目标与指标 |
| 2. 收集数据 | 内部/外部数据清单 | 10X/VCF/mzML、ChEMBL、ZINC、ClinVar、NCBI Gene | 数据目录 |
| 3. 质控清洗 | 过滤坏数据、统一格式 | Scanpy、pyOpenMS、pysam | 干净数据 |
| 4. 注释整合 | 映射到基因/蛋白/通路 | UniProt、KEGG、Reactome、STRING | 合并表 |
| 5. 核心分析 | 差异/对接/建模 | RDKit、DiffDock、PyDESeq2、scikit-learn、Arboreto | 结果与模型 |
| 6. 验证核查 | 复现与稳健性检查 | statsmodels、PyMC、外部数据(GEO/Cellxgene) | 验证结论 |
| 7. 可视化解释 | 图表+通俗说明 | matplotlib、seaborn | 可视化图 |
| 8. 报告归档 | 写报告、打包复现 | ReportLab、README/配置 | PDF/代码包 |
按这八步走,把数据从哪里来、怎么变干净、怎么得到结论、如何让人看懂并复现讲清楚,再配上对应工具与证据,就能把复杂的生物学科研流程形成人人能懂的流程。
所以,这样的一个 Skills,实际上并不是一个简单的模型提示词的交互,更不是说一次简单的数据分析。更像是一个完整的项目。
- AI 能够结合内外部现状形成最佳方案;
- AI 能够获取内外部数据,并且能够对内外部结构&非结构化数据进行 ETL 处理,形成高质量的数据。
- AI 能够根据方案进行分析、标注、建模、反馈、验证等等工作,实现 PDCA 流程的闭环。
即便是 Gemini 3、GPT 5.2、Opus 4.5 这种具备深度思考(Thinking)能力的模型,针对很多有一定难度的问题,也很难一次性解决。显然,这种以结构化编程步骤为核心的方法是完全可行的,而且很多人,如 Every 的 CEO 实际经验证明这样的流程可以提升一位专家工程师 5 倍以上的速度。虽然,当前的还需要人参与在流程中做方案确认、经验的积累。已然是大幅提升效能的神器了。
实际例子分析
既然是一篇学习笔记/文章,不能没有实际 skill 分析。计划分析下明确目标、收集数据,核心分析,验证核查四个步骤。
原因在于,如何明确目标是一个大家都知道的结合内外部信息让最佳实践落地过程,但具体来说如何利用 LLM 形成最佳实践呢?收集数据是一个需要定义什么数据对当前有用,如何收集更为高效?核心分析涉及到的建模与循环过程,这中间是否需要人的参与?最后的验证,是关于如何核查才是当下顶尖大模型的最佳实践?
首先下载这个 repository 到本地,然后用一个程序把这个 repository 整合为一份文本文件,方便放到 Gemini3进行详细分析。
方案撰写 Skill
这一步需要写清问题与交付物,是通过 Research Methodology & Planning(研究方法与规划) 和 Scientific Communication(科学传播) 这两大类 Skills 中的多个工具协同完成的。
先看下 Research Methodology & Planning,
结构上采用固定双层结构:YAML 头部元信息,这样方便机器可读,然后是 Markdown 正文方便人类阅读。 Markdown 正文骨架从概述开始,回答问题并阐述价值,整个 Skill 详细涵盖了背景、触发、约束、任务、评审、策略、执行、资产及迭代的每一个环节。
关键章节与最佳实践对照表
| 章节 | 作用 | 最佳实践 | 常见坑 |
|---|---|---|---|
| 概述 | 定义问题与价值 | 第一段回答“为谁、在何时、为何重要” | 抽象空泛、无用户收益 |
| When to Use | 触发边界 | 正/反向触发并列;可量化前提 | 无边界、处处适用 |
| 约束 | 合规与硬限制 | MUST/SHOULD/MAY 分层;可视化最低配 | 约束隐性、分散难遵循 |
| 核心组件 | 执行主体 | 统一子模板;列验收标准 | 仅列清单,无可执行性 |
| 评审标准 | 反向对齐 | 用评审问题驱动写作;给证据样例 | 只罗列,不可实操 |
| 工作流 | 时程化管理 | 分阶段“活动/产出/时间”三联表达 | 仅甘特图,无验收口径 |
| 资源资产 | 复用与降本 | 模板/脚本/参考分目录管理 | 链接失效、冗余重复 |
| 迭代策略 | 提升与再投 | 逐条回应、变更可见、设不再投条件 | 情绪化反驳、无结构 |
- 技能以主要面向提交给 NSF、NIH、DOE、DARPA 的资助申请为要求,但并非局限在这四家,主要目的是为了满足 FOA/BAA 与指南进行合规适配。
- 严格说明了什么时候使用这个 Skill,譬如准备项目说明、乃至个人简历、个人履历。
- 强制必须要使用 scientific-schematics 技能生成的 AI 图,Nano Banana Pro 将自动生成、审阅并优化示意图。画的图非常具体:如工作流程图、项目甘特图、概念框架插图、系统架构图、实验设计流程图、影响活动图、协作网络图,乃至复杂的概念
- 接下来开始介绍各个机构的不同要求,从机构选择—叙事适配—合规落地。
- 研究提案的核心组成从项目摘要开始,包含了基本要素、写作策略、常见错误避免。
- 开始项目研究策略说明,包括背景、研究目标、计划与方法、里程碑
- 具体目标需要定义明确、可检验的目标
- 更广泛影响
- 创新
- 研究方法与路线
- 初步数据与可行性
- 时间线、里程碑与管理计划
- 团队资历与协作
- 10.预算说明
- 各机构评审标准 详细描述各个机构评审的标准。如重要性、创新性、方法等
- 撰写具有竞争力提案的原则,乃至包括再次提交、常见错误等等。
- 多技能写作
- 科学写作:用于清晰、有说服力的文稿
- 文献综述:用于全面的背景部分
- 同行评审:用于提交前的自我评估
- 研究检索:用于查找相关引文与既往工作
- 数据可视化:用于创建有效图表
- 其他技能,引用和介绍了 references 目录下边 11 个技能,如创建时间线和里程碑、呈现高效团队、回应评审等等。
数据获取 skill
数据获取主要是使用 literature-review 和 research-lookup (Perplexity) 技能。不仅是搜关键词,还能进行系统性综述,去 PubMed, bioRxiv, Google Scholar 抓取最新的论文和预印本。 这里只大致分析下 literature-review Skill。
如何打通内部数据和外部数据,这不仅扩大了数据获取的范围,更是在编写代码(Coding)时参考最佳实践,从而产出高质量成果的必要前提。对于 K-Dense 这个项目,它不只是搜索外部数据,而且直接调用 API 访问专业数据库,包括生物信息、专利查询等。然后通过 document-skills 解析 PDF/Word/Excel 等非结构化数据,提取你上传的实验记录或报告中的文字和表格,最终把用户的数据和世界的数据对齐。
其核心难点在于如何实现内外部数据的标准化映射与对齐通过映射与标准化对齐,能够通过多模态整合多维数据,并且还具备了初步的智能代理编排能力。譬如用户一句模糊信息输入 “分析这个癌症数据集”,就需要调用 Python 清洗和关联外部数据,得出综合结论。
那么回到 literature-review 这个 Skill 上,我们看看这个 Skill 的构成和上边的方案撰写有哪些区别?
首先,如果说 research-grants 科研基金申请的核心是说服与推销,那么 literature-review 文献综述的核心是系统与综合。它要求客观、全面、可复现地搜集证据,并进行公正的评判和总结。
其次,这个 skill 的规划与界定思维模式完全相反。 research-grants 需要组建团队,所以需要分析、吸引评审的痛点,进行差距分析。 而 literature-review 强调人群、干预、对照、结果,所以需要设定严格的纳入与排除标准。
然后,Research Grants 是针对性查找支持自己假设的文献,或查找竞争对手的弱点。而 literature-review 是多数据库联合检索如 PubMed, arXiv, bioRxiv 等,使用复杂的布尔逻辑。
关键一步,QA 质量评估。literature-review 需要使用 cochrane 或 Newcastle-Ottawa 量表给别人的研究打分,然后再进行综合与分析,要横向对比不同研究的共识与分歧。
最终交付专业的 PDF 或 Markdown,包含目录、正文、引用列表。
literature-review Skill 大纲
- 首先界定这是一个检索多个文献数据库,按主题综合研究结果核验准确性返回文档的 skill。
- 说明使用场景,如文献综述、论文、未来研究方向等
- 示意图生成 / Nano Banana Pro
- 规划与界定范围 定义清晰的问题、范围、策略等
- 多数据库检索 使用 gget 命令根据研究领域选择合适数据库,通过关键词检索
- 筛选与选择
- 初步数据质量评估 关键时间、作者、主题等初级评估
- 数据获取方法:
- arXiv 用API 或 WebFetch 访问
- Semantic Scholar可以使用免费的接口
- 其它专项数据库
- 最佳实践
- 常见误区、工作流样本
- 与其他技能的集成 包括数据库访问、分析可视化、写作等
- 其他可参考的数据源
- 详细定义了需要用到的 python 包,以及如何安装
Statsmodels 统计建模 skill
Statsmodels提供从估计、推断到诊断的一整套工具,覆盖广泛的统计方法。可以完成从简单线性回归到复杂时间序列模型与计量经济学分析。
在 Quick Start Guide 中直接给出了示例程序:
- 线性回归(OLS)
- 逻辑回归(Binary Outcomes)
- 时间序列(ARIMA)
- 广义线性模型(GLM)
- 线性回归模型针对连续型因变量,支持多种误差结构的线性模型。
- 广义线性模型(GLM)将线性模型拓展到非正态分布的灵活框架。
- 离散选择模型用于分类与计数型结果。
- 时间序列分析 可以针对全面的时间序列建模与预测能力。
- 统计检验与诊断用于模型验证的广泛检验与诊断能力。
然后提供了一套如数据准备、模型构建、推断、模型评估和报告的注意事项。 参考文档中有线性回归模型、广义线性模型、离散结果模型等详细指南,以及常见陷阱与如何避免。
K-Dense-AI 这类项目将诸多技能硬编码在单一仓库中,但是当前 MCP 生态已经成熟,业界已经开始把 MCP servers 当成可插拔技能。这也是我们正在经历从语言对话到工具调用,再到系统集成 / MCP的范式转移。
2026 未来已至,后续计划基于这篇文章的基础上再详细分析:
- 结构化输出: 不依赖模糊的 Prompt,而是强制 LLM 返回符合的结构化数据
- Search-as-a-Tool,超越 K-Dense 基于静态数据的检索分析
- 如何让 LLM 写一段 Python 代码并在沙盒里运行
- 如何封装为一个 MCP Server,让工具拥有读取实时数据能力
为了详细分析如何用现代 MCP + Skills 架构,把领域的能力做成可装配、可迭代、可评测、可上架分发的技能包体系,并且不锁定 Claude 平台。
From Prompts to MCP Skills
Since late last year, I’ve been exploring how to assemble a practical set of Skills to support daily reading and writing, and gradually, how to stand up local MCP services. The challenge: lacking a full picture of the architecture—what it should look like, how it differs from traditional prompts, and what truly defines a “Skill.”
After two years of writing prompts, my view is simple:
- Prompts delegate a narrow, one-off task—ask a model a question and get an answer.
- Skills define a repeatable, structured workflow with domain-aware guidance, enabling the model to handle specialized sub-tasks professionally.
Recently, I came across the K-Dense team’s comprehensive collection of 138 ready-to-use scientific skills. It’s ideal as an initial analysis sample to clarify some remaining ambiguities around what Skills are and aren’t.
The project README is explicit: 138 skills, 28+ databases, 55+ Python packages—this is a coherent project system, not a loose bundle of prompts.
- MCP can be hosted on K-Dense’s servers.
- Or on a self-hosted MCP server.
Which also implies Docker deployment is feasible.
A Full Scientific Data Analysis Workflow
- Define objectives and deliverables: the problem to solve and outputs (figures, conclusions, report, code).
- Collect data (internal and external): internal experimental/clinical raw files; external public repositories and literature. Establish a data inventory.
- QC and cleaning: ETL to produce clean data and QC records.
- Annotation and integration: merge data tables with consistent mapping.
- Core analysis and modeling: ranking/scoring, differential expression, cohort stratification, trial matching, hypothesis testing.
- Validation and cross-checks: reproduction, robustness tests, literature comparison, independent validation.
- Visualization and interpretation: clear, readable visuals and narrative.
- Reporting and archiving: abstract, methods, key findings, limitations, next steps.
This example matters because it pairs the workflow with concrete data sources and tools—what to find and how to use it. External data is organized by domain (chemistry, structure, genes/proteins, etc.), with implemented access/query patterns. Tools are cataloged by function: statsmodels, scikit-learn for stats/ML; matplotlib, seaborn, ReportLab for figures and documents.
It also provides scenario-driven methods. Example: searching for lung cancer inhibitors.
- Identify cell types and genes that change across conditions.
- Predict prognosis or stratify patients; find stable feature sets.
- Discover small molecules interacting with target proteins.
- Assess hereditary cancer risk clinically.
- Apply systems biology network analysis.
- Deliver a clear, visually compelling report.
Summary Table: Workflow, Actions, Tools, Outputs
| Stage | What’s done | Common Tools/Databases | Typical Output |
|---|---|---|---|
| 1. Define Objectives | Specify problem and deliverables | — | Goals & Metrics |
| 2. Collect Data | Internal/external data inventory | 10X/VCF/mzML, ChEMBL, ZINC, ClinVar, NCBI Gene | Data Catalog |
| 3. QC & Cleaning | Filter/unify formats | Scanpy, pyOpenMS, pysam | Clean Data |
| 4. Annotation & Integration | Map to genes/proteins/pathways | UniProt, KEGG, Reactome, STRING | Merged Tables |
| 5. Core Analysis | Diff/docking/modeling | RDKit, DiffDock, PyDESeq2, scikit-learn, Arboreto | Results & Models |
| 6. Validation & Checks | Reproduction/robustness | statsmodels, PyMC, external GEO/Cellxgene | Validated Conclusions |
| 7. Visualization & Interpretation | Charts + narrative | matplotlib, seaborn | Visualizations |
| 8. Reporting & Archiving | Write/package for repro | ReportLab, README/config | PDF / Code Package |
Follow these eight steps—trace data provenance, clean and align it, derive and verify conclusions, and make them understandable and reproducible with the right tools and evidence—and complex biomedical R&D becomes a workflow anyone can follow.
In short, a Skill is not a prompt exchange nor a single analysis—it’s a project scaffold:
- Combine internal/external context to propose the best approach.
- Acquire and ETL both structured and unstructured data to high quality.
- Analyze, annotate, model, iterate, and validate—closing the PDCA loop.
Even state-of-the-art models struggle to solve complex tasks in one pass. A structured, stepwise approach is both feasible and proven in practice to accelerate expert workflows—often by multiples—while keeping humans in the loop for plan confirmation and domain judgment.
Practical Example Analysis
To make this concrete, I’ll examine four steps: defining objectives, data collection, core analysis, and validation.
Why these?
- Defining objectives: translating mixed internal/external signals into best-practice plans using LLMs.
- Data collection: deciding what’s useful now, and collecting efficiently.
- Core analysis: modeling loops and when human oversight is needed.
- Validation: practical patterns for top-tier model workflows.
I first download the repository locally and consolidate it into a single text file for detailed analysis in Gemini 3.
Proposal Writing Skill
Clarifying the problem and deliverables is achieved via collaboration between two skill families: Research Methodology & Planning and Scientific Communication.
See: Research Methodology & Planning.
Structure: a fixed two-layer design—YAML metadata for machine readability, followed by Markdown for human readability. The Markdown body starts with an overview that frames the problem and value, then details background, triggers, constraints, tasks, review, strategy, execution, assets, and iteration.
Key Sections and Best Practices
| Section | Purpose | Best Practice | Common Pitfalls |
|---|---|---|---|
| Overview | Define problem & value | Answer “for whom, when, why it matters” upfront | Abstract, no user benefit |
| When to Use | Trigger boundaries | Parallel positive/negative triggers; quantifiable prerequisites | Universal applicability claims |
| Constraints | Compliance & hard limits | Layered MUST/SHOULD/MAY; visualize minimum viable spec | Implicit constraints, scattered |
| Core Components | Execution building blocks | Unified sub-templates; explicit acceptance criteria | Checklists without executability |
| Review Standards | Reverse alignment | Use review questions to drive drafting; sample evidence | Lists without actionability |
| Workflow | Timeline & gating | Phased “Activity/Output/Time” triplets | Gantt-only, no acceptance gates |
| Resources & Assets | Reuse & cost control | Templates/scripts/refs, organized directories | Broken links, redundancy |
| Iteration Strategy | Improve & resubmit | Itemized responses, visible changes, stop conditions | Emotional rebuttals, unstructured |
- Primarily aimed at NSF/NIH/DOE/DARPA proposals, but adaptable beyond. The goal is aligning with FOA/BAA and compliance guidelines.
- Explicit usage boundaries: from project narratives to CVs and bios.
- Requires AI diagrams via the
scientific-schematicsskill; Nano Banana Pro auto-generates, reviews, and optimizes schematics—workflow diagrams, Gantt charts, conceptual frames, system architectures, experimental flows, impact maps, collaboration networks, and more. - Institution-specific requirements: choosing the right venue, narrative adaptation, and compliance.
- Project abstract: essential elements, writing strategies, common errors to avoid.
- Research strategy: background, objectives, methods, milestones.
- Specific aims: clear, testable goals.
- Broader impacts.
- Innovation.
- Methods & roadmap.
- Preliminary data & feasibility.
- Timeline, milestones, management plan.
- Team qualifications & collaboration.
- Budget narrative.
- Review standards by institution: significance, innovation, methods, etc.
- Principles for competitive proposals: including resubmissions and common mistakes.
- Multi-skill writing:
- Scientific Writing (clarity, persuasion)
- Literature Review (comprehensive background)
- Peer Review (pre-submission self-assessment)
- Research Lookup (citations and prior art)
- Data Visualization (effective figures)
- References to 11 complementary skills (timelines/milestones, high-performing teams, review responses, etc.).
Data Acquisition Skill
Data acquisition relies on the literature-review and research-lookup (Perplexity) skills. This goes beyond keyword searches to systematic reviews, pulling latest papers/preprints from PubMed, bioRxiv, and Google Scholar. Here I focus on literature-review.
Bridging internal and external data expands coverage and grounds coding practices in evidence, yielding higher-quality outputs. K-Dense not only searches external sources but also calls APIs for specialist databases (bioinformatics, patents, etc.). With document-skills, it parses PDFs/Word/Excel, extracting text/tables from user experiments or reports—aligning user data with public data.
The core difficulty: standardized mapping/alignment between internal and external data. Done well, you can integrate multi-dimensional data via multi-modality and orchestrate agent workflows. A vague instruction like “analyze this cancer dataset” triggers Python-based cleaning and external data linking to produce a synthesized conclusion.
Compared to research-grants, literature-review differs fundamentally:
- Research grants prioritize persuasion toward a hypothesis and venue fit; literature reviews prioritize objectivity, comprehensiveness, and reproducibility.
- Planning mindsets diverge: grants lean into reviewer pain points and gap analysis; literature reviews follow PICO (Population, Intervention, Comparison, Outcome), with strict inclusion/exclusion criteria.
- Retrieval patterns differ: targeted supportive citations vs. multi-database joint queries (PubMed, arXiv, bioRxiv) with Boolean logic.
- QA is essential: apply Cochrane or Newcastle–Ottawa scales; synthesize across studies and compare consensus vs. divergence.
Final delivery: a professional PDF or Markdown with ToC, body, and references.
literature-review Skill Outline
- Define: multi-database retrieval, topic synthesis, accuracy verification, documented output.
- Use cases: literature surveys, papers, future research directions.
- Schematics via Nano Banana Pro.
- Planning & scoping: clear questions, scope, and strategy.
- Multi-database retrieval: use
ggetto choose appropriate databases and query by keywords. - Screening & selection.
- Preliminary data quality checks: dates, authors, topics, basic signals.
- Data acquisition methods:
- arXiv via API or WebFetch
- Semantic Scholar via free endpoints
- Specialized domain databases
- Best practices.
- Common pitfalls; sample workflows.
- Integration with other skills: database access, analytical visualization, writing.
- Additional data sources.
- Python dependencies and installation details.
Statsmodels Statistical Modeling Skill
Statsmodels spans estimation, inference, and diagnostics across a wide statistical toolkit—from linear models to time series and econometrics.
Quick Start examples include:
- Linear Regression (OLS)
- Logistic Regression (Binary outcomes)
- Time Series (ARIMA)
- Generalized Linear Models (GLM)
Highlights:
- Linear models for continuous outcomes, with flexible error structures.
- GLM extends linear modeling to non-normal distributions.
- Discrete choice models for categorical and count data.
- Comprehensive time series modeling and forecasting.
- Broad statistical tests and diagnostics for validation.
It also offers guidance on data preparation, model construction, inference, evaluation, and reporting—plus references on common traps and how to avoid them.
From Dialogue to MCP: A Paradigm Shift
K-Dense-AI hardcodes many skills in a single repository. Meanwhile, the MCP ecosystem has matured; MCP servers are increasingly treated as pluggable skills. We’re moving from language-only dialogue to tool invocation, and onward to system integration via MCP.
Next Steps
Planned follow-ups based on this analysis:
- Structured outputs: require compliant schemas, not free-form prompts.
- Search-as-a-Tool: go beyond static retrieval to live, tool-integrated search.
- Have LLMs write Python and execute in sandboxes.
- Package capabilities as an MCP Server to access real-time data.
Goal: demonstrate how modern MCP + Skills architectures turn domain capabilities into assemblable, iterable, evaluable, and distributable skill packages—without locking to any specific model platform.
讨论
登录后可评论、点赞、引用回复。