project-wiki
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
Chinese<!--
This source file is part of the Stanford Spezi open-source project.
SPDX-FileCopyrightText: 2026 Stanford University and the project authors (see CONTRIBUTORS.md)
SPDX-License-Identifier: MIT
-->
<!--
This source file is part of the Stanford Spezi open-source project.
SPDX-FileCopyrightText: 2026 Stanford University and the project authors (see CONTRIBUTORS.md)
SPDX-License-Identifier: MIT
-->
Project Wiki
项目Wiki
Set up and maintain a persistent knowledge base for a digital health project. Instead of scattering notes across chat histories, Google Docs, and Slack threads, the wiki gives your project a single, structured home for everything you learn — and the AI keeps it current.
Inspired by Andrej Karpathy's LLM Wiki pattern, adapted for the Stanford Biodesign innovation process.
为数字医疗项目搭建并维护持久化知识库。无需再将笔记散落在聊天记录、Google Docs和Slack线程中,该wiki为项目所有沉淀的知识提供了唯一的结构化存放空间,并且由AI自动保持内容更新。
该方案受Andrej Karpathy的LLM Wiki模式启发,适配斯坦福Biodesign创新流程定制。
When to Use
适用场景
Use this skill when you want to:
- accumulate project knowledge that compounds over time instead of scattering across tools
- keep clinical observations, user interviews, literature, and competitive intelligence organized and cross-referenced
- maintain a living evidence base that updates as new information arrives
- give new team members a structured way to get up to speed on a project
Do not use this skill for one-off questions or quick research. It is designed for projects where knowledge accumulates over weeks or months.
当你有以下需求时可使用该工具:
- 沉淀随时间累积的项目知识,避免内容散落在不同工具中
- 有序管理临床观察、用户访谈、文献资料和竞争情报,支持内容交叉引用
- 维护动态更新的证据库,随新信息流入自动同步
- 为新团队成员提供结构化的项目快速上手路径
请勿将该工具用于一次性问题或快速调研,它专为知识需要数周/数月持续累积的项目设计。
The Core Idea
核心思路
Most teams use AI the way they use search — ask a question, get an answer, move on. The AI rediscovers context from scratch every time. Nothing compounds.
A project wiki is different. When you add a new source — an interview transcript, a paper, a clinical observation — the AI does not just file it. It reads it, extracts what matters, and integrates it into the existing wiki: updating stakeholder pages, revising the evidence landscape, flagging where new data contradicts earlier assumptions, strengthening or challenging the evolving picture.
The knowledge is compiled once and kept current, not re-derived on every conversation.
You curate sources, direct the analysis, and ask the right questions.
The AI does the summarizing, cross-referencing, filing, and bookkeeping.
大多数团队使用AI的方式和使用搜索引擎一样:提问、获得答案、结束对话。AI每次都需要从零重新梳理上下文,没有任何知识沉淀。
项目Wiki的逻辑完全不同。当你添加新的数据源(访谈转录稿、论文、临床观察记录)时,AI不只是简单归档,它会读取内容、提取核心信息、并整合到现有wiki中:更新利益相关方页面、修正证据全景、标记新数据与旧假设冲突的位置、补充或修正当前的知识体系。
知识只需编译一次即可保持更新,无需在每次对话中重新推导。
你负责筛选数据源、指导分析方向、提出正确的问题。
AI负责内容总结、交叉引用、归档和记录维护工作。
Working Style
工作方式
You are a wiki maintainer and research partner. When ingesting sources, be thorough — touch every page that the new information affects. When answering queries, cite specific wiki pages and source documents. When something contradicts existing knowledge, flag it clearly rather than silently overwriting.
Keep the wiki useful, not exhaustive. A concise page with good cross-references beats a long page that nobody reads.
你是wiki维护者和研究伙伴。摄入数据源时要全面:更新所有受新信息影响的页面。回答查询时,引用具体的wiki页面和源文档。如果新内容与现有知识冲突,要明确标记,不要静默覆盖。
保证wiki实用性而非追求面面俱到:一份带有完善交叉引用的简洁页面远好于没人愿意读的冗长页面。
Three-Layer Architecture
三层架构
Layer 1: Raw Sources
第一层:原始数据源
wiki/raw/
interviews/
papers/
observations/
competitors/
regulatory/
media/Your curated collection of source documents. Articles, papers, interview transcripts, clinical observation notes, screenshots, data files. These are immutable — the AI reads from them but never modifies them. This is your source of truth.
Organize by type. Use descriptive filenames with dates: .
2026-04-05-cardiac-rehab-patient-interview-03.mdwiki/raw/
interviews/
papers/
observations/
competitors/
regulatory/
media/这是你筛选的源文档集合,包括文章、论文、访谈转录稿、临床观察笔记、截图、数据文件。这些内容是不可修改的:AI仅读取内容,永远不会修改该目录下的文件,是项目的可信数据源。
按内容类型组织文件,使用带日期的描述性文件名:。
2026-04-05-cardiac-rehab-patient-interview-03.mdLayer 2: The Wiki
第二层:Wiki内容层
wiki/pages/
index.md
log.md
overview.md
stakeholders/
evidence/
landscape/
design/
regulatory/
questions/A directory of AI-generated and AI-maintained markdown files. Summaries, entity pages, concept pages, comparisons, evidence tables, and synthesis documents. The AI owns this layer entirely. It creates pages, updates them when new sources arrive, maintains cross-references, and keeps everything consistent. You read it; the AI writes it.
wiki/pages/
index.md
log.md
overview.md
stakeholders/
evidence/
landscape/
design/
regulatory/
questions/这是由AI生成和维护的Markdown文件目录,包含总结、实体页面、概念页面、对比分析、证据表格、合成文档。AI完全拥有该层的所有权:它负责创建页面、新数据源流入时更新页面、维护交叉引用、保证内容一致性。你仅读取该层内容,由AI负责写入。
Layer 3: The Schema
第三层:规则配置层
The (or ) file at the project root. It tells the AI how the wiki is structured, what the conventions are, and what workflows to follow when ingesting sources, answering questions, or maintaining the wiki. You and the AI co-evolve this over time as you figure out what works for your project.
AGENTS.mdCLAUDE.md项目根目录下的(或)文件,它告知AI wiki的结构、命名规范、以及摄入数据源、回答问题、维护wiki时需要遵循的工作流。你可以和AI随项目推进共同迭代该文件,找到最适配项目的规则。
AGENTS.mdCLAUDE.mdSetup
搭建步骤
When a user asks to set up a project wiki, do the following:
当用户要求搭建项目Wiki时,按以下步骤操作:
1. Understand the Project
1. 了解项目信息
Ask:
- "What is the project about? What problem are you working on?"
- "What stage are you in — early exploration, needs-finding, prototyping, validation?"
- "What kinds of sources do you already have — interviews, papers, clinical observations, competitive research?"
询问以下问题:
- "项目的核心方向是什么?你正在解决什么问题?"
- "项目当前处于什么阶段:早期探索、需求挖掘、原型开发、验证阶段?"
- "你目前已经有哪些类型的数据源:访谈、论文、临床观察、竞争研究?"
2. Scaffold the Wiki
2. 搭建Wiki目录结构
Create the directory structure inside the project repository:
wiki/
raw/
interviews/
papers/
observations/
competitors/
regulatory/
media/
pages/
index.md
log.md
overview.md
stakeholders/
evidence/
landscape/
design/
regulatory/
questions/在项目仓库中创建如下目录结构:
wiki/
raw/
interviews/
papers/
observations/
competitors/
regulatory/
media/
pages/
index.md
log.md
overview.md
stakeholders/
evidence/
landscape/
design/
regulatory/
questions/3. Seed from Existing Planning Docs
3. 导入现有规划文档
If the project has SpeziVibe planning documents, ingest them as the wiki's first sources:
| Planning Document | Wiki Pages Created |
|---|---|
| |
| |
| |
| |
| |
| |
| |
Do not duplicate content — extract key facts, relationships, and open questions into wiki pages with links back to the original planning docs.
如果项目有SpeziVibe规划文档,将其作为wiki的首批数据源导入:
| 规划文档 | 创建的Wiki页面 |
|---|---|
| |
| |
| |
| |
| |
| |
| |
不要重复内容:提取核心事实、关联关系和待解决问题到wiki页面,并添加回原规划文档的链接。
4. Write the Schema
4. 编写规则配置
Check whether an or already exists at the project root.
AGENTS.mdCLAUDE.md- If one exists: append a clearly scoped section to the existing file. Do not overwrite or remove any existing content — the file may contain instructions for other tools or skills.
## Project Wiki Schema - If neither exists: create a new at the project root.
AGENTS.md
The Project Wiki Schema section should describe:
- the wiki directory structure and conventions
- the project domain and key terminology
- page naming conventions
- how to handle ingestion, queries, and maintenance
- which Biodesign stage the project is in and what that means for page priorities
- cross-referencing conventions (use standard Markdown links)
Tailor the schema to the specific project. A cardiac rehab app wiki has different page categories than a surgical device wiki.
检查项目根目录下是否已经存在或:
AGENTS.mdCLAUDE.md- 如果已存在:在现有文件末尾添加范围明确的章节,不要覆盖或删除任何现有内容,该文件可能包含其他工具或技能的指令。
## 项目Wiki规则 - 如果不存在:在项目根目录下创建新的文件。
AGENTS.md
项目Wiki规则章节需要描述:
- wiki目录结构和规范
- 项目领域和核心术语
- 页面命名规范
- 内容摄入、查询、维护的处理方式
- 项目当前所处的Biodesign阶段,以及该阶段对应的页面优先级
- 交叉引用规范(使用标准Markdown链接)
根据具体项目定制规则:心脏康复应用wiki的页面分类和手术设备wiki的页面分类完全不同。
5. Create the Index and Log
5. 创建索引和日志
wiki/pages/index.mdwiki/pages/log.md## [YYYY-MM-DD] action | descriptionmarkdown
undefinedwiki/pages/index.mdwiki/pages/log.md## [YYYY-MM-DD] 操作 | 描述markdown
undefined[2026-04-05] setup | Wiki initialized for CardioTrack project
[2026-04-05] setup | CardioTrack项目Wiki初始化
[2026-04-05] seed | Ingested need-statement.md, ux-brief.md, compliance-brief.md
[2026-04-05] seed | 导入need-statement.md、ux-brief.md、compliance-brief.md
[2026-04-07] ingest | Patient interview #3 — post-discharge cardiac rehab
[2026-04-07] ingest | 患者访谈#3 — 出院后心脏康复
[2026-04-07] query | "What are the main barriers to exercise adherence?" → filed as evidence/adherence-barriers.md
[2026-04-07] query | "运动依从性的主要障碍有哪些?" → 归档为evidence/adherence-barriers.md
[2026-04-10] lint | Found 3 orphan pages, 1 contradicted claim, 2 missing cross-references
[2026-04-10] lint | 发现3个孤立页面、1个矛盾声明、2个缺失的交叉引用
undefinedundefinedCore Workflows
核心工作流
Ingest
内容摄入
The user should never have to think about folder structure. When they want to add something to the wiki, they can:
- paste text directly into the conversation
- share or upload a file (PDF, image, markdown, transcript)
- share a URL
- dictate or describe an observation
The AI handles everything from there:
- Save the source — classify the source type (interview, paper, observation, competitor, regulatory) and save it to the correct subfolder with a descriptive, dated filename. The user never needs to navigate the folder structure manually.
wiki/raw/ - Read the source completely
- Discuss key takeaways with the user — what surprised them, what confirms existing thinking, what challenges it
- Write or update a summary page in the wiki
- Update
wiki/pages/index.md - Update every relevant entity, concept, and evidence page across the wiki
- Flag contradictions with existing wiki content explicitly — do not silently overwrite
- Append an entry to
wiki/pages/log.md
A single source may touch 5–15 wiki pages. That is expected.
The user's job is just: "Add this to my wiki" + share the content. One step.
Prefer ingesting sources one at a time with the user involved. Batch ingestion is fine for catching up, but interactive ingestion produces better results.
用户无需关心文件夹结构,当他们想要向wiki添加内容时,可以:
- 直接粘贴文本到对话
- 分享或上传文件(PDF、图片、Markdown、转录稿)
- 分享URL
- 口述或描述观察内容
AI负责后续所有工作:
- 保存源文件:分类数据源类型(访谈、论文、观察、竞品、监管),保存到下对应的子文件夹,使用带日期的描述性文件名,用户无需手动导航文件夹结构。
wiki/raw/ - 完整读取源文件
- 和用户讨论核心要点:哪些内容出人意料、哪些验证了现有想法、哪些带来了挑战
- 编写或更新wiki中的总结页面
- 更新
wiki/pages/index.md - 更新wiki中所有相关的实体、概念、证据页面
- 明确标记和现有wiki内容的冲突,不要静默覆盖
- 追加记录到
wiki/pages/log.md
单个数据源可能会影响5-15个wiki页面,这是正常情况。
用户只需要做:"把这个添加到我的wiki" + 分享内容,仅需一步。
优先单次摄入单个数据源并让用户参与,批量摄入适合补录历史数据,但交互式摄入效果更好。
Query
问题查询
When the user asks a question:
- Read to find relevant pages
wiki/pages/index.md - Read those pages and synthesize an answer with citations to specific wiki pages and raw sources
- If the answer reveals a useful synthesis, comparison, or connection — offer to file it as a new wiki page so it compounds rather than disappearing into chat history
当用户提出问题时:
- 读取找到相关页面
wiki/pages/index.md - 读取这些页面,合成答案,并引用具体的wiki页面和原始数据源
- 如果答案形成了有用的合成内容、对比分析或关联关系,主动提议将其归档为新的wiki页面,让知识沉淀而不是消失在聊天记录中。
Lint
质量检查
Periodically (or when the user asks), health-check the wiki:
- Contradictions — pages that disagree with each other or with newer sources
- Stale claims — assertions that newer evidence has superseded
- Orphan pages — pages with no inbound links from other pages
- Missing pages — important concepts mentioned on other pages but lacking their own page
- Missing cross-references — pages that should link to each other but don't
- Evidence gaps — questions or claims that lack supporting sources
- Stage alignment — whether the wiki's depth matches the project's current Biodesign stage
Present findings as a checklist and offer to fix each one.
定期(或用户要求时)对wiki进行健康检查:
- 内容冲突:页面之间相互矛盾,或与更新的数据源冲突
- 过时声明:已经被新证据取代的断言
- 孤立页面:没有其他页面入链的页面
- 缺失页面:其他页面提到了重要概念,但没有对应的独立页面
- 缺失交叉引用:应该互相链接的页面没有添加链接
- 证据缺口:缺乏源文件支撑的问题或声明
- 阶段匹配度:wiki的内容深度是否匹配项目当前的Biodesign阶段
将检查结果以清单形式呈现,并主动提议修复每一个问题。
Biodesign-Specific Page Types
Biodesign专属页面类型
These are starting categories. Adapt them to the project:
以下是初始分类,可根据项目适配调整:
Stakeholders
利益相关方
One page per key stakeholder group. Include:
- who they are
- how the problem affects them
- their role in adoption decisions (decision-maker, influencer, user, payer)
- evidence from interviews or observations
- links to relevant raw sources
每个核心利益相关方群体对应一个页面,包含:
- 群体身份
- 问题对他们的影响
- 他们在采用决策中的角色(决策者、影响者、用户、付费方)
- 访谈或观察得到的证据
- 相关原始源文件的链接
Evidence
证据
Pages that synthesize what is known about specific topics:
- clinical evidence for the problem
- current standard of care and its limitations
- quantitative burden (prevalence, cost, outcomes data)
- key studies and their findings
合成特定主题已知信息的页面:
- 问题相关的临床证据
- 当前标准治疗方案及其局限性
- 量化负担(患病率、成本、预后数据)
- 核心研究及其发现
Landscape
行业全景
Competitive and market analysis:
- existing solutions and their strengths and limitations
- adjacent technologies
- market size and dynamics
- IP considerations
竞争和市场分析:
- 现有解决方案及其优劣势
- 相邻技术
- 市场规模和动态
- 知识产权考量
Design
设计
Product and technical decisions:
- user journeys and workflows
- data model choices
- architecture decisions and rationale
- implementation milestones
产品和技术决策:
- 用户旅程和工作流
- 数据模型选型
- 架构决策和依据
- 落地里程碑
Regulatory
监管
Compliance and regulatory pathway:
- applicable regulations and standards
- classification decisions
- submission pathway
- open regulatory questions
合规和监管路径:
- 适用的法规和标准
- 分类决策
- 提交路径
- 待解决的监管问题
Questions
待解决问题
A living list of open questions, organized by category. When a question gets answered, move it to the relevant wiki page and note the resolution.
动态更新的开放问题列表,按类别组织。问题得到解答后,将其移动到对应的wiki页面并记录解决方案。
Integration with build-an-app
与应用构建流程的集成
When completes its planning phase, it should offer:
build-an-app"Your planning documents are ready. Would you like to set up a project wiki to keep accumulating knowledge as you build? The wiki will seed from your planning docs and grow as you add interviews, papers, and clinical observations."
If the user accepts, hand off to this skill.
当完成规划阶段后,应该主动询问:
build-an-app"你的规划文档已经准备就绪。是否要搭建项目wiki,在构建过程中持续沉淀知识?wiki会自动导入你的规划文档,并随着你添加访谈、论文和临床观察内容持续增长。"
如果用户同意,切换到该技能处理。
Guardrails
使用准则
- Raw sources are immutable. Never modify anything in . The AI reads from raw sources but only writes to
wiki/raw/.wiki/pages/ - Flag contradictions, don't hide them. When new information conflicts with existing wiki content, note both positions and the evidence for each. Let the user decide what to believe.
- Cite everything. Every claim in the wiki should trace back to a raw source or a planning document. Use Markdown links.
- Keep pages concise. A wiki page should be readable in 2–3 minutes. Split long pages into focused sub-pages.
- Don't fabricate evidence. If the wiki has gaps, say so. Add the gap to rather than filling it with speculation.
questions/open-questions.md - Respect the user's domain expertise. The AI maintains the wiki; the user directs the analysis. Ask before making judgment calls about clinical significance or research direction.
- The wiki is a git repo. Encourage commits after significant updates. The version history is valuable.
- 原始数据源不可修改:永远不要修改下的任何内容,AI仅读取原始数据源,仅写入
wiki/raw/目录。wiki/pages/ - 标记冲突不要隐藏:当新信息和现有wiki内容冲突时,同时记录两种观点及其支撑证据,让用户决定采信哪一方。
- 所有内容标注来源:wiki中的每一个声明都应该能追溯到原始数据源或规划文档,使用Markdown链接。
- 保持页面简洁:单个wiki页面应该能在2-3分钟内读完,将长页面拆分为聚焦的子页面。
- 不要编造证据:如果wiki存在内容缺口,明确说明,将缺口添加到,不要用猜测填充。
questions/open-questions.md - 尊重用户的领域专业知识:AI负责维护wiki,用户负责指导分析方向,涉及临床意义或研究方向的判断时要提前询问用户。
- wiki是git仓库:重大更新后鼓励提交commit,版本历史很有价值。
Checklist
检查清单
- Project scope and stage understood
- Directory structure created
- Existing planning docs seeded into wiki
- Schema file (AGENTS.md or CLAUDE.md) written and tailored to the project
- Index and log files initialized
- At least one source ingested interactively to demonstrate the workflow
- User understands ingest, query, and lint workflows
- 已明确项目范围和所处阶段
- 已创建目录结构
- 已将现有规划文档导入wiki
- 已编写适配项目的规则配置文件(AGENTS.md或CLAUDE.md)
- 已初始化索引和日志文件
- 已交互式摄入至少一个数据源演示工作流
- 用户已理解内容摄入、查询和质量检查工作流