Token导航 LogoToken导航

Agent Skill · 研究检索

karpathy-wiki(卡帕蒂维基)

karpathy-wiki 用于查找、检索和筛选相关信息,适合在 OpenClaw 中需要根据关键词、任务场景或来源线索快速定位候选结果时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

karpathy-wiki 是研究检索类 Agent Skill,可用宿主为 OpenClaw,通过ClawHub获取,权限为只读;许可证为 MIT-0,第三方扫描已通过,按安装量在研究检索类的 73,675 个 Skill 里排第 24,678。

总安装 3,951周安装 163GitHub Stars 公开资料未说明下载 1,291核验 2026-05-01

安装说明

本站只整理中文说明,不托管安装包,也不代用户安装。
GitHub
许可证
MIT-0
版本
0.1.1

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:karpathy-wiki(卡帕蒂维基)
来源仓库:https://github.com/teki-ai/karpathy-wiki
安装命令:
openclaw skills install karpathy-wiki
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 OpenClaw 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

ClawHubOpenClaw
openclaw skills install karpathy-wiki

这个 Skill 是做什么的

研究检索类 Skill 主要解决「找资料、核对事实、把来源整理清楚」的问题,通常会让助手先检索再总结。

karpathy-wiki 用于构建和维护持久化的 Markdown 知识库,支持在 OpenClaw 中快速检索和筛选相关信息。适用于需要根据关键词、任务场景或来源线索定位候选结果的场景,如个人知识管理或团队协作。通过关键词查询、来源筛选和结果验证来组织信息,具体用法需结合原始 README 进一步确认。

安装命令为 openclaw skills install karpathy-wiki,建议提前确认权限范围和维护状态。使用前需注意是否涉及联网、文件读写或命令执行,避免触发不必要的安全限制。

上手步骤

4 步
  1. 1

    确认自己的宿主工具

    公开资料标注支持 OpenClaw;先确认自己用的是其中之一,再按对应工具的 Skills 目录放置。

  2. 2

    拿到安装命令

    可以直接复制的命令是 openclaw skills install karpathy-wiki;命令里的占位符换成自己的路径或密钥再执行。

  3. 3

    安装后重启宿主

    命令执行完要完全退出宿主工具再打开,让新装的 Skill 被重新扫描到,否则列表里可能看不到。

  4. 4

    按最小权限试用

    它标注为只读,先用只读任务试一次,确认行为符合预期再扩大使用范围。

常见问题

6 条

karpathy-wiki 怎么安装?

在终端执行 openclaw skills install karpathy-wiki。装完完全退出宿主工具再打开,让它重新扫描本地 Skill。

karpathy-wiki 支持哪些 AI 工具?

公开资料标注支持 OpenClaw。其他工具只要能读取本地 Skill 目录,一般也能手动放进去,但要自己确认触发方式是否一致。

karpathy-wiki 需要什么权限?

权限级别是只读。该 Skill 主要提供规则、说明或参考内容,本身偏只读;真正读写文件、联网或执行命令仍取决于宿主 Agent 的任务。

karpathy-wiki 是免费的吗?

来源标注的许可证是 MIT-0。本站不代收任何费用,也不托管安装包;能否商用、要不要署名,以许可证原文为准。

karpathy-wiki 用的人多吗?

公开数据是累计安装 3,951 次,近一周安装 163 次,下载 1,291 次。这些数字来自第三方目录统计,只能当作热度参考,不代表质量或安全性。

karpathy-wiki 安全吗?

第三方扫描结果为通过,但这只覆盖已知的静态检查项,不能替代自己读一遍源码和安装命令。

值得选的地方和要先想清楚的地方

值得选的地方

  • 专注 OpenClaw 一种宿主,触发方式和使用习惯不用来回切换。
  • 许可证明确为 MIT-0,商用和分发边界好判断。
  • 第三方扫描已通过,常见的高风险行为没有命中。
  • 标注为只读权限,试用阶段不容易误改本地文件。

先想清楚的地方

  • 第三方目录整理,版本更新和行为细节以来源仓库为准。

适不适合你

更适合

  • 已经在用 OpenClaw,想直接扩展它能力的人。
  • 需要处理研究检索类任务、又不想每次都手写提示词的人。
  • 需要许可证清楚、并且有第三方扫描结果的团队项目。

不太适合

  • 需要厂商级技术支持和 SLA 的生产系统。

研究检索类 Skill 对比

6 个同类
Skill总安装GitHub 星权限许可证扫描
find-skills

寻找技能 · Codex、Claude

31,199,81316,486操作浏览器MIT有提醒
self-improving-agent

自我改进 Agent · OpenClaw

16,464,0023,396接触敏感数据MIT-0通过
soultrace

灵魂追踪 · Codex、Claude

7,588,94412需要联网未明确有提醒
self-improving

自我完善 · OpenClaw

6,829,2841,046需要联网MIT-0通过
microsoft-foundry

微软代工厂 · Codex、Claude

6,614,144764调用外部服务未明确有提醒
karpathy-wiki

卡帕蒂维基 · 本页

3,951—只读MIT-0通过

按安装量取同分类前几名;数字来自第三方目录统计,只作热度参考,不代表质量结论。

SKILL.md 原文

name
karpathy-wiki
description
Build, maintain, query, and lint a persistent markdown knowledge wiki that sits between raw sources and final answers. Use when managing a personal or team wiki, ingesting notes, articles, transcripts, reports, or research into a structured knowledge base, updating index and log pages, creating or revising topic, entity, concept, source, or analysis pages, answering questions from an existing wiki, or checking a wiki for contradictions, stale claims, missing links, orphan pages, duplicate pages, and schema drift. Best for Obsidian-friendly, plain markdown, or git-backed wikis where the agent should incrementally compile knowledge instead of re-deriving it from raw documents on every query.

Karpathy Wiki

Maintain a persistent markdown wiki that compiles knowledge over time. Treat raw sources as immutable, treat the wiki as the maintained artifact, and keep structure and conventions consistent so future sessions can continue the work.

Core model

Operate with three layers:

  1. raw sources/ or equivalent input area, read-only source material
  2. wiki/ or equivalent markdown knowledge base, editable compiled knowledge
  3. schema document that defines folder layout, naming, citation style, and workflows

Prefer updating the wiki over answering from scratch. When useful work is produced during analysis or Q&A, file the result back into the wiki as a reusable page.

Default wiki structure

If the user does not already have a schema, propose a simple markdown-first layout like:

wiki/
  index.md
  log.md
  schema.md
  sources/
  pages/
    topics/
    concepts/
    entities/
    analyses/

Adapt to the user's existing layout instead of forcing this one.

Operating modes

1. Ingest

Use when a new source is added.

Workflow:

  1. Read the source
  2. Identify key topics, entities, concepts, dates, claims, and open questions
  3. Create or update a source summary page if the wiki uses source pages
  4. Update affected topic, entity, concept, or analysis pages
  5. Add or repair internal links
  6. Update index.md
  7. Append a concise entry to log.md
  8. Record uncertainties, contradictions, or follow-up questions explicitly

During ingest, prefer touching a small number of clearly relevant pages over creating a large number of weak pages.

Create a new page only when at least one of these is true:

  • the concept or entity is likely to recur
  • the page would receive meaningful links from multiple places
  • the content would otherwise make an existing page too broad or noisy

Otherwise, expand an existing page.

2. Query

Use when the user asks a question against the wiki.

Workflow:

  1. Read index.md first when available
  2. Identify the most relevant wiki pages
  3. Synthesize from the wiki before falling back to raw sources
  4. Cite the wiki pages and, when appropriate, the underlying sources
  5. If the answer creates durable value, save it as a new or updated analysis page
  6. Update index.md if a new page is created
  7. Append to log.md if the wiki treats queries as first-class events

Prefer answers that preserve distinctions between:

  • facts directly supported by sources
  • synthesis across multiple pages
  • speculation or open questions

3. Lint

Use when checking wiki health.

Look for:

  • orphan pages with few or no inbound references
  • duplicate pages with overlapping scope
  • stale claims superseded by newer sources
  • contradictions between pages
  • missing cross-links
  • important concepts mentioned repeatedly without their own page
  • schema drift, inconsistent titles, inconsistent frontmatter, broken naming
  • analysis pages that should have been linked from topic/entity pages but were not

When linting, prefer producing an actionable list of fixes grouped by severity:

  • critical consistency issues
  • likely quality improvements
  • optional structural improvements

Page conventions

Favor concise pages with clear structure. A useful default shape is:

# Page Title

## Summary
Short synthesis of what this page is about.

## Key points
- Bullet points of durable knowledge

## Details
Longer notes, evidence, chronology, or structured sections

## Related
- [[Other Page]]

## Sources
- [[Source Page A]]
- [[Source Page B]]

If the wiki uses frontmatter, keep it minimal and consistent. Good optional fields include:
- `type`
- `aliases`
- `status`
- `updated`
- `source_count`
- `tags`

Do not invent elaborate metadata unless the user actually benefits from it.

## Naming and linking rules

Use stable, human-readable file names.

Prefer:
- one canonical page per concept/entity/topic
- redirects or aliases only when the wiki supports them
- explicit wiki-links between related pages
- consistent singular vs plural naming

When unsure whether two pages should merge, keep both only if they have clearly different scope. Otherwise merge and leave one canonical page.

## Index and log rules

### `index.md`

Treat `index.md` as the navigational catalog.

Include:
- page link
- one-line summary
- optional grouping by category
- optional source counts or update dates if the wiki uses them

Keep it skimmable. It should help future sessions decide what to read next.

### `log.md`

Treat `log.md` as append-only chronology.

Use a parseable heading style such as:

[2026-04-12] ingest | Source title


Keep entries concise:
- what was ingested, queried, or linted
- which pages were created or updated
- any unresolved issues

## Quality bar

A good wiki update should:
- preserve source fidelity
- surface contradictions instead of hiding them
- strengthen cross-links
- reduce future work
- make later questions cheaper to answer

Do not overwrite uncertainty with confident prose. When the evidence is mixed, say so clearly.

## Working with existing wikis

If a wiki already exists:

1. Inspect its schema, folder layout, and naming style
2. Follow the existing conventions unless they are clearly harmful
3. Repair inconsistencies gradually instead of rewriting the whole wiki at once
4. Propose larger schema changes before making them

## Obsidian-friendly guidance

For Obsidian-style vaults:
- prefer markdown files and wiki-links like `[[Page Name]]`
- keep filenames readable
- avoid fragile generated syntax unless the user already uses it
- if frontmatter exists, preserve formatting and field order when practical
- make pages pleasant to browse by humans, not only optimized for machine parsing

## Deliverables by task

### For ingest requests
Deliver:
- source summary or source page update
- updated related pages
- `index.md` update
- `log.md` entry
- short note on contradictions or open questions

### For query requests
Deliver:
- answer with citations
- optional durable analysis page if worth keeping
- any relevant index or log updates

### For lint requests
Deliver:
- prioritized issue list
- concrete proposed edits
- optional patch plan for high-value fixes

## References

- Read `references/getting-started.md` when the user needs a minimal starter schema for a new wiki.
- Read `references/wiki-patterns.md` for core page templates and structural heuristics.
- Read `references/ingest-patterns.md` when ingesting a new source into the wiki.
- Read `references/query-patterns.md` when answering questions from the wiki or deciding whether to save a durable analysis page.
- Read `references/lint-checklist.md` when checking the wiki for contradictions, stale claims, weak links, duplicates, or structural drift.

平台分布

OpenClaw

87.7%
按下载量换算1,132

安全审计

VirusTotal

通过

ClawScan

通过

Static analysis

通过

权限和风险

只读

该 Skill 主要提供规则、说明或参考内容,本身偏只读;真正读写文件、联网或执行命令仍取决于宿主 Agent 的任务。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。目前只有一个公开来源,建议同时核对源仓库。

继续浏览同类 Skills