流动蓝色
复现方式:依赖 skill html-imagegen(来源 github.com/lindong28/lindong-skills)——把 Prompt 原样粘贴给 Claude / Codex 即可:Prompt 开头写明了 skill 地址,agent 会自行安装,并向你确认要表达的内容与目标平台。
双平台效果图
小红书封面
按四块阅读输入
角色与任务
你是中文封面编辑与排版设计师。根据本次完整原文制作一张小红书封面,输出完整HTML。
基本要求
输出独立完整HTML,导出区域唯一 id="cover",尺寸600×800像素,3:4,不透明底板,浏览器原生截图为同尺寸PNG。文字不越界、不重叠、不截断。使用系统字体、内嵌CSS与装饰,不调用图片模型或外部API。署名显示 Andrej Karpathy / @karpathy,不添加发布账号。转述忠实于下方原文,不伪装成直接引语。
视觉要求
近白灰蓝底,右上蓝雾与左下浅蓝渐变,柔和弧光贯穿中下部;左上小型来源,左对齐大标题用蓝色强调其中一段,摘要与蓝色小标签接在标题下方,空心环和淡三角点缀。种子标题从左49px、上311px、53px起排,三行只是起始组合;保留流动弧光与上下呼吸感,随内容调整强调位置、换行和摘要。
种子是排版起点,不是不可变的填槽表。保持本款风格与核心构图,按内容调整文字模块、数量、换行、字号、行距、间距和模块尺寸;代码中的像素值、行数及槽位数量均为起始值,不是硬内容预算。提炼一个清楚的中心主张,再选择支撑它的必要信息;不为填满空槽捏造事实、引语、标签、英文、账号、期号或步骤。无内容依据的可选模块及其分隔符一并移除,保留有意的留白,不用重复文案填空间。
纵向保留原种子的上下层级与有意留白,内容模块按本次文案适配。
本次平台完整HTML骨架(以下任务局部CSS已确定本次排版):
```html
<!DOCTYPE html>
<html lang="zh-CN"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>封面</title><style>*{box-sizing:border-box}html,body{margin:0}body{font-family:"PingFang SC","Microsoft YaHei",Arial,sans-serif}#cover{width:600px;height:800px;position:relative;overflow:hidden}h1,h2,p{margin:0}#cover{background:radial-gradient(ellipse at 91% 13%,#bedbed 0,transparent 45%),radial-gradient(ellipse at 11% 92%,#c6e2ee 0,transparent 31%),linear-gradient(145deg,#fbfbfb,#e4e8ed 75%,#e6edf4);color:#2f2f2f}.source{position:absolute;left:49px;top:48px;display:flex;gap:9px;align-items:flex-start}.initial{width:26px;height:26px;background:linear-gradient(140deg,#0875d2,#00beee);border-radius:50%;color:white;font-size:17px;display:flex;align-items:center;justify-content:center;font-weight:700}.author{font-size:15px;font-weight:700;margin-top:3px}.handle{font-size:14px;color:#0572d2;margin-top:12px}.arcs{position:absolute;inset:0;width:600px;height:800px;pointer-events:none}.ring{position:absolute;left:518px;top:159px;width:20px;height:20px;border-radius:50%;border:2px solid #75b1f5a0}.triangle{position:absolute;left:91px;bottom:121px;width:0;height:0;border-left:12px solid transparent;border-right:12px solid transparent;border-bottom:20px solid #90c8fa77}h1{position:absolute;left:49px;top:311px;font-size:53px;line-height:1.14;letter-spacing:-1.7px;font-weight:750}h1 em{font-style:normal;color:#046acb}.summary{position:absolute;left:49px;right:49px;top:516px;font-size:17px;line-height:1.5;color:#555}.tags{position:absolute;left:49px;bottom:65px;display:flex;gap:9px}.tags span{border-radius:5px;padding:6px 11px;color:#006bcc;background:#bdd9f0;font-size:14px;font-weight:650}</style><style>/* 本次平台的局部排版 */h1{top:278px;font-size:56px;line-height:1.16;letter-spacing:-1px}.summary{top:514px;font-size:19px;line-height:1.65}.initial{font-size:11px}.tags{bottom:60px}</style></head><body><main id="cover"><svg class="arcs" viewBox="0 0 600 800" xmlns="http://www.w3.org/2000/svg"><defs><filter id="blur"><feGaussianBlur stdDeviation="2.7"/></filter></defs><g fill="none" stroke="#ffffff" opacity=".30" filter="url(#blur)"><ellipse cx="345" cy="388" rx="337" ry="210" transform="rotate(-14 345 388)" stroke-width="3"/><ellipse cx="355" cy="389" rx="325" ry="210" transform="rotate(-21 355 389)" stroke-width="3"/><path d="M-40 293C130 484 406 662 679 654" stroke-width="4"/></g><path d="M25 441C92 590 359 666 622 596" fill="none" stroke="#ffffff" stroke-width="1.7" opacity=".25"/></svg><div class="source"><div class="initial">{{source_initial}}</div><div><div class="author">{{author}}</div><div class="handle">{{source_caption}}</div></div></div><div class="ring"></div><div class="triangle"></div><h1>{{title1}}<br><em>{{title2}}</em><br>{{title3}}</h1><p class="summary">{{summary}}</p><div class="tags"><span>{{tag1}}</span><span>{{tag2}}</span></div></main></body></html>
```内容与素材
中心主张:LLM 将原始资料整理为 wiki,并通过研究问答和输出回存持续积累知识。封面只呈现这一主张,不把全文塞进画面。 图片素材:无。 本次文案映射(br 表示语义换行): source_initial:AK author:Andrej Karpathy source_caption:@karpathy title1:让知识 title2:在每次提问后 title3:继续生长 summary:LLM 把原始资料编成 wiki,<br>再把问答结果存回知识库。<br>在 Obsidian 中查看与研究。 tag1:Markdown wiki tag2:持续积累 完整原文: 作者:Andrej Karpathy(@karpathy) 来源:https://x.com/karpathy/status/2039805659525644595 LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.
# 角色与任务
你是中文封面编辑与排版设计师。根据本次完整原文制作一张小红书封面,输出完整HTML。
# 基本要求
输出独立完整HTML,导出区域唯一 id="cover",尺寸600×800像素,3:4,不透明底板,浏览器原生截图为同尺寸PNG。文字不越界、不重叠、不截断。使用系统字体、内嵌CSS与装饰,不调用图片模型或外部API。署名显示 Andrej Karpathy / @karpathy,不添加发布账号。转述忠实于下方原文,不伪装成直接引语。
# 视觉要求
近白灰蓝底,右上蓝雾与左下浅蓝渐变,柔和弧光贯穿中下部;左上小型来源,左对齐大标题用蓝色强调其中一段,摘要与蓝色小标签接在标题下方,空心环和淡三角点缀。种子标题从左49px、上311px、53px起排,三行只是起始组合;保留流动弧光与上下呼吸感,随内容调整强调位置、换行和摘要。
种子是排版起点,不是不可变的填槽表。保持本款风格与核心构图,按内容调整文字模块、数量、换行、字号、行距、间距和模块尺寸;代码中的像素值、行数及槽位数量均为起始值,不是硬内容预算。提炼一个清楚的中心主张,再选择支撑它的必要信息;不为填满空槽捏造事实、引语、标签、英文、账号、期号或步骤。无内容依据的可选模块及其分隔符一并移除,保留有意的留白,不用重复文案填空间。
纵向保留原种子的上下层级与有意留白,内容模块按本次文案适配。
本次平台完整HTML骨架(以下任务局部CSS已确定本次排版):
```html
<!DOCTYPE html>
<html lang="zh-CN"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>封面</title><style>*{box-sizing:border-box}html,body{margin:0}body{font-family:"PingFang SC","Microsoft YaHei",Arial,sans-serif}#cover{width:600px;height:800px;position:relative;overflow:hidden}h1,h2,p{margin:0}#cover{background:radial-gradient(ellipse at 91% 13%,#bedbed 0,transparent 45%),radial-gradient(ellipse at 11% 92%,#c6e2ee 0,transparent 31%),linear-gradient(145deg,#fbfbfb,#e4e8ed 75%,#e6edf4);color:#2f2f2f}.source{position:absolute;left:49px;top:48px;display:flex;gap:9px;align-items:flex-start}.initial{width:26px;height:26px;background:linear-gradient(140deg,#0875d2,#00beee);border-radius:50%;color:white;font-size:17px;display:flex;align-items:center;justify-content:center;font-weight:700}.author{font-size:15px;font-weight:700;margin-top:3px}.handle{font-size:14px;color:#0572d2;margin-top:12px}.arcs{position:absolute;inset:0;width:600px;height:800px;pointer-events:none}.ring{position:absolute;left:518px;top:159px;width:20px;height:20px;border-radius:50%;border:2px solid #75b1f5a0}.triangle{position:absolute;left:91px;bottom:121px;width:0;height:0;border-left:12px solid transparent;border-right:12px solid transparent;border-bottom:20px solid #90c8fa77}h1{position:absolute;left:49px;top:311px;font-size:53px;line-height:1.14;letter-spacing:-1.7px;font-weight:750}h1 em{font-style:normal;color:#046acb}.summary{position:absolute;left:49px;right:49px;top:516px;font-size:17px;line-height:1.5;color:#555}.tags{position:absolute;left:49px;bottom:65px;display:flex;gap:9px}.tags span{border-radius:5px;padding:6px 11px;color:#006bcc;background:#bdd9f0;font-size:14px;font-weight:650}</style><style>/* 本次平台的局部排版 */h1{top:278px;font-size:56px;line-height:1.16;letter-spacing:-1px}.summary{top:514px;font-size:19px;line-height:1.65}.initial{font-size:11px}.tags{bottom:60px}</style></head><body><main id="cover"><svg class="arcs" viewBox="0 0 600 800" xmlns="http://www.w3.org/2000/svg"><defs><filter id="blur"><feGaussianBlur stdDeviation="2.7"/></filter></defs><g fill="none" stroke="#ffffff" opacity=".30" filter="url(#blur)"><ellipse cx="345" cy="388" rx="337" ry="210" transform="rotate(-14 345 388)" stroke-width="3"/><ellipse cx="355" cy="389" rx="325" ry="210" transform="rotate(-21 355 389)" stroke-width="3"/><path d="M-40 293C130 484 406 662 679 654" stroke-width="4"/></g><path d="M25 441C92 590 359 666 622 596" fill="none" stroke="#ffffff" stroke-width="1.7" opacity=".25"/></svg><div class="source"><div class="initial">{{source_initial}}</div><div><div class="author">{{author}}</div><div class="handle">{{source_caption}}</div></div></div><div class="ring"></div><div class="triangle"></div><h1>{{title1}}<br><em>{{title2}}</em><br>{{title3}}</h1><p class="summary">{{summary}}</p><div class="tags"><span>{{tag1}}</span><span>{{tag2}}</span></div></main></body></html>
```
# 内容与素材
中心主张:LLM 将原始资料整理为 wiki,并通过研究问答和输出回存持续积累知识。封面只呈现这一主张,不把全文塞进画面。
图片素材:无。
本次文案映射(br 表示语义换行):
source_initial:AK
author:Andrej Karpathy
source_caption:@karpathy
title1:让知识
title2:在每次提问后
title3:继续生长
summary:LLM 把原始资料编成 wiki,<br>再把问答结果存回知识库。<br>在 Obsidian 中查看与研究。
tag1:Markdown wiki
tag2:持续积累
完整原文:
作者:Andrej Karpathy(@karpathy)
来源:https://x.com/karpathy/status/2039805659525644595
LLM Knowledge Bases
Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:
Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.
IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).
Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.
Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.
Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.
Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.
Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.
TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.请用 skill「html-imagegen」完成下面的任务。若尚未安装:
git clone https://github.com/lindong28/lindong-skills
然后把其中的 html-imagegen/ 目录放到 Claude Code 的 ~/.claude/skills/ 或 Codex 的 ~/.agents/skills/(复制或符号链接均可),首次运行按 SKILL.md 安装 uv、Playwright 与 Chromium。
这是网站样例「流动蓝色 · 小红书封面(600×800)」实际使用的完整输入,内容已含,直接生成并导出 PNG 即可复现。
---- Prompt ----
# 角色与任务
你是中文封面编辑与排版设计师。根据本次完整原文制作一张小红书封面,输出完整HTML。
# 基本要求
输出独立完整HTML,导出区域唯一 id="cover",尺寸600×800像素,3:4,不透明底板,浏览器原生截图为同尺寸PNG。文字不越界、不重叠、不截断。使用系统字体、内嵌CSS与装饰,不调用图片模型或外部API。署名显示 Andrej Karpathy / @karpathy,不添加发布账号。转述忠实于下方原文,不伪装成直接引语。
# 视觉要求
近白灰蓝底,右上蓝雾与左下浅蓝渐变,柔和弧光贯穿中下部;左上小型来源,左对齐大标题用蓝色强调其中一段,摘要与蓝色小标签接在标题下方,空心环和淡三角点缀。种子标题从左49px、上311px、53px起排,三行只是起始组合;保留流动弧光与上下呼吸感,随内容调整强调位置、换行和摘要。
种子是排版起点,不是不可变的填槽表。保持本款风格与核心构图,按内容调整文字模块、数量、换行、字号、行距、间距和模块尺寸;代码中的像素值、行数及槽位数量均为起始值,不是硬内容预算。提炼一个清楚的中心主张,再选择支撑它的必要信息;不为填满空槽捏造事实、引语、标签、英文、账号、期号或步骤。无内容依据的可选模块及其分隔符一并移除,保留有意的留白,不用重复文案填空间。
纵向保留原种子的上下层级与有意留白,内容模块按本次文案适配。
本次平台完整HTML骨架(以下任务局部CSS已确定本次排版):
```html
<!DOCTYPE html>
<html lang="zh-CN"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>封面</title><style>*{box-sizing:border-box}html,body{margin:0}body{font-family:"PingFang SC","Microsoft YaHei",Arial,sans-serif}#cover{width:600px;height:800px;position:relative;overflow:hidden}h1,h2,p{margin:0}#cover{background:radial-gradient(ellipse at 91% 13%,#bedbed 0,transparent 45%),radial-gradient(ellipse at 11% 92%,#c6e2ee 0,transparent 31%),linear-gradient(145deg,#fbfbfb,#e4e8ed 75%,#e6edf4);color:#2f2f2f}.source{position:absolute;left:49px;top:48px;display:flex;gap:9px;align-items:flex-start}.initial{width:26px;height:26px;background:linear-gradient(140deg,#0875d2,#00beee);border-radius:50%;color:white;font-size:17px;display:flex;align-items:center;justify-content:center;font-weight:700}.author{font-size:15px;font-weight:700;margin-top:3px}.handle{font-size:14px;color:#0572d2;margin-top:12px}.arcs{position:absolute;inset:0;width:600px;height:800px;pointer-events:none}.ring{position:absolute;left:518px;top:159px;width:20px;height:20px;border-radius:50%;border:2px solid #75b1f5a0}.triangle{position:absolute;left:91px;bottom:121px;width:0;height:0;border-left:12px solid transparent;border-right:12px solid transparent;border-bottom:20px solid #90c8fa77}h1{position:absolute;left:49px;top:311px;font-size:53px;line-height:1.14;letter-spacing:-1.7px;font-weight:750}h1 em{font-style:normal;color:#046acb}.summary{position:absolute;left:49px;right:49px;top:516px;font-size:17px;line-height:1.5;color:#555}.tags{position:absolute;left:49px;bottom:65px;display:flex;gap:9px}.tags span{border-radius:5px;padding:6px 11px;color:#006bcc;background:#bdd9f0;font-size:14px;font-weight:650}</style><style>/* 本次平台的局部排版 */h1{top:278px;font-size:56px;line-height:1.16;letter-spacing:-1px}.summary{top:514px;font-size:19px;line-height:1.65}.initial{font-size:11px}.tags{bottom:60px}</style></head><body><main id="cover"><svg class="arcs" viewBox="0 0 600 800" xmlns="http://www.w3.org/2000/svg"><defs><filter id="blur"><feGaussianBlur stdDeviation="2.7"/></filter></defs><g fill="none" stroke="#ffffff" opacity=".30" filter="url(#blur)"><ellipse cx="345" cy="388" rx="337" ry="210" transform="rotate(-14 345 388)" stroke-width="3"/><ellipse cx="355" cy="389" rx="325" ry="210" transform="rotate(-21 355 389)" stroke-width="3"/><path d="M-40 293C130 484 406 662 679 654" stroke-width="4"/></g><path d="M25 441C92 590 359 666 622 596" fill="none" stroke="#ffffff" stroke-width="1.7" opacity=".25"/></svg><div class="source"><div class="initial">{{source_initial}}</div><div><div class="author">{{author}}</div><div class="handle">{{source_caption}}</div></div></div><div class="ring"></div><div class="triangle"></div><h1>{{title1}}<br><em>{{title2}}</em><br>{{title3}}</h1><p class="summary">{{summary}}</p><div class="tags"><span>{{tag1}}</span><span>{{tag2}}</span></div></main></body></html>
```
# 内容与素材
中心主张:LLM 将原始资料整理为 wiki,并通过研究问答和输出回存持续积累知识。封面只呈现这一主张,不把全文塞进画面。
图片素材:无。
本次文案映射(br 表示语义换行):
source_initial:AK
author:Andrej Karpathy
source_caption:@karpathy
title1:让知识
title2:在每次提问后
title3:继续生长
summary:LLM 把原始资料编成 wiki,<br>再把问答结果存回知识库。<br>在 Obsidian 中查看与研究。
tag1:Markdown wiki
tag2:持续积累
完整原文:
作者:Andrej Karpathy(@karpathy)
来源:https://x.com/karpathy/status/2039805659525644595
LLM Knowledge Bases
Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:
Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.
IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).
Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.
Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.
Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.
Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.
Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.
TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.公众号封面
按四块阅读输入
角色与任务
你是中文封面编辑与排版设计师。根据本次完整原文制作一张微信公众号封面,输出完整HTML。
基本要求
输出独立完整HTML,导出区域唯一 id="cover",尺寸1050×450像素,21:9,不透明底板,浏览器原生截图为同尺寸PNG。文字不越界、不重叠、不截断。使用系统字体、内嵌CSS与装饰,不调用图片模型或外部API。署名显示 Andrej Karpathy / @karpathy,不添加发布账号。转述忠实于下方原文,不伪装成直接引语。
视觉要求
近白灰蓝底,右上蓝雾与左下浅蓝渐变,柔和弧光贯穿中下部;左上小型来源,左对齐大标题用蓝色强调其中一段,摘要与蓝色小标签接在标题下方,空心环和淡三角点缀。种子标题从左49px、上311px、53px起排,三行只是起始组合;保留流动弧光与上下呼吸感,随内容调整强调位置、换行和摘要。
种子是排版起点,不是不可变的填槽表。保持本款风格与核心构图,按内容调整文字模块、数量、换行、字号、行距、间距和模块尺寸;代码中的像素值、行数及槽位数量均为起始值,不是硬内容预算。提炼一个清楚的中心主张,再选择支撑它的必要信息;不为填满空槽捏造事实、引语、标签、英文、账号、期号或步骤。无内容依据的可选模块及其分隔符一并移除,保留有意的留白,不用重复文案填空间。
横向独立重排;保持配色、装饰语言与标题层级,不对竖图做缩放、裁切或拉伸。 左侧两行标题,右侧摘要与标签,弧光按横画幅重新布置。
本次平台完整HTML骨架(以下任务局部CSS已确定本次排版):
```html
<!DOCTYPE html>
<html lang="zh-CN"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>封面</title><style>*{box-sizing:border-box}html,body{margin:0}body{font-family:"PingFang SC","Microsoft YaHei",Arial,sans-serif}#cover{width:600px;height:800px;position:relative;overflow:hidden}h1,h2,p{margin:0}#cover{background:radial-gradient(ellipse at 91% 13%,#bedbed 0,transparent 45%),radial-gradient(ellipse at 11% 92%,#c6e2ee 0,transparent 31%),linear-gradient(145deg,#fbfbfb,#e4e8ed 75%,#e6edf4);color:#2f2f2f}.source{position:absolute;left:49px;top:48px;display:flex;gap:9px;align-items:flex-start}.initial{width:26px;height:26px;background:linear-gradient(140deg,#0875d2,#00beee);border-radius:50%;color:white;font-size:17px;display:flex;align-items:center;justify-content:center;font-weight:700}.author{font-size:15px;font-weight:700;margin-top:3px}.handle{font-size:14px;color:#0572d2;margin-top:12px}.arcs{position:absolute;inset:0;width:600px;height:800px;pointer-events:none}.ring{position:absolute;left:518px;top:159px;width:20px;height:20px;border-radius:50%;border:2px solid #75b1f5a0}.triangle{position:absolute;left:91px;bottom:121px;width:0;height:0;border-left:12px solid transparent;border-right:12px solid transparent;border-bottom:20px solid #90c8fa77}h1{position:absolute;left:49px;top:311px;font-size:53px;line-height:1.14;letter-spacing:-1.7px;font-weight:750}h1 em{font-style:normal;color:#046acb}.summary{position:absolute;left:49px;right:49px;top:516px;font-size:17px;line-height:1.5;color:#555}.tags{position:absolute;left:49px;bottom:65px;display:flex;gap:9px}.tags span{border-radius:5px;padding:6px 11px;color:#006bcc;background:#bdd9f0;font-size:14px;font-weight:650}</style><style>/* 本次平台的局部排版 */#cover{width:1050px;height:450px}.arcs{width:1050px;height:450px}.source{left:54px;top:38px}.initial{font-size:11px}.author{font-size:17px}.handle{font-size:15px;margin-top:9px}h1{left:54px;top:166px;font-size:65px;line-height:1.15;letter-spacing:-1px}.summary{left:622px;right:54px;top:182px;font-size:22px;line-height:1.65}.tags{left:622px;bottom:69px}.tags span{font-size:15px}.ring{left:947px;top:64px;width:24px;height:24px}.triangle{left:519px;bottom:65px}</style></head><body><main id="cover"><svg class="arcs" viewBox="0 0 1050 450" xmlns="http://www.w3.org/2000/svg"><defs><filter id="blur"><feGaussianBlur stdDeviation="2.7"/></filter></defs><g fill="none" stroke="#ffffff" opacity=".30" filter="url(#blur)"><ellipse cx="600" cy="210" rx="590" ry="170" transform="rotate(-14 600 210)" stroke-width="3"/><ellipse cx="610" cy="200" rx="575" ry="165" transform="rotate(-21 610 200)" stroke-width="3"/><path d="M-40 120C220 285 690 450 1090 350" stroke-width="4"/></g><path d="M25 215C200 405 680 445 1050 325" fill="none" stroke="#ffffff" stroke-width="1.7" opacity=".25"/></svg><div class="source"><div class="initial">{{source_initial}}</div><div><div class="author">{{author}}</div><div class="handle">{{source_caption}}</div></div></div><div class="ring"></div><div class="triangle"></div><h1>{{title1}}<br><em>{{title3}}</em></h1><p class="summary">{{summary}}</p><div class="tags"><span>{{tag1}}</span><span>{{tag2}}</span></div></main></body></html>
```内容与素材
中心主张:LLM 将原始资料整理为 wiki,并通过研究问答和输出回存持续积累知识。封面只呈现这一主张,不把全文塞进画面。 图片素材:无。 本次文案映射(br 表示语义换行): source_initial:AK author:Andrej Karpathy source_caption:@karpathy title1:让知识 title2:在每次提问后 title3:继续生长 summary:LLM 把原始资料编成 wiki,<br>再把问答结果存回知识库。<br>在 Obsidian 中查看与研究。 tag1:Markdown wiki tag2:持续积累 完整原文: 作者:Andrej Karpathy(@karpathy) 来源:https://x.com/karpathy/status/2039805659525644595 LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.
# 角色与任务
你是中文封面编辑与排版设计师。根据本次完整原文制作一张微信公众号封面,输出完整HTML。
# 基本要求
输出独立完整HTML,导出区域唯一 id="cover",尺寸1050×450像素,21:9,不透明底板,浏览器原生截图为同尺寸PNG。文字不越界、不重叠、不截断。使用系统字体、内嵌CSS与装饰,不调用图片模型或外部API。署名显示 Andrej Karpathy / @karpathy,不添加发布账号。转述忠实于下方原文,不伪装成直接引语。
# 视觉要求
近白灰蓝底,右上蓝雾与左下浅蓝渐变,柔和弧光贯穿中下部;左上小型来源,左对齐大标题用蓝色强调其中一段,摘要与蓝色小标签接在标题下方,空心环和淡三角点缀。种子标题从左49px、上311px、53px起排,三行只是起始组合;保留流动弧光与上下呼吸感,随内容调整强调位置、换行和摘要。
种子是排版起点,不是不可变的填槽表。保持本款风格与核心构图,按内容调整文字模块、数量、换行、字号、行距、间距和模块尺寸;代码中的像素值、行数及槽位数量均为起始值,不是硬内容预算。提炼一个清楚的中心主张,再选择支撑它的必要信息;不为填满空槽捏造事实、引语、标签、英文、账号、期号或步骤。无内容依据的可选模块及其分隔符一并移除,保留有意的留白,不用重复文案填空间。
横向独立重排;保持配色、装饰语言与标题层级,不对竖图做缩放、裁切或拉伸。 左侧两行标题,右侧摘要与标签,弧光按横画幅重新布置。
本次平台完整HTML骨架(以下任务局部CSS已确定本次排版):
```html
<!DOCTYPE html>
<html lang="zh-CN"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>封面</title><style>*{box-sizing:border-box}html,body{margin:0}body{font-family:"PingFang SC","Microsoft YaHei",Arial,sans-serif}#cover{width:600px;height:800px;position:relative;overflow:hidden}h1,h2,p{margin:0}#cover{background:radial-gradient(ellipse at 91% 13%,#bedbed 0,transparent 45%),radial-gradient(ellipse at 11% 92%,#c6e2ee 0,transparent 31%),linear-gradient(145deg,#fbfbfb,#e4e8ed 75%,#e6edf4);color:#2f2f2f}.source{position:absolute;left:49px;top:48px;display:flex;gap:9px;align-items:flex-start}.initial{width:26px;height:26px;background:linear-gradient(140deg,#0875d2,#00beee);border-radius:50%;color:white;font-size:17px;display:flex;align-items:center;justify-content:center;font-weight:700}.author{font-size:15px;font-weight:700;margin-top:3px}.handle{font-size:14px;color:#0572d2;margin-top:12px}.arcs{position:absolute;inset:0;width:600px;height:800px;pointer-events:none}.ring{position:absolute;left:518px;top:159px;width:20px;height:20px;border-radius:50%;border:2px solid #75b1f5a0}.triangle{position:absolute;left:91px;bottom:121px;width:0;height:0;border-left:12px solid transparent;border-right:12px solid transparent;border-bottom:20px solid #90c8fa77}h1{position:absolute;left:49px;top:311px;font-size:53px;line-height:1.14;letter-spacing:-1.7px;font-weight:750}h1 em{font-style:normal;color:#046acb}.summary{position:absolute;left:49px;right:49px;top:516px;font-size:17px;line-height:1.5;color:#555}.tags{position:absolute;left:49px;bottom:65px;display:flex;gap:9px}.tags span{border-radius:5px;padding:6px 11px;color:#006bcc;background:#bdd9f0;font-size:14px;font-weight:650}</style><style>/* 本次平台的局部排版 */#cover{width:1050px;height:450px}.arcs{width:1050px;height:450px}.source{left:54px;top:38px}.initial{font-size:11px}.author{font-size:17px}.handle{font-size:15px;margin-top:9px}h1{left:54px;top:166px;font-size:65px;line-height:1.15;letter-spacing:-1px}.summary{left:622px;right:54px;top:182px;font-size:22px;line-height:1.65}.tags{left:622px;bottom:69px}.tags span{font-size:15px}.ring{left:947px;top:64px;width:24px;height:24px}.triangle{left:519px;bottom:65px}</style></head><body><main id="cover"><svg class="arcs" viewBox="0 0 1050 450" xmlns="http://www.w3.org/2000/svg"><defs><filter id="blur"><feGaussianBlur stdDeviation="2.7"/></filter></defs><g fill="none" stroke="#ffffff" opacity=".30" filter="url(#blur)"><ellipse cx="600" cy="210" rx="590" ry="170" transform="rotate(-14 600 210)" stroke-width="3"/><ellipse cx="610" cy="200" rx="575" ry="165" transform="rotate(-21 610 200)" stroke-width="3"/><path d="M-40 120C220 285 690 450 1090 350" stroke-width="4"/></g><path d="M25 215C200 405 680 445 1050 325" fill="none" stroke="#ffffff" stroke-width="1.7" opacity=".25"/></svg><div class="source"><div class="initial">{{source_initial}}</div><div><div class="author">{{author}}</div><div class="handle">{{source_caption}}</div></div></div><div class="ring"></div><div class="triangle"></div><h1>{{title1}}<br><em>{{title3}}</em></h1><p class="summary">{{summary}}</p><div class="tags"><span>{{tag1}}</span><span>{{tag2}}</span></div></main></body></html>
```
# 内容与素材
中心主张:LLM 将原始资料整理为 wiki,并通过研究问答和输出回存持续积累知识。封面只呈现这一主张,不把全文塞进画面。
图片素材:无。
本次文案映射(br 表示语义换行):
source_initial:AK
author:Andrej Karpathy
source_caption:@karpathy
title1:让知识
title2:在每次提问后
title3:继续生长
summary:LLM 把原始资料编成 wiki,<br>再把问答结果存回知识库。<br>在 Obsidian 中查看与研究。
tag1:Markdown wiki
tag2:持续积累
完整原文:
作者:Andrej Karpathy(@karpathy)
来源:https://x.com/karpathy/status/2039805659525644595
LLM Knowledge Bases
Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:
Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.
IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).
Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.
Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.
Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.
Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.
Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.
TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.请用 skill「html-imagegen」完成下面的任务。若尚未安装:
git clone https://github.com/lindong28/lindong-skills
然后把其中的 html-imagegen/ 目录放到 Claude Code 的 ~/.claude/skills/ 或 Codex 的 ~/.agents/skills/(复制或符号链接均可),首次运行按 SKILL.md 安装 uv、Playwright 与 Chromium。
这是网站样例「流动蓝色 · 公众号封面(1050×450)」实际使用的完整输入,内容已含,直接生成并导出 PNG 即可复现。
---- Prompt ----
# 角色与任务
你是中文封面编辑与排版设计师。根据本次完整原文制作一张微信公众号封面,输出完整HTML。
# 基本要求
输出独立完整HTML,导出区域唯一 id="cover",尺寸1050×450像素,21:9,不透明底板,浏览器原生截图为同尺寸PNG。文字不越界、不重叠、不截断。使用系统字体、内嵌CSS与装饰,不调用图片模型或外部API。署名显示 Andrej Karpathy / @karpathy,不添加发布账号。转述忠实于下方原文,不伪装成直接引语。
# 视觉要求
近白灰蓝底,右上蓝雾与左下浅蓝渐变,柔和弧光贯穿中下部;左上小型来源,左对齐大标题用蓝色强调其中一段,摘要与蓝色小标签接在标题下方,空心环和淡三角点缀。种子标题从左49px、上311px、53px起排,三行只是起始组合;保留流动弧光与上下呼吸感,随内容调整强调位置、换行和摘要。
种子是排版起点,不是不可变的填槽表。保持本款风格与核心构图,按内容调整文字模块、数量、换行、字号、行距、间距和模块尺寸;代码中的像素值、行数及槽位数量均为起始值,不是硬内容预算。提炼一个清楚的中心主张,再选择支撑它的必要信息;不为填满空槽捏造事实、引语、标签、英文、账号、期号或步骤。无内容依据的可选模块及其分隔符一并移除,保留有意的留白,不用重复文案填空间。
横向独立重排;保持配色、装饰语言与标题层级,不对竖图做缩放、裁切或拉伸。 左侧两行标题,右侧摘要与标签,弧光按横画幅重新布置。
本次平台完整HTML骨架(以下任务局部CSS已确定本次排版):
```html
<!DOCTYPE html>
<html lang="zh-CN"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>封面</title><style>*{box-sizing:border-box}html,body{margin:0}body{font-family:"PingFang SC","Microsoft YaHei",Arial,sans-serif}#cover{width:600px;height:800px;position:relative;overflow:hidden}h1,h2,p{margin:0}#cover{background:radial-gradient(ellipse at 91% 13%,#bedbed 0,transparent 45%),radial-gradient(ellipse at 11% 92%,#c6e2ee 0,transparent 31%),linear-gradient(145deg,#fbfbfb,#e4e8ed 75%,#e6edf4);color:#2f2f2f}.source{position:absolute;left:49px;top:48px;display:flex;gap:9px;align-items:flex-start}.initial{width:26px;height:26px;background:linear-gradient(140deg,#0875d2,#00beee);border-radius:50%;color:white;font-size:17px;display:flex;align-items:center;justify-content:center;font-weight:700}.author{font-size:15px;font-weight:700;margin-top:3px}.handle{font-size:14px;color:#0572d2;margin-top:12px}.arcs{position:absolute;inset:0;width:600px;height:800px;pointer-events:none}.ring{position:absolute;left:518px;top:159px;width:20px;height:20px;border-radius:50%;border:2px solid #75b1f5a0}.triangle{position:absolute;left:91px;bottom:121px;width:0;height:0;border-left:12px solid transparent;border-right:12px solid transparent;border-bottom:20px solid #90c8fa77}h1{position:absolute;left:49px;top:311px;font-size:53px;line-height:1.14;letter-spacing:-1.7px;font-weight:750}h1 em{font-style:normal;color:#046acb}.summary{position:absolute;left:49px;right:49px;top:516px;font-size:17px;line-height:1.5;color:#555}.tags{position:absolute;left:49px;bottom:65px;display:flex;gap:9px}.tags span{border-radius:5px;padding:6px 11px;color:#006bcc;background:#bdd9f0;font-size:14px;font-weight:650}</style><style>/* 本次平台的局部排版 */#cover{width:1050px;height:450px}.arcs{width:1050px;height:450px}.source{left:54px;top:38px}.initial{font-size:11px}.author{font-size:17px}.handle{font-size:15px;margin-top:9px}h1{left:54px;top:166px;font-size:65px;line-height:1.15;letter-spacing:-1px}.summary{left:622px;right:54px;top:182px;font-size:22px;line-height:1.65}.tags{left:622px;bottom:69px}.tags span{font-size:15px}.ring{left:947px;top:64px;width:24px;height:24px}.triangle{left:519px;bottom:65px}</style></head><body><main id="cover"><svg class="arcs" viewBox="0 0 1050 450" xmlns="http://www.w3.org/2000/svg"><defs><filter id="blur"><feGaussianBlur stdDeviation="2.7"/></filter></defs><g fill="none" stroke="#ffffff" opacity=".30" filter="url(#blur)"><ellipse cx="600" cy="210" rx="590" ry="170" transform="rotate(-14 600 210)" stroke-width="3"/><ellipse cx="610" cy="200" rx="575" ry="165" transform="rotate(-21 610 200)" stroke-width="3"/><path d="M-40 120C220 285 690 450 1090 350" stroke-width="4"/></g><path d="M25 215C200 405 680 445 1050 325" fill="none" stroke="#ffffff" stroke-width="1.7" opacity=".25"/></svg><div class="source"><div class="initial">{{source_initial}}</div><div><div class="author">{{author}}</div><div class="handle">{{source_caption}}</div></div></div><div class="ring"></div><div class="triangle"></div><h1>{{title1}}<br><em>{{title3}}</em></h1><p class="summary">{{summary}}</p><div class="tags"><span>{{tag1}}</span><span>{{tag2}}</span></div></main></body></html>
```
# 内容与素材
中心主张:LLM 将原始资料整理为 wiki,并通过研究问答和输出回存持续积累知识。封面只呈现这一主张,不把全文塞进画面。
图片素材:无。
本次文案映射(br 表示语义换行):
source_initial:AK
author:Andrej Karpathy
source_caption:@karpathy
title1:让知识
title2:在每次提问后
title3:继续生长
summary:LLM 把原始资料编成 wiki,<br>再把问答结果存回知识库。<br>在 Obsidian 中查看与研究。
tag1:Markdown wiki
tag2:持续积累
完整原文:
作者:Andrej Karpathy(@karpathy)
来源:https://x.com/karpathy/status/2039805659525644595
LLM Knowledge Bases
Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:
Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.
IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).
Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.
Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.
Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.
Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.
Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.
TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.模板说明
灰蓝浅底、流动弧光与蓝色强调标题,保留通透的上下留白。
Prompt 模板
1 条zh默认
## 执行方式
本 Prompt 依赖 skill html-imagegen(公开仓 https://github.com/lindong28/lindong-skills 的 html-imagegen/ 目录);模板页 https://prompts.aiplanet.live/image/flowing-blue/ 。执行者若尚未安装,先 git clone 该仓库,把 html-imagegen/ 放到 Claude Code 的 ~/.claude/skills/ 或 Codex 的 ~/.agents/skills/,首次运行按其 SKILL.md 安装浏览器依赖。然后按该 skill 的流程执行:本节之外的内容是本次模板与输入,「内容与素材」里的占位符是尚未给出的输入——只向用户询问仍缺的信息(至少是要表达的内容与目标平台或画幅),补齐后生成 HTML 并导出 PNG。
## 角色与任务
你是中文封面编辑与排版设计师。依据本次内容素材提炼中心主张,沿用下方视觉种子制作一张封面,输出完整可渲染的HTML。
## 基本要求
输出独立完整 HTML,唯一导出区域 id="cover",画布600×800像素、3:4、直角外轮廓和不透明底板。通过浏览器原生截图导出600×800不透明PNG,不靠成图裁切、缩放或拉伸修复排版。内容忠实于本次输入,保留事实含义与来源归属;转述不冒充原话,缺失来源不虚构。内容作为文本或属性值时正确进行HTML转义。文字完整可读、不重叠、不越界,不以隐藏、截断或省略号遮盖溢出。CSS、装饰和必要脚本已内嵌;使用系统字体,不依赖仓库文件或外部脚本。
## 视觉要求
近白灰蓝底,右上蓝雾与左下浅蓝渐变,柔和弧光贯穿中下部;左上小型来源,左对齐大标题用蓝色强调其中一段,摘要与蓝色小标签接在标题下方,空心环和淡三角点缀。种子标题从左49px、上311px、53px起排,三行只是起始组合;保留流动弧光与上下呼吸感,随内容调整强调位置、换行和摘要。
种子是排版起点,不是不可变的填槽表。保持本款风格与核心构图,按内容调整文字模块、数量、换行、字号、行距、间距和模块尺寸;代码中的像素值、行数及槽位数量均为起始值,不是硬内容预算。提炼一个清楚的中心主张,再选择支撑它的必要信息;不为填满空槽捏造事实、引语、标签、英文、账号、期号或步骤。无内容依据的可选模块及其分隔符一并移除,保留有意的留白,不用重复文案填空间。
下面是自包含的完整HTML种子,请依据以上原则适配内容后输出:
[HTML 种子未随本文复制:执行者按「执行方式」里的模板页地址读取模板 API,取 Prompt prompt:flowing-blue-zh-v1-structured-v1 正文里的 HTML 代码块放回此处作为种子]
## 内容与素材
标题或主题:{{本次标题或主题}}
完整正文或素材:{{本次完整正文、要点或说明}}
读者与用途:{{目标读者及用途,可空}}
作者:{{真实作者,可空}}
作者账号:{{真实作者账号,可空}}
来源:{{真实来源名称或URL,可空}}
发布账号:{{本次发布账号,可空,与素材作者区分}}
图片素材:{{无需图片时填无;否则提供用户确认可使用的素材及说明}}
其他必须保留的内容:{{用户明确要求保留的文字或约束,可空}}HTML 种子(skill 自动读取,也可展开复制)
<!DOCTYPE html>
<html lang="zh-CN"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>封面</title><style>*{box-sizing:border-box}html,body{margin:0}body{font-family:"PingFang SC","Microsoft YaHei",Arial,sans-serif}#cover{width:600px;height:800px;position:relative;overflow:hidden}h1,h2,p{margin:0}#cover{background:radial-gradient(ellipse at 91% 13%,#bedbed 0,transparent 45%),radial-gradient(ellipse at 11% 92%,#c6e2ee 0,transparent 31%),linear-gradient(145deg,#fbfbfb,#e4e8ed 75%,#e6edf4);color:#2f2f2f}.source{position:absolute;left:49px;top:48px;display:flex;gap:9px;align-items:flex-start}.initial{width:26px;height:26px;background:linear-gradient(140deg,#0875d2,#00beee);border-radius:50%;color:white;font-size:17px;display:flex;align-items:center;justify-content:center;font-weight:700}.author{font-size:15px;font-weight:700;margin-top:3px}.handle{font-size:14px;color:#0572d2;margin-top:12px}.arcs{position:absolute;inset:0;width:600px;height:800px;pointer-events:none}.ring{position:absolute;left:518px;top:159px;width:20px;height:20px;border-radius:50%;border:2px solid #75b1f5a0}.triangle{position:absolute;left:91px;bottom:121px;width:0;height:0;border-left:12px solid transparent;border-right:12px solid transparent;border-bottom:20px solid #90c8fa77}h1{position:absolute;left:49px;top:311px;font-size:53px;line-height:1.14;letter-spacing:-1.7px;font-weight:750}h1 em{font-style:normal;color:#046acb}.summary{position:absolute;left:49px;right:49px;top:516px;font-size:17px;line-height:1.5;color:#555}.tags{position:absolute;left:49px;bottom:65px;display:flex;gap:9px}.tags span{border-radius:5px;padding:6px 11px;color:#006bcc;background:#bdd9f0;font-size:14px;font-weight:650}</style></head><body><main id="cover"><svg class="arcs" viewBox="0 0 600 800" xmlns="http://www.w3.org/2000/svg"><defs><filter id="blur"><feGaussianBlur stdDeviation="2.7"/></filter></defs><g fill="none" stroke="#ffffff" opacity=".30" filter="url(#blur)"><ellipse cx="345" cy="388" rx="337" ry="210" transform="rotate(-14 345 388)" stroke-width="3"/><ellipse cx="355" cy="389" rx="325" ry="210" transform="rotate(-21 355 389)" stroke-width="3"/><path d="M-40 293C130 484 406 662 679 654" stroke-width="4"/></g><path d="M25 441C92 590 359 666 622 596" fill="none" stroke="#ffffff" stroke-width="1.7" opacity=".25"/></svg><div class="source"><div class="initial">{{source_initial}}</div><div><div class="author">{{author}}</div><div class="handle">{{source_caption}}</div></div></div><div class="ring"></div><div class="triangle"></div><h1>{{title1}}<br><em>{{title2}}</em><br>{{title3}}</h1><p class="summary">{{summary}}</p><div class="tags"><span>{{tag1}}</span><span>{{tag2}}</span></div></main></body></html>