You keep saving good articles and never reopening them: copy the link, paste into a notes app, tag it by hand, forget it. An agent can do that loop for you - send a link, get the body text summarised and archived as Markdown, with the most token-hungry step (stuffing a whole page of HTML into context) removed. This explains the save2kb Agent Skill: how it works, how to install it, what the knowledge base looks like.

💡 What You Will Learn

You keep saving good articles and never reopening them: copy the link, paste into a notes app, tag it by hand, forget it. An agent can do that loop for you - send a link, get the body text summarised

📜 Table of Contents

• 1. The three problems it targets
• 2. Why it saves tokens: three-tier reading
• 3. Install: standard Agent Skills package
• 4. What using it looks like
• 5. What the knowledge base becomes
• 6. Versus bookmarks and clippers
• FAQ

save2kb: Teach Your Agent to Read Articles, Save a Knowledge Base, and Waste Far Fewer Tokens

Bookmarks are easy to add and hard to revisit. Clippers save HTML, so body text arrives mixed with navigation and ads. And when you let a model read a page, the expensive part is rarely comprehension - it is the raw markup you pushed into the context window. save2kb is an Agent Skill that turns "read, summarise, archive" into a standard routine.

1. The three problems it targets

• Read and lost. Articles read inside a chat vanish with the window.
• You saved a web page, not content. Full HTML where the body is a few percent of the bytes.
• Context full of markup. Many quick implementations fetch an entire page into the prompt; tokens disappear into tags.

2. Why it saves tokens: three-tier reading

The first tier loads no images, renders no layout and parses no framework code - only the words. In the project's own measurements, one WeChat article came to about 4.05 million characters of source versus about 2,838 characters of body text; a clean English tech site went from roughly 630k characters of HTML to 250k.

Read that ratio correctly: it measures source bytes to body bytes, not a 99% cut on your bill. Your actual saving depends on how you read pages today.

3. Install: standard Agent Skills package

save2kb follows the SKILL.md convention (repo: simumu1/save2kb, MIT) and works with Hermes Agent, Claude Code, Codex CLI and other SKILL.md-compatible agents. Installation means placing the skill directory where your agent looks for skills:

git clone https://github.com/simumu1/save2kb
cp -r save2kb/skills/save2kb ~/.hermes/skills/   # confirm layout in the repo README

Update by pulling and overwriting. Check the repo README and your agent's skill docs for exact paths.

4. What using it looks like

1. You send a link.
2. The agent reads the body and reports key points, source and date.
3. You say "save to the knowledge base" (or pre-approve), and it writes a Markdown file.

Files carry YAML frontmatter - the part clippers get wrong, and the reason the archive stays machine-searchable:

---
title: "Article title"
source: "Site/Author"
url: "https://..."
date: 2026-07-09
tags: [AI, quant, SEO]
summary: "under 200 characters"
status: raw
---

5. What the knowledge base becomes

One file per article with a date in the name; automatic topic tags; an INDEX.md grouped by topic; and a status field of raw or classified, where a weekly job sweeps new items into a digest and marks them classified. That raw-to-classified loop is what stops a knowledge base from becoming write-only.

6. Versus bookmarks and clippers

FAQ

Do I need to code? No - install is copying a directory, and daily use is sending links.

Chinese sites? Supported, including WeChat and Zhihu-style pages; anti-scrape sites fall through to browser rendering.

Where does the KB live? A local directory you choose the first time (or your existing structure) and reuse afterwards.

Paywalled content? The skill extracts public pages only; respect each site's terms and copyright.

Can I convert this into money saved? Yes, with your own baseline: input tokens per article before versus after, times calls, times unit price.

Note: exact directory layout, commands and supported agents are governed by the GitHub README and the Agent Skills specification.

❓ FAQ

Do I need to code?

No - install is copying a skill directory; daily use is sending links.

How should I read the 99% claim?

It compares full source bytes to body bytes. Your real saving depends on your current reading method.

Where does the knowledge base live?

A local directory you specify on first use, or your existing folder structure.