Knowledge Plugin for Claude
You read an article, learn something new, move on. A month later you've forgotten half of it. Claude forgot all of it the moment the conversation ended.
This plugin gives you a shared knowledge base. When you learn something, Claude learns it too, checks it against what you both already know, and keeps it around for next time.
Source code and tests: knowledge-skills.
Install
From marketplace
/plugin marketplace add farzadshbfn/knowledge-plugin
/plugin install kb@farzadshbfn
Standalone
claude --plugin-dir /path/to/knowledge-plugin
First run
/kb-bootstrap
Creates the config and directory structure. Run once per project.
Why This Exists
Tech moves fast. New frameworks, new conventions, a constant stream of articles and release notes. Even when you do read something useful, it's gone from your head a month later.
LLMs can help, but they forget everything between conversations and their training data lags behind. Context windows are expensive and finite. You end up re-explaining the same things, re-looking-up the same references, losing the insights you already earned.
This plugin fills that hole:
- You only learn what you don't already know. New information is checked against your existing knowledge, so you're not re-reading things you've already internalized.
- Knowledge survives across conversations, projects, and machines. Plain markdown files you own.
- Topics you keep coming back to get flagged as skill candidates. Your most-used knowledge turns into dedicated, reusable skills without manual curation.
- Retrieval is progressive. It reads just enough to answer the question, not your entire KB. Context stays cheap.
Skills
| Skill | Purpose |
|---|
/kb-learn | Learn from articles, research topics, fix KB errors |
/kb-find | Look up existing knowledge with progressive, token-efficient loading |
/kb-compact | Split oversized notes, unify terminology, fix indexes |
/kb-mint | Convert KB topics into skills, package as distributable plugins |
/kb-view | Local web renderer with fuzzy search, markdown rendering, and knowledge graph |
/kb-monitor | Track access patterns, surface skill candidates, check skill health |
/kb-bootstrap | First-time setup, run once per project |
How It Works
Learn
/kb-learn runs a multi-agent pipeline:
- Scouter finds related notes in your KB (including counter-evidence via
--challenge)
- Searcher pulls context from the web
- Assessor evaluates claims against your existing knowledge + web evidence
- New or updated notes are written with proper cross-references and changelog entries
/kb-learn article <url> # Extract and assess claims from an article
/kb-learn topic "raft consensus" # Research a topic, fill gaps in your KB
/kb-learn fix "X actually works like Y" # Correct a mistake
After learning, ask Claude to summarize what it found, explain trade-offs, or compare against what you already knew.
Find
/kb-find progressively loads only what's relevant:
Tier 1 → Topic tree (names + descriptions only)
Tier 2 → Frontmatter scan (first 15 lines of candidates)
Tier 3 → TOC + targeted sections
Tier 4 → Full note read (only when necessary)
A lookup in a 10,000-line KB might only load 50 lines into context. Challenge mode (--challenge) looks for contradicting evidence.
View
/kb-view starts a local HTTP server with a single-page app:
- Sidebar tree navigation across all configured KBs
- Fuzzy search across all notes
- Markdown rendering with syntax highlighting and Mermaid diagrams
- Interactive knowledge graph (d3-based) showing connections between notes
- Light/dark/system theme
Compact and Evolve
/kb-compact keeps your KB healthy as it grows. It splits large notes, merges duplicates, unifies terminology, and rebuilds indexes. --deep mode traverses the entire KB bottom-up.
From knowledge to skills
Access patterns are tracked across conversations. When topics keep coming back, they're flagged as skill candidates. These are topics mature enough to be a dedicated skill rather than a loose collection of notes.
/kb-mint converts those candidates into standalone skills or packaged plugins for distribution. The pipeline: raw notes, then structured knowledge, then reusable skill, then distributable plugin.
Multi-KB support — the proximity principle
A single centralized KB works until you're juggling multiple domains, then you're loading irrelevant context every time.
Multi-KB lets you co-locate knowledge with the domain it serves:
