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.
/plugin marketplace add farzadshbfn/knowledge-plugin
/plugin install kb@farzadshbfnclaude --plugin-dir /path/to/knowledge-plugin/kb-bootstrapCreates the config and directory structure. Run once per project.
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.
| 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 |
/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.
/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.
/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
/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.
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.
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:
// .claude/knowledge-base/config.json
{
"kb_roots": [
{ "name": "core", "path": "./knowledge" },
{ "name": "frontend", "path": "./frontend/knowledge" },
{ "name": "infra", "path": "./infra/knowledge" }
]
}Each KB lives next to its domain. /kb-find searches the most relevant KB first, so you get fewer tokens loaded and more relevant results. When a concept spans domains, @kb-name/path cross-references link them without duplicating content.
Tighter scoping means less noise for the progressive loader to filter through. Three focused domain KBs with cross-links are faster and more precise than one 200-note general KB. Everything stays connected, but retrieval is cheaper.
skills/
learn/ Multi-agent learning pipeline (scouter, searcher, assessor)
find/ Progressive 4-tier loader with deterministic Python scripts
compact/ KB restructuring and health checks
mint/ Skill conversion and plugin packaging
view/ Local web renderer (HTTP server + SPA + d3 graph)
monitor/ Access tracking and skill candidate analysis
bootstrap/ First-time project setup
Data retrieval is near-deterministic. Python scripts handle indexing, tree building, validation, and access tracking. The LLM orchestrates workflows and makes judgment calls. The scripts do the rest.
Built by Farzad Sharbafian
