learn
Research any topic online and create comprehensive learning guides with RAG-optimized indexes for your AI agents.
Why
AI agents work better when they have curated, pre-researched knowledge to draw from instead of searching the web on every question. /learn builds that knowledge base systematically - gathering sources, scoring them for quality, and synthesizing structured guides that agents can reference instantly.
Use cases:
- Learning a new technology before starting implementation
- Building a shared knowledge base across a team's AI tools
- Creating authoritative reference material from scattered online sources
- Producing guides that work as RAG context for Claude Code, OpenCode, and Codex
Installation
agentsys install learn
Requires agentsys to be set up in your project.
Quick Start
/learn react hooks
This searches the web for ~20 sources on React hooks, scores each source for authority and depth, fetches the top results, and writes a synthesized guide to agent-knowledge/react-hooks.md with a companion resources/react-hooks-sources.json containing full source metadata.
How It Works
The learn skill follows a six-stage methodology:
-
Progressive discovery - Funnel approach: broad queries for landscape mapping, focused queries for core content, deep queries for advanced material. Avoids noise from dumping all queries at once.
-
Quality scoring - Each source is scored on a 100-point scale across five dimensions: authority (3x weight), recency (2x), depth (2x), examples (2x), and uniqueness (1x). Official docs score highest; undated blog posts score lowest.
-
Just-in-time extraction - Only high-scoring sources get fetched. Summaries and key insights are extracted - never full content. This keeps token usage predictable and respects copyright.
-
Synthesis - A structured learning guide is generated with prerequisites, core concepts, code examples, common pitfalls, best practices, and further reading. Content is cross-referenced across sources, not copied from any single one.
-
RAG index - The master index (agent-knowledge/CLAUDE.md and AGENTS.md) is updated with the new topic, trigger phrases, and keyword mappings so agents can find relevant guides automatically.
-
Enhancement - Runs enhance:enhance-docs and enhance:enhance-prompts on the output to improve RAG retrieval quality. Skip with --no-enhance.
Usage
# Default depth (20 sources)
/learn recursion
# Deep research (40 sources)
/learn kubernetes networking --depth=deep
# Quick overview (10 sources)
/learn python decorators --depth=brief
# Skip enhancement pass
/learn typescript generics --no-enhance
Depth Levels
| Level | Sources | When to Use |
|---|
brief | 10 | Quick overview, time-sensitive topics |
medium | 20 | Balanced coverage (default) |
deep | 40 | Comprehensive research, complex topics |
Output Files
Each run creates or updates:
agent-knowledge/
CLAUDE.md # Master index (updated)
AGENTS.md # Master index for OpenCode/Codex (updated)
<topic-slug>.md # Synthesized learning guide
resources/
<topic-slug>-sources.json # Source metadata with quality scores
Existing Topics
If a guide already exists for the topic, you are prompted to either update the existing guide with new sources or start fresh.
Architecture
| Component | Type | Model | Role |
|---|
learn | command | - | Entry point, argument parsing |
learn-agent | agent | sonnet | Research coordination, web search, synthesis |
learn | skill | - | Research methodology, scoring rubric, templates |
Requirements
- agentsys runtime
- Web access (WebSearch and WebFetch tools)
- An
agent-knowledge/ directory in the workspace (created automatically)
Related Plugins
- agent-knowledge - Where guides are stored; contains existing research
- enhance - Post-processing for RAG optimization
- consult - For getting a second opinion on specific questions instead of building a full guide
License
MIT
agent-sh