deep-learn

name	deep-learn
description	Deep research a topic using parallel agents, then synthesize and save comprehensive knowledge to ~/.claude/learnings/.
argument-hint	<topic> - e.g., 'TanStack Router', 'Kubernetes networking', 'Rust ownership'
allowed-tools	["Bash","Read","Write","Edit","Glob","Grep","WebSearch","WebFetch","Task","TeamCreate","SendMessage","TaskCreate","TaskUpdate","TaskList","TaskGet","mcp__context7__resolve-library-id","mcp__context7__query-docs","mcp__gh_grep__searchGitHub","mcp__memory__create_entities","mcp__memory__search_nodes"]

deep-learn

Deep research a topic using parallel agents. Each agent explores a different dimension simultaneously, then results are synthesized into a comprehensive knowledge file.

Input

Topic string from $ARGUMENTS.

Process

1. Generate Slug & Check Existing

Convert topic to slug (lowercase, hyphens). Check ~/.claude/learnings/<slug>.md. If exists, read frontmatter. Ask user: Update, Replace, or Cancel.

mkdir -p ~/.claude/learnings

2. Analyze Topic & Choose Strategy

Determine the topic category, research dimensions, and agent count. Then decide how to orchestrate.

You have two orchestration modes — choose based on the situation:

Mode A: Subagents (Task tool, no team)

Fire-and-forget parallel agents. Each works independently, you synthesize after all return.

Best for:

• Independent research dimensions that don't build on each other
• Well-defined topics where you know what to look for upfront
• 2-4 agents with clear, non-overlapping assignments

Mode B: Team Agents (TeamCreate + coordinated tasks)

Create a team with shared task list. Agents can communicate, report partial findings, and you can dynamically assign follow-up work.

Best for:

• Broad or unfamiliar topics where initial research may reveal unexpected subtopics
• Iterative deepening: first wave scouts the landscape, second wave digs into what matters
• Complex topics where one agent's findings should influence another's direction
• 4+ agents or multi-round research

Decision guideline:

Signal	→ Mode
You can enumerate all research dimensions upfront	Subagents
Topic is well-scoped (e.g., "React useEffect cleanup")	Subagents
Topic is broad or vague (e.g., "Kubernetes networking")	Team
You might need a second research round based on findings	Team
Agent findings may overlap or conflict and need real-time coordination	Team

Agent count guideline:

Topic Type	Suggested Agents
Library/Framework	3-4: docs, code patterns, gotchas, ecosystem
Concept/Pattern	2-3: theory, implementations, comparisons
Tool/CLI	2-3: official docs, config recipes, troubleshooting
Language Feature	2-3: spec/docs, adoption patterns, edge cases
Complex/Broad Topic	4-5: split by subtopic areas

You are NOT locked into these. Use your judgment. Scale up or down as needed.

3. Execute Research

If using Subagents (Mode A):

Use the Task tool with subagent_type: "general-purpose" to spawn agents in parallel (multiple Task calls in a single message). Each gets a focused prompt and returns structured findings.

If using Team Agents (Mode B):

1. TeamCreate — create a research team
2. TaskCreate — create tasks for each research dimension
3. Task with team_name — spawn named agents, assign tasks
4. Monitor progress via TaskList. When an agent reports interesting findings via SendMessage, decide whether to:
   • Create follow-up tasks for deeper investigation
   • Redirect other agents to explore related areas
   • Spawn additional agents for newly discovered subtopics
5. When all tasks are complete, shut down agents and delete the team

Agent prompt template (both modes):

Research "<topic>" focusing on: <specific dimension>.

Use these tools as needed:
- WebSearch / WebFetch for web sources
- mcp__context7__resolve-library-id + query-docs for library docs
- mcp__gh_grep__searchGitHub for real-world code

Return your findings as structured markdown:
## <Dimension Name>
### Key Findings
- ...
### Code Examples
```lang
...

Sources

1. Title

Be thorough but concise. Prioritize accuracy and code correctness. Focus ONLY on your assigned dimension — other agents cover the rest.

### 4. Synthesize Results

After all agents return, merge their findings:

1. **Deduplicate** — remove overlapping content, keep the better version
2. **Cross-validate** — if agents disagree, flag the conflict or verify
3. **Organize** — structure into the standard knowledge format
4. **Attribute** — collect all sources from all agents

### 5. Save Knowledge File

Write to `~/.claude/learnings/<slug>.md`:

```markdown
---
topic: "<Original Topic Name>"
slug: "<slug>"
category: "<library|concept|tool|language-feature>"
created: "<YYYY-MM-DD>"
last_verified: "<YYYY-MM-DD>"
confidence: "<high|medium|low>"
tags: [<relevant, tags>]
sources_count: <N>
research_depth: "deep"
agents_used: <N>
strategy: "<subagents|team>"
---

# <Topic Name>

## TL;DR
<2-4 sentences: what it is, key capabilities, primary use case.>

## Core APIs / Concepts

### <Name>
- **Signature/Usage**: `<code>`
- **Purpose**: <one line>
- **Example**:
```<lang>
<minimal working example>

Patterns & Recipes

Gotchas

• : <What happens + fix/workaround>

Advanced Topics

<Deeper material that basic /learn would skip: internals, performance, edge cases, architecture decisions.>

Quick Reference

Sources

1. — <which agent found this>

### 6. Knowledge Graph Indexing (Optional)

If `mcp__memory__create_entities` is available:

```json
{
  "name": "<Topic Name>",
  "entityType": "learning",
  "observations": [
    "Saved to ~/.claude/learnings/<slug>.md",
    "Category: <category>",
    "Research depth: deep (<N> agents)",
    "Tags: <tag1>, <tag2>"
  ]
}

7. Report

Tell the user:

• File path
• How many agents were used and their dimensions
• Source count
• Key highlights (3-5 bullet points of most valuable findings)

Quality Rules

• Each agent's findings are cross-checked during synthesis. Conflicting info is resolved or flagged.
• Code examples must be correct. Mark untested code with // untested.
• Set confidence high only when multiple agents' sources agree.
• Include version numbers for libraries.
• The research_depth: "deep" and agents_used fields in frontmatter distinguish deep-learn output from regular /learn.