research

Research: $ARGUMENTS

Turn "$ARGUMENTS" into genuine expertise using the unified research pipeline.

What This Does

1. Parses your input (topic, paragraph, URLs, or any mix)
2. Builds a structured question tree (research methodology before searching)
3. Discovers and fetches sources (content never enters context — goes to disk → .mv2)
4. Distills indexed knowledge into a compact expertise artifact via systematic querying
5. Loads the expertise into your context — you now know the topic

Output lives at ~/.claude/research/<topic-slug>/.

When This Skill Triggers

• "research X", "learn about X", "look up X before we start"
• "I need to understand X" (with or without URLs/context)
• Pasting a paragraph of context about something to learn
• "/research " explicitly

Step 1: Check for Existing Research

HOME_DIR=$(echo ~)
ls "$HOME_DIR/.claude/research/" 2>/dev/null

If the topic (or something close) already has a directory with expertise.md:

• Read and present the existing expertise doc
• Ask: "I have existing research on this. Load it, or re-research with fresh sources?"
• If load → read expertise.md, done
• If re-research → continue to Step 2

Step 2: Spawn the Research Agent

The research agent handles the full pipeline. Spawn it with the user's input:

Task(
  subagent_type="general-purpose",
  name="researcher",
  description="Research pipeline for topic",
  model="sonnet",
  prompt="""
You are the research-agent. Follow the pipeline in agents/research-agent.md exactly.

## Input
$ARGUMENTS

## Instructions
Run all 6 phases (Phase 0 is new and critical):
0. Search existing knowledge stores FIRST (~/.neo-research/knowledge/*.mv2)
1. Parse input → build question tree, annotate branches as [COVERED]/[PARTIAL]/[MISSING]
2. Discover sources (WebSearch) ONLY for [PARTIAL] and [MISSING] branches
3. Fetch → disk → index into .mv2 (zero context cost)
4. Distill: query .mv2 systematically → write expertise.md
5. Report results

Write all artifacts to ~/.claude/research/<slug>/.
Return the expertise.md content and a summary report when done.

## MCP Tools Available
Use ToolSearch to load: rlm_search, rlm_ask, rlm_ingest, rlm_exec, rlm_knowledge_status

## BM25 Query Rules (CRITICAL)
The .mv2 stores use Tantivy BM25. Multi-word queries silently return 0 hits
because Tantivy treats them as boolean AND. Use SINGLE KEYWORDS or OR-joined terms:
  - GOOD: mem.find("MeshResource", k=5)
  - GOOD: mem.find("MeshResource OR generateSphere OR texture", k=5)
  - BAD:  mem.find("MeshResource generateSphere texture", k=5) → 0 results
  - BAD:  mem.find("how to create a sphere mesh?", k=5) → 0 results

## Rules
- Search existing knowledge stores before any web research.
- Never read fetched content. Files go disk → knowledge store.
- Never use WebFetch. Use curl via Bash for fetching, WebSearch for discovery.
- Question tree before searching. Structure first.
- Be honest about gaps.
"""
)

Step 3: Load Expertise

When the agent returns:

1. Read ~/.claude/research/<slug>/expertise.md
2. Present it to the user
3. The agent (you) now has the expertise loaded in context

Step 4: Report

Research complete: <topic>
- Expertise: ~/.claude/research/<slug>/expertise.md
- Knowledge store: ~/.claude/research/<slug>/knowledge.mv2
- Deep-dive: rlm_search(query="...", project="<slug>")
- Reload later: /research load <topic>

After the agent returns, check sources.json for a coupling assessment. If coupling.recommendation == "skill-graph", append to your report:

Domain coupling: high (score N/5)
→ Create navigable skill graph: /create-skill-graph <slug>

Loading Existing Research

If the user says /research load <topic>:

1. Find the matching directory in ~/.claude/research/
2. Read expertise.md
3. Present it — agent now has the expertise

No need to re-fetch or re-index. The knowledge store is also available for rlm_search queries.

Listing Available Research

If the user says /research with no arguments or /research list:

HOME_DIR=$(echo ~)
ls -1 "$HOME_DIR/.claude/research/" 2>/dev/null

List what topics have been researched with their dates and sizes.

Rules

• One fetch attempt per URL. Fail → skip → move on.
• Don't paste full doc content. The expertise doc is the output, not raw pages.
• Check before re-fetching. Existing research is reusable.
• Stay on topic. Research what was asked. Don't branch into related topics uninvited.
• Quality over quantity. 10 good sources beat 50 mediocre ones.