Research

Research Skill

Multi-backend research orchestrator. Routes to the cheapest effective tool for each task.

Modes

Workflow: Default Research

0. Research prompt gate — before --deep or external research dispatch, run /research-prompt-quality (CHECK-FIRST + GROUND). Do not ship flat or redundant prompts.
1. Check BrainLayer first — brain_search(query) may already have what you need
2. Launch researcher subagent in background:

   Task(subagent_type: "researcher", prompt: "Research: {topic}", run_in_background: true)
   

3. Continue working while research runs
4. Read results when notified — researcher saves to docs.local/research/[date]-[slug].md
5. Digest to BrainLayer if worth keeping: brain_digest(content) then delete the file

Workflow: Quick Research (inline)

No subagent — run directly in current context:

1. Run 3-5 WebSearch queries in parallel
2. WebFetch top 3-5 results
3. Synthesize inline
4. Continue working

Use when you need a quick answer, not a report.

Workflow: Company Research

For job leads, freelance prospects, meeting prep:

1. company_research_exa(company_name) — get company overview
2. web_search_exa("company_name funding team size tech stack") — deeper intel
3. brain_search("company_name") — check if we've seen them before
4. Output structured brief:
   • What they do (1 sentence)
   • Tech stack / relevant tech
   • Recent news / funding
   • Connection to our skills
   • Red flags

Workflow: Code Research

For library comparison, pattern discovery, architecture decisions:

1. BrainLayer check: brain_search("topic") — past decisions?
2. Exa code context: get_code_context_exa("pattern/library") — real code examples
3. CLI agent (Gemini): run.sh gemini "Compare X vs Y for [use case]" — free, detailed analysis
4. Optional: Cursor audit: run.sh cursor "Review this code pattern: ..." — GPT-5.2 perspective

Workflow: External Research (offload from Opus)

When the main Claude Code session is expensive Opus and you want cheap research:

# Gemini does the research, saves to file (FREE)
~/.claude/commands/cli-agents/scripts/run.sh gemini "Research: {full prompt}" docs.local/research/$(date +%Y%m%d)-research.md

Then read the output file. Gemini is free (1K/day) and good for general research.

Workflow: Deep Research (Claude Web-quality)

For comprehensive research matching Claude Web's 463-source depth:

1. Launch researcher subagent with explicit depth:

   Task(subagent_type: "researcher", prompt: "COMPREHENSIVE deep research on: {topic}. Target 40+ sources. Run 25+ search queries. Cross-reference all claims.", run_in_background: true)
   

2. Supplement with CLI agents for extra perspectives:

   run.sh gemini "Deep research on {topic} — focus on {angle A}" /tmp/research-gemini.md
   

3. Merge results — researcher report + Gemini output = comprehensive coverage
4. Store in BrainLayer — brain_store the synthesis for future retrieval

Workflow: Pre-PR Audit

Extra eyes before PR loop:

1. Gemini review: run.sh gemini "Review this diff for bugs, security issues, and missed edge cases: $(git diff main..HEAD)" — free
2. Cursor review: run.sh cursor "Audit this code change: ..." — GPT-5.2 perspective
3. Both in parallel — compare their findings
4. Fix issues before pushing

Workflow: Unified Multi-Platform Research

Use this when you want Claude Web and Gemini to research the same question against the same Drive-backed context.

1. Verify the project folders exist in Drive:

   python3 skills/golem-powers/research/_shared/drive-paths.py ensure-project-folders <project>
   

2. Compute the shared dispatch plan:

   bash skills/golem-powers/research/scripts/unified-dispatch.sh --project <project> --topic "Compare BrainLayer vs Mem0 for long-term memory"
   

3. Dispatch Claude Desktop with the shared Drive project:
   • /claude-desktop-research (was /claude-web-research — renamed 2026-04-30; alias active until 2026-05-30)
   • reads Drive/Research/<project>/context/
   • writes Drive/Research/<project>/results/R{NN}-claude-desktop-result.md
4. Dispatch Gemini with the same Drive project:
   • /gemini-research
   • reads the identical Drive folder ID for Drive/Research/<project>/context/
   • writes Drive/Research/<project>/results/R{NN}-gemini-result.md
5. Hand the paired results to /research-ab-test without platform guessing.

Example:

/research --unified "Compare BrainLayer vs Mem0 for long-term memory" --project brainlayer

Unified mode is successful only when both dispatches reference the same Drive folder ID and produce pre-attributed paired result paths.

Integration Points

Cost Summary

Default stack is 100% free: WebSearch + Exa free credits + Gemini CLI.

Output Location

Future: n8n Deep Research Pipeline

When n8n orchestrator is set up (packages/orchestrator), add:

• Scheduled research (e.g., weekly job market scan)
• 400+ source deep research via recursive search loops
• Automated BrainLayer ingestion of research results
• Research templates (company intel, tech comparison, market scan)

Template exists: n8n deep research workflow