Research Agent — Deep Investigation on Any Topic

A structured research workflow that turns a vague question into a comprehensive analysis. 5 research modes, each with a clear output format. Supports web search, source evaluation, and structured reporting.

Research Modes

How to Use

Quick Research (30 seconds)

"What is gstack?"
"Tell me about Claude Code skills"

→ Web search, extract key facts, 1-paragraph summary. No fluff.

Deep Dive (2-5 minutes)

"Research the AI coding agent landscape"
"Deep dive into Agent Skills standard"

→ Spawn subagent (Sonnet) with the Deep Dive prompt. Searches multiple sources, cross-references, identifies patterns, writes RESEARCH.md.

Compare (1-3 minutes)

"Claude Code vs Cursor vs Codex"
"RICE vs Kano vs ICE for prioritization"
"Notion vs Linear vs Jira"

→ Side-by-side comparison table with scoring across key dimensions. Includes a recommendation with reasoning.

Landscape Analysis (3-5 minutes)

"What open source projects exist for X?"
"Map the competitive landscape for X"
"What tools do PMs use for X?"

→ Categorized map of existing solutions. For each: what it does, what it misses, where the gap is.

Evaluate (2-3 minutes)

"Should we build on X or Y?"
"Is it worth adopting X?"
"Pros and cons of using X for our case"

→ Decision matrix scoring across dimensions (cost, effort, risk, fit, longevity). Recommendation with confidence level.

Phase Details

Deep Dive Prompt

Spawn a subagent (Sonnet) with this research methodology:

1. Define the question. Restate the research question. What specifically are we trying to find out?

1. Source gathering. Search for:
• Official docs / primary sources (most reliable)
• Community discussions (Reddit, HN, Discord — real user opinions)
• Technical analysis (blog posts, benchmarks, comparisons)
• GitHub metrics (stars, activity, issues, contributors)
• Commercial context (funding, team, business model)

1. Source evaluation. For each source:
• Credibility: official vs community vs opinion
• Recency: when was this published/updated?
• Bias: does the author have a stake in the outcome?

1. Pattern extraction. What themes emerge across sources?
• Points of agreement (high confidence)
• Points of disagreement (needs further investigation)
• Gaps in available information

1. Structured output. Write RESEARCH.md with:
• Executive summary (3-5 sentences)
• Key findings (numbered, with sources)
• Detailed analysis (organized by theme)
• Gaps and caveats (what we couldn't verify)
• Recommendation (if applicable)
• Sources (with URLs)

Compare Prompt

For comparing N items across M dimensions:

1. Define comparison axis. What dimensions matter for this decision?
• Functional: what can it do?
• Performance: how fast/reliable?
• Cost: pricing model, free tier?
• Ecosystem: integrations, community, docs?
• Maturity: how battle-tested?

1. Score each item (1-5 per dimension):

```

| Dimension | Option A | Option B | Option C |

|---------------|----------|----------|----------|

| Feature set | ⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐⭐⭐ |

| Ease of use | ⭐⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐ |

```

1. Context-specific recommendation. Not "A is best" but "A is best IF you need X, B if you need Y."

Landscape Prompt

For mapping a space:

1. Categorize solutions:
• Direct competitors (same approach, same users)
• Adjacent tools (different approach, overlapping use case)
• Workarounds (not products, but how people solve it today)
• Emerging (new, not proven yet)

1. For each solution:
• What it does (1 sentence)
• What it does well (strength)
• What it misses (gap)
• Who should use it (ideal user)

1. Identify the gap. Where is nobody doing a good job? That's the opportunity.

Output Files

• RESEARCH.md — Deep dive report (full analysis with sources)
• Comparison results go to stdout (capture in conversation)
• Landscape maps go to stdout or LANDSCAPE.md if long

Model Selection

Tips

• Be specific. "Research AI" is too broad. "Research AI coding agents for solo developers" is actionable.
• State your goal. "I need to decide between X and Y" gives the research direction.
• Time-box it. "Give me the top 5, not top 50" keeps it focused.
• Ask for sources. "Show me where you found this" for verification.