Deep Research

Deep Research Skill - Lutum Veritas Inspired

Trigger

• "deep research on [topic]"
• "thorough research on [topic]"
• "/deep-research [topic]"
• Any request requiring comprehensive, verified answers with sources

Core Philosophy

Key innovations from Lutum Veritas:

1. Recursive context passing - Each research point knows what previous ones discovered
2. Claim Audits - Self-verification instead of blind assertions
3. Dual-scraping phases - First for answer, second for verification
4. Source transparency - Every claim traced to its source

Mode Selection

Mode A: Quick Verify (Ask Mode)

When: User needs a verified answer, not a full research report Time: ~60-90 seconds Use for: Factual questions, single-topic verification

Mode B: Deep Research

When: User needs comprehensive analysis Time: ~5-15 minutes Use for: Complex topics, multiple angles, full reports

Ask user which mode if unclear:

• "Do you want a quick verified answer (2 min) or comprehensive deep research (10 min)?"

Mode A: Quick Verify Pipeline (6 Stages)

C1: Intent Analysis → C2: Knowledge Gap → C3: Search Strategy → C4: Synthesis → C5: Claim Audit → C6: Verification

Stage C1: Intent Analysis

Goal: Understand what the user really needs

Use Perplexity Ask to analyze the question:

Analyze this research question and identify:
1. Core intent - what is the user actually trying to understand?
2. Implicit requirements - what does "good answer" look like for this query?
3. Success criteria - how will we know the answer is complete?

Question: {user_question}

Stage C2: Knowledge Gap Analysis

Goal: Determine what must be searched vs. what model knows

For this question: {user_question}

What does the model's training data likely cover well?
What aspects definitely need fresh web search?
Be specific about knowledge gaps.

Stage C3: Search Strategy

Goal: Create targeted search plan

Execute parallel searches using Perplexity Search:

• Primary angle: {main_topic}
• Counter角度: {opposing_view}
• Latest updates: {topic} 2024 2025

Use Kagi as secondary source for:

• Technical docs, GitHub, forums
• Alternative viewpoints not in Perplexity

Stage C4: Answer Synthesis

Goal: Generate initial answer with inline citations

Format:

• Start with direct answer
• Use [1], [2], [3] for sources
• Distinguish established facts from interpretations
• Flag areas of uncertainty

Stage C5: Claim Audit (KEY INNOVATION)

Goal: Self-verify every claim

For each major claim in C4 output, run verification:

Claim: {specific_claim_from_answer}

Verify this claim. Search for contradictory evidence.
Return:
- VERIFIED: Source confirms claim
- CONTRADICTED: Sources disagree  
- UNCERTAIN: Insufficient evidence
- NUANCED: Claim is partially true but needs qualification

This is the Lutum Veritas magic - forces self-reflection instead of blind assertions.

Stage C6: Final Verification

Goal: Cross-check against second set of sources

Search for:

• "debunks [claim]"
• "myth about [topic]"
• "[topic] false"

Final output format:

## Answer

[Synthesized answer with verified claims]

## Sources
[1] Source title - URL
[2] Source title - URL

## Verification Notes
- Claim X: VERIFIED [source]
- Claim Y: NUANCED - [qualification]
- Claim Z: UNCERTAIN - [explanation]

Mode B: Deep Research Pipeline

Clarification → Research Plan → Deep Research (per point) → Final Synthesis

Step 1: Clarification

Ask smart follow-up questions to understand scope:

To research this topic effectively, I need to clarify:

1. Scope: Are we covering history, current state, future, or all?
2. Depth: Surface overview or detailed analysis?
3. Angle: Any specific perspective (pro/con, technical, business)?
4. Audience: Expert or general?
5. Time sensitivity: Need latest 2024/2025 info?

Topic: {topic}

Use AskUserQuestion tool.

Step 2: Research Plan

Generate structured investigation points:

Create a research plan for: {topic}

Break this into 4-8 independent research points that:
- Can be researched in parallel
- Together cover the topic comprehensively
- Have clear, distinct angles
- Build on each other (later points can reference earlier)

Format:
## Research Points

### Point 1: [Title]
- Focus: [what to investigate]
- Key questions: [2-3 questions]
- Expected sources: [type of sources]

...etc

Step 3: Deep Research (Per Point)

For EACH research point, execute:

3a. Think: Generate search queries for this point 3b. Search: Run Perplexity + Kagi searches
3c. Select: Pick best 3-5 URLs 3d. Synthesize: Create "dossier" for this point

CRITICAL: Pass context forward:

Context from previous research points:
{summary_of_points_1_to_n-1}

Now researching point {n}: {title}

How does this connect to previous findings?
What new angles does this point add?

Step 4: Final Synthesis

Cross-reference all dossiers:

Synthesize all research points into a comprehensive report:

## Previous Findings (Context)
{all_point_summaries}

## New Research: Point {n}
{current_dossier}

## Cross-Reference Task
- What connects this to previous findings?
- Any contradictions?
- What's the unified picture?

Final output structure:

# Deep Research: {Topic}

## Executive Summary
{2-3 paragraph overview}

## Key Findings
### Finding 1
- Evidence: [source]
- Confidence: HIGH/MEDIUM/LOW

### Finding 2
...

## Source Registry
| # | Source | Type | Relevance |
|---|--------|------|-----------|
| 1 | Title | article | 9/10 |
...

## Areas of Uncertainty
- [List claims with lower confidence]

## Conclusion
{ synthesised answer }

Source Quality Grading

Apply Lutum Veritas quality levels:

Cost Optimization (from Lutum Veritas philosophy)

• Use Perplexity Search for initial discovery (cheap)
• Use Perplexity Ask for synthesis (moderate)
• Use Kagi for technical/deep sources (secondary verification)
• Reserve Deep Research for complex topics only

Target costs:

• Quick Verify: ~5k tokens
• Deep Research: ~30-50k tokens

Verification Checklist

Before presenting any research, confirm:

• Each factual claim has inline citation
• Contradicting evidence sought and noted
• Confidence levels assigned to claims
• Sources are recent (unless historical topic)
• Alternative viewpoints represented
• Uncertainty explicitly acknowledged

Examples

Quick Verify:

• "Deep research: Is AI coding actually faster?"
• "Verify: What's the current state of Rust in web dev?"

Deep Research:

• "Do a deep research on: The true cost of AI coding assistants in enterprise"
• "Thorough research: Compare React vs Vue vs Svelte in 2025"

Fallback

If research is inconclusive:

## Limitations

- [Aspect that couldn't be verified]
- [Reason for uncertainty]
- [Recommendations for further research]

Always be honest about what couldn't be determined. This is the Lutum Veritas way - truth over confidence.