Krudony Deep Research
Research engine with Graph-of-Thoughts for topics where training data is outdated.
Quick Start
Standard Mode
- CLASSIFY → Type A/B/C/D
- SCOPE → Boundaries + success criteria
- HYPOTHESIZE → 3-5 testable hypotheses
- PLAN → Search queries + source strategy
- RETRIEVE → Parallel execution (ALL searches in ONE message)
- TRIANGULATE → Cross-verify, assign C1/C2/C3 claims
- SYNTHESIZE → Insights + SO WHAT/NOW WHAT
- RED TEAM → Counter-evidence (depth 3+)
- PACKAGE → Report with citations
Creative Mode
- ABSTRACT → Topic to core function
- MAP → 3-5 analogous domains
- SEARCH → Cross-industry parallel
- GENERALIZE → Transferable principles
- SYNTHESIZE → Apply to original topic
Trigger: "creative mode", "cross-industry", "what do others do", innovation needed
Phase 0: Fast-Path Routing
| Type | Characteristics | Process |
|---|---|---|
| A (LOOKUP) | Single fact | WebSearch → Answer |
| B (SYNTHESIS) | Multi-fact aggregation | 3 phases |
| C (ANALYSIS) | Judgment needed | 6 phases |
| D (INVESTIGATION) | Novel/conflicting | Full 8 phases + Red Team |
Classify BEFORE proceeding.
Intensity Tiers
| Tier | Sources | When |
|---|---|---|
| Quick | 5-10 | Known answer, single source |
| Standard | 10-20 | Multi-faceted, moderate |
| Deep | 20-30 | Novel, high stakes |
| Exhaustive | 30+ | Critical decision |
| Creative | 15-25 | Cross-industry innovation |
Parallel Search (MANDATORY)
WRONG: WebSearch #1 → wait → WebSearch #2 → wait
CORRECT (single message):
WebSearch: "AI coding tools December 2025"
WebSearch: "Claude Code vs Cursor comparison"
WebSearch: "[topic] limitations challenges"
Task(agent): Academic paper analysis
Claim Types
| Type | Requirements |
|---|---|
| C1 Critical | Quote + citation + 2+ independent sources + confidence |
| C2 Supporting | Citation required |
| C3 Context | Cite if contested |
Independence Rule: 5 articles citing same report = ONE source.
Confidence Levels
| Level | Criteria |
|---|---|
| HIGH (90%+) | 3+ A/B sources, n>1000, replicated |
| MEDIUM (60-90%) | Single strong OR multiple weaker |
| LOW (30-60%) | Preliminary, expert opinion |
| SPECULATIVE (<30%) | Single weak, theoretical |
Anti-Hallucination
- Every C1 MUST cite [N] immediately
- Use "According to [1]..." format
- Never "research suggests..." without citation
- Admit: "No sources found for X"
Quality Gates
| Gate | Requirement |
|---|---|
| Source count | Per intensity tier |
| Independence | C1 claims in 2+ independent sources |
| Recency | < 3 months for AI/tech |
| Citation coverage | Every C1 has [N] + confidence |
| Implications | Every finding has SO WHAT/NOW WHAT |
Termination
Stop when any 2 are true:
- Each subquestion meets source minimums
- Last 5 queries yield <10% new info
- All C1 claims meet independence rule
- Budget caps hit
Autonomy
Default: Proceed autonomously. Only ask if query is incomprehensible or contradictory.
Resources
| Resource | Purpose |
|---|---|
| methodology.md | Full 8-phase details, Creative Mode, agent templates |
| report_template.md | Thai report structure + output format |
source_evaluator.py |
Credibility scoring (0-100) |
validate_report.py |
9-check quality validation |
Related Skills (Optional)
| When | Suggest |
|---|---|
| Cross-industry innovation | /generate-creative-ideas - Combinatorial Engine |
| Technical contradiction found | /triz - systematic innovation |
| Need to explain findings | /explain-concepts - teach concepts |
| Strategic analysis needed | /manage-business-strategy - SWOT, Porter's |
| Business model research | /design-business-model - BMC, Lean Canvas |
Note: These skills are optional. Deep-research works standalone for fact-finding.