idea-checking

name	idea-checking
description	Validate research idea novelty and feasibility via deep search + Codex MCP (GPT-4.5) cross-review. Use after /initial-idea-generation when user says "检查idea", "validate idea", "check novelty", or wants to verify the idea before proceeding.
argument-hint	["idea-summary-or-reference-to-skill-1-output"]
allowed-tools	Bash(*), Read, Write, Grep, Glob, WebSearch, WebFetch, Agent, mcp__codex__codex, mcp__codex__codex-reply

Skill-2: Idea Checking

Validate and finalize the research idea: $ARGUMENTS

Overview

This skill takes the user-approved idea from skill-1 and performs two-stage validation:

1. Claude deep search — novelty and significance check, modify if needed (with user consent)
2. GPT-4.5 one-shot review — independent cross-check via Codex MCP

After both stages, combine feedback with user input, make final adjustments, and output idea01.md. No repeated review loops — one pass through each stage.

Workflow

Phase 1: Idea Intake & Confirmation

1. Retrieve the idea from current conversation (produced by skill-1)
2. Display to user: title, research question, innovation points, proposed method
3. Ask: "Is this the idea you want me to validate?"
4. Wait for user confirmation

Phase 2: Claude Deep Search — Novelty & Significance

2.1 Novelty Search

Run at least 8 distinct searches using WebSearch:

1. "[idea title keywords]" + "knowledge graph" + "graph RAG"
2. "[proposed method]" + "ontology" + "[domain]"
3. "[application domain]" + "chatbot" + "knowledge graph"
4. site:arxiv.org + key terms (last 12 months)
5. site:sciencedirect.com OR site:springer.com + key terms
6. Semantic Scholar: "[core concepts]" + "graph-based retrieval"
7. Chinese sources (if relevant): 知识图谱 + 检索增强 + [领域]
8. Patent/industry: "[key technology]" + system + implementation

For each search: review top 10 results, flag overlapping papers, record how our idea differs.

2.2 Significance Assessment

• Active research interest in the area? (recent publications)
• Does it fill a clearly identified gap?
• Useful to practitioners, not just academics?
• Well-scoped enough for clear results?

2.3 Issue Classification

Category	Description	Action
BLOCKER	Nearly identical work published	Must pivot or differentiate
WARNING	Partial overlap	Need to clarify differentiation
MINOR	Similar approach in different domain	Note in related work, not a problem
CLEAR	No significant overlap	Proceed

If BLOCKER or WARNING found:

• Report to user with specific papers and overlap analysis
• Suggest modifications
• Wait for user approval → apply changes
• Proceed to Phase 3 with the modified idea

If CLEAR: Proceed to Phase 3 directly.

Phase 3: GPT-4.5 One-Shot Review via Codex MCP

Submit the idea (after any Phase 2 modifications) to GPT-4.5 for a single independent review.

mcp__codex__codex:
  model: gpt-4.5
  prompt: |
    You are an independent academic reviewer for a Q1 journal (Advanced Engineering Informatics level).

    ## Research Idea

    **Title**: [title]
    **Research Question**: [question]
    **Innovation Points**: [list]
    **Proposed Method**: [method summary]
    **Domain**: [domain]
    **Tech Stack**: Ontology → Knowledge Graph → Graph RAG → Chatbot

    ## Review Task

    Give a single, thorough review. Be critical but constructive.

    1. **Novelty** (1-10): Is this genuinely new? Closest existing work?
    2. **Technical Feasibility** (1-10): Can this be implemented as described?
    3. **Research Significance** (1-10): Does this matter? Who benefits?
    4. **Methodology Soundness** (1-10): Is the approach logically sound?
    5. **Data Availability** (1-10): Can required data be obtained?
    6. **Overall Verdict**: ACCEPT / REVISE / REJECT
    7. **Top 3 Concerns**: Specific weaknesses
    8. **Top 3 Suggestions**: What would make it stronger

    Be honest. No flattery.

Save the GPT-4.5 response. Do not reply to it or start a loop.

Phase 4: Synthesize & Present to User

Combine Claude search results and GPT-4.5 review into one report:

## Validation Report

### Claude Deep Search
- Papers reviewed: [N]
- Novelty status: [CLEAR / WARNING / BLOCKER]
- Closest existing work: [paper + how ours differs]
- Modifications already made: [if any, from Phase 2]

### GPT-4.5 Review
- Novelty: [X/10], Feasibility: [X/10], Significance: [X/10]
- Verdict: [ACCEPT / REVISE / REJECT]
- Key concerns: [list]
- Suggestions: [list]

### Combined Assessment
[Areas where Claude and GPT-4.5 agree / disagree]
[Final recommendation: PROCEED / MODIFY FURTHER / RECONSIDER]
[If MODIFY: specific changes suggested]

Present to user. Wait for user feedback:

• User may accept as-is → go to Phase 5
• User may request specific changes → apply changes, then go to Phase 5
• User may reject → end, suggest re-running skill-1

Phase 5: Output — Write idea01.md

After user approval, write the comprehensive idea document to output/ideas/idea01.md:

# Research Idea: [Selected Title]

**Candidate Paper Titles** (user selects final title during paper writing):
1. [Title option 1]
2. [Title option 2]
3. [Title option 3]

**Generated**: [date]
**Validated**: [date]
**Status**: Validated — passed deep search + GPT-4.5 review
**Target Journal**: Advanced Engineering Informatics (or similar Q1)
**Innovation Type**: [Application / Method / System]

## 1. Research Background
[Why this problem matters, what gap exists, 2-3 paragraphs]

## 2. Core Research Question
[Clear statement of the main research question / hypothesis]

## 3. Innovation Points
- **Innovation 1**: [description]
- **Innovation 2**: [description]
- **Innovation 3**: [description]
[Why each is novel, with references to search results]

## 4. Proposed Methodology Overview
### 4.1 Data Collection Strategy
- Data sources, types, expected volume

### 4.2 Ontology Design Direction
- Domain concepts, key relationships, design approach

### 4.3 Knowledge Graph Construction
- Extraction method, graph database, expected scale

### 4.4 Graph RAG Strategy
- Retrieval approach, LLM integration, query handling

### 4.5 Chatbot / Q&A System
- Interface, interaction flow, target users

### 4.6 Evaluation Plan
- Metrics, baselines, ablation studies

## 5. Expected Results & Contributions
- Theoretical, practical contributions, expected findings

## 6. Data Requirements
| Data Type | Source | Format | Volume | Purpose |
|-----------|--------|--------|--------|---------|

## 7. Technical Feasibility
- Tools/APIs, compute, timeline, risks

## 8. Literature Review Framework
[Topics to cover, for skill-5]

## 9. Significance & Impact
[Who benefits, broader impact]

## 10. Validation Record
### Deep Search Results
| # | Paper/Source | Year | Venue | Overlap | Differentiation |
|---|-------------|------|-------|---------|-----------------|

### GPT-4.5 Review
- Scores: Novelty [X], Feasibility [X], Significance [X], Methodology [X], Data [X]
- Verdict: [ACCEPT/REVISE/REJECT]
- Key feedback: [summary]

### Modifications During Validation
| Issue | Source | Change Made | User Approved |
|-------|--------|-------------|---------------|

## 11. Idea Development Log
| Date | Event | Detail |
|------|-------|--------|

Key Rules

1. No modification without user consent: Report issues, wait for approval, then change.
2. No review loops: Claude searches once, GPT-4.5 reviews once. Combine results, present to user, finalize.
3. GPT-4.5 is a guard, not a generator: It verifies Claude's idea, nothing more.
4. This skill writes the file: idea01.md is ONLY produced here, not in skill-1.
5. Honest reporting: If the idea has fatal flaws, say so clearly.

Composing with Other Skills

/initial-idea-generation  → user-approved idea (no file)
/idea-checking            → search + GPT-4.5 review → idea01.md  ← YOU ARE HERE
/ontology-design          → domain ontology from idea01.md
/research-planning        → plan from idea01.md + ontology.md