initial-idea-generation

name	initial-idea-generation
description	Generate and validate Graph RAG research ideas for risk/disaster/threat/event domains. Use when user says "找idea", "generate idea", "我有一个想法", "brainstorm", or provides a raw research concept.
argument-hint	["user-idea-or-research-direction"]
allowed-tools	Bash(*), Read, Write, Grep, Glob, WebSearch, WebFetch, Agent

Skill-1: Initial Idea Generation

Generate, validate, and refine Graph RAG research ideas for: $ARGUMENTS

Overview

This skill takes the user's raw idea or research direction, validates whether it aligns with Graph RAG implementation, then generates 3 candidate ideas with full detail. The goal is to produce ideas publishable in Q1 journals such as Advanced Engineering Informatics.

This skill composes with /idea-checking (skill-2) for deep novelty validation and cross-LLM review (via Codex MCP + GPT-4.5). Its output feeds into /ontology-design (skill-4) and /research-planning (skill-3).

Workflow

Phase 1: Graph RAG Alignment Check

Evaluate whether the user's idea can be implemented through Graph RAG technology.

Acceptance criteria (flexible, not all required):

• The problem involves structured/semi-structured knowledge that benefits from graph representation
• There are identifiable entities and relationships (e.g., risks→causes, disasters→responses, threats→mitigations)
• A knowledge graph can meaningfully enhance retrieval compared to flat document search
• The domain has sufficient data sources (papers, news, reports) for KG construction
• The scenario involves Q&A, decision support, or information retrieval where graph context adds value

If the idea FITS Graph RAG:

• Confirm alignment to the user with a brief explanation of why it works
• Proceed to Phase 2

If the idea DOES NOT FIT Graph RAG:

• Explain specifically why it doesn't align (e.g., "This problem is better suited to traditional NLP classification rather than graph-based retrieval")
• Suggest how the idea could be modified to fit the Graph RAG direction
• Do NOT proceed until the user explicitly agrees to either:
  • Modify their idea to align with Graph RAG, OR
  • Confirm they want to proceed with the original idea anyway (with documented risks)

Phase 2: Landscape Quick Survey

Before generating ideas, understand what exists in the area.

1. Search recent literature using WebSearch:

   • Query the user's topic + "knowledge graph" + "graph RAG" in multiple formulations
   • Check recent papers (last 2-3 years) from venues relevant to the domain
   • Focus on Advanced Engineering Informatics, Expert Systems with Applications, Knowledge-Based Systems, Information Sciences, and similar Q1 journals
   • Read abstracts of 8-12 relevant papers

2. Map the landscape:

   • What Graph RAG / KG approaches exist in this domain?
   • What data sources are commonly used?
   • What ontologies have been proposed?
   • What are the stated limitations and future work directions?

3. Identify gaps:

   • Domains where Graph RAG hasn't been applied yet
   • Ontology designs that are incomplete or domain-specific limitations
   • Retrieval strategies that haven't been explored
   • Multi-source data integration opportunities (news + papers + reports)

Phase 3: Generate 3 Candidate Ideas

Based on the landscape survey and user's direction, Claude generates exactly 3 ideas with distinct innovation angles.

The 3 ideas MUST cover different innovation types:

Idea A — Application/Scenario Innovation: Apply Graph RAG to a novel domain or scenario that hasn't been explored. Focus on new use cases, novel data sources, or unique problem formulations.

Idea B — Method/Technical Innovation: Propose a new or improved method for ontology design, knowledge extraction, graph construction, or retrieval strategy. Could involve novel graph structures, hybrid retrieval, or multi-hop reasoning.

Idea C — System/Integration Innovation: Design a novel system architecture that integrates multiple components in a new way — e.g., multi-source data fusion, real-time KG updates, cross-domain transfer, or human-in-the-loop design.

For each idea, provide:

1. 2-3 candidate paper titles (concise, publication-ready, suitable for Q1 journal)
2. One-paragraph summary
3. Core research question / hypothesis
4. Innovation points (what is new compared to existing work)
5. Data sources needed (papers, news, reports — be specific)
6. Ontology scope (what concepts and relations would the KG cover)
7. Graph RAG strategy (how the graph enhances retrieval)
8. Chatbot/system design (what the end-user interaction looks like)
9. Expected contribution to the field
10. Feasibility assessment (can this be done with LLM APIs + web data + standard tools?)
11. Risk level: LOW / MEDIUM / HIGH

Important: All 3 ideas must be ones that Claude can realistically help execute through the downstream skills and subagents in this project. Do not propose ideas requiring resources beyond what is available.

Phase 4: Q1 Journal Evaluation

For each of the 3 ideas, evaluate against Advanced Engineering Informatics standards:

Criterion	Weight	Evaluation
Novelty	High	Is this substantially different from existing work?
Technical depth	High	Does the methodology show sufficient engineering/informatics rigor?
Practical significance	High	Does it solve a real-world problem in risk/disaster/event management?
Reproducibility	Medium	Can the experiments be replicated with described data and methods?
Completeness	Medium	Does it cover the full pipeline (data → KG → RAG → evaluation)?
Writing potential	Medium	Can this fill a full journal paper with substantial content?

Rate each idea: Strong Accept / Accept / Borderline / Reject for Q1 potential.

Phase 5: Present to User & Interactive Refinement

Present the 3 ideas to the user in a structured format:

## Idea A: Application Innovation
**Candidate Titles:**
1. [Title option 1]
2. [Title option 2]
3. [Title option 3]
[Summary, innovation, feasibility, Q1 evaluation]

## Idea B: Method Innovation
**Candidate Titles:**
1. [Title option 1]
2. [Title option 2]
[Summary, innovation, feasibility, Q1 evaluation]

## Idea C: System Innovation
**Candidate Titles:**
1. [Title option 1]
2. [Title option 2]
[Summary, innovation, feasibility, Q1 evaluation]

## Recommendation
[Which idea I recommend and why]

Wait for user selection. Do NOT proceed without explicit user choice.

After user selects an idea, enter free-form dialogue to refine:

• Answer user's questions about the chosen idea
• Deepen specific aspects (methodology, data, evaluation design)
• Adjust scope based on user feedback
• Ensure the user is satisfied with all aspects before finalizing

Critical rule: Every refinement must be confirmed by the user. Never auto-advance.

Phase 6: User Approval — Handoff to Skill-2

Once the user confirms the final idea through dialogue:

1. Summarize the approved idea clearly in conversation (title, innovation points, method outline, data needs)
2. Do NOT write idea01.md yet — file output is handled by skill-2 (/idea-checking) after cross-LLM validation
3. Inform the user that the next step is /idea-checking to validate novelty and feasibility via deep search + Codex MCP (GPT-4.5)
4. Wait for user to invoke /idea-checking — do not auto-trigger

Key Rules

1. User permission is paramount: Never generate ideas without user approval to proceed. Never auto-select an idea.
2. Graph RAG alignment: Every idea must be implementable through ontology → KG → Graph RAG → chatbot pipeline.
3. Q1 standard: Ideas must be assessed against real Q1 journal criteria, not inflated.
4. No file output: skill-1 does NOT write idea01.md. It produces a user-approved idea in dialogue. The file is written by skill-2 after validation.
5. Honesty over enthusiasm: If an idea is weak, say so. Better to refine now than fail at review.
6. Development log: Record every decision and its reason in the idea file.
7. Free dialogue: After presenting 3 ideas, engage naturally with the user — answer questions, take suggestions, iterate until they are satisfied.
8. No Codex MCP here: Idea generation is done by Claude. Codex MCP validation happens in skill-2 (/idea-checking).

Composing with Other Skills

/initial-idea-generation "direction"  → 3 ideas → user selects → user-approved idea (no file yet)
/idea-checking                        → deep novelty + Codex MCP (GPT-4.5) validation → outputs idea01.md
/ontology-design                      → domain ontology from idea
/research-planning                    → executable research plan