[🔨활용] 그래프 창의력 증진기 스킬md임 쓸사람쓰셈

[믕붕이(prior3567)]

간단히 말해서

사용자가 목표 질문만 주면, 관련 검색 쿼리 3~5개를 생성합니다.

검색 결과를 출처, 요약, 특징, 한계가 포함된 근거 단위로 정리합니다.

각 근거에서 구체적인 특징을 뽑아 노드로 만들고, 같은 근거에서 함께 등장한 특징을 엣지로 연결합니다.

그래프에서 연결 부족, 구분 부족, 모순 가능성, 중심적이지만 근거가 빈 연결, 검색 부족, 출처 편향을 찾습니다.

그 결과를 바탕으로 정보 가치가 큰 다음 질문 또는 탐구 주제 3~5개를 우선순위와 함께 제시합니다.

각 추천에는 중요성, 근거, 다음 행동, 그래프 개선 예상, Curiosity Score를 포함합니다.

사람이 검증한 내용은 Verified Graph, AI가 제안한 연결과 가설은 Candidate Graph로 분리합니다.

ㅡㅡㅡㅡㅡㅡ

---

name: graph-curiosity-finder

deion: Use when the user provides a goal question or raw materials and wants the system to search, structure evidence, build a feature graph, find information gaps, and recommend what to investigate next. This skill can start from only a goal question by generating search queries before graph analysis.

---

# Graph Curiosity Finder

## Purpose

Graph Curiosity Finder is not a generic summarizer. It can start from either a user's goal question or supplied materials, gather or structure evidence as a feature graph, and find:

- missing or weak relationships

- areas where labels or cases are hard to distinguish

- possible contradictions

- high-value questions that would improve the knowledge structure

- candidate links the user may approve later

The final goal is to tell the user what to investigate next, even when the user did not provide source documents.

## Operating Modes

Prefer Goal Question Mode when the user gives a question, problem, topic, hypothesis, or decision they want to explore.

Use Direct Material Mode when the user has already provided documents, notes, reports, interview excerpts, paper summaries, CSV extracts, or case data.

## Goal Question Mode

When the user only provides a goal question, work in this order:

1. Restate the goal question in one sentence.

2. Generate 3 to 5 search queries.

3. Run web search or available local/source search for those queries.

4. Select diverse, relevant, and reasonably authoritative results.

5. Summarize each result into evidence units.

6. Extract concrete features from the evidence units.

7. Turn features into nodes and co-occurrences into edges.

8. Analyze the graph for connection gaps, contradiction possibilities, distinction gaps, and important missing links.

9. Recommend 3 to 5 next questions or exploration topics.

10. For each recommendation, explain:

    - why it matters

    - what evidence or graph structure produced it

    - what the user should do next

If web access is unavailable, say so briefly and continue with query generation plus a research plan. Do not pretend search results were retrieved.

## Search Query Generation

Generate queries that cover different angles of the goal question:

- overview query: broad context and terminology

- mechanism query: causes, drivers, or relationships

- evidence query: data, studies, cases, benchmarks, or reports

- contrast query: competing explanations, failure cases, counterexamples, or risks

- action query: methods, experiments, evaluation, or implementation guidance

Use 3 queries for narrow questions and 5 queries for broad or ambiguous questions.

Queries should be specific enough to retrieve evidence, not just restate the user's wording.

## Search Result Structuring

For each useful result, create an evidence unit:

- source_id

- title

- url or locator when available

- source_type

- short_summary

- extracted_features

- uncertainty_or_limitations

Favor primary or authoritative sources when the topic affects money, health, law, safety, science, or current events. If using secondary sources, label them as such.

## Direct Material Mode

When the user provides source material directly, work in this order:

1. Read the user's input material.

2. Extract concrete, reusable features.

3. Merge duplicate or very similar features into nodes.

4. Create edges between features that appear in the same document, source, case, or example.

5. Treat user-approved information as Graph A, the verified graph.

6. Treat AI-generated questions, inferred gaps, and proposed new links as Graph B, the candidate graph.

7. Analyze edge structure before generating questions.

8. Generate 3 to 10 prioritized questions.

9. Explain the reason for each question in easy Korean unless the user asks for another language.

## Data Model

Node:

- id

- name

- source_ids

- labels

- deion

Edge:

- source_node_id

- target_node_id

- source_ids

- labels

- evidence_text

- evidence_source

- search_query, when the evidence came from search

Do not add AI-generated relationships directly to Graph A. Keep them in Graph B until the user approves them.

## Feature Extraction Rules

Good features are specific and reusable:

- "20대 여성 사용자 반응률 높음"

- "모바일 결제 단계에서 이탈률 증가"

- "고온 조건에서 성능 저하"

- "신규 기기에서 고액 현금서비스 이용"

Avoid vague features:

- "좋다"

- "중요하다"

- "문제 있음"

- "뭔가 이상함"

## Gap Types

Look for these four gap types:

1. 연결 부족 영역: related nodes or clusters that should probably connect but lack direct evidence.

2. 구분 부족 영역: different labels or outcomes share similar features and edges, making them hard to distinguish.

3. 모순 가능 영역: similar feature combinations point to different outcomes.

4. 중요하지만 비어 있는 연결: central features or bridge candidates lack evidence for key relationships.

For Goal Question Mode, also track:

5. 검색 부족 영역: the graph suggests an important subtopic, but search results did not provide enough evidence.

6. 출처 불균형 영역: one side of a debate or one type of evidence dominates the retrieved material.

## Question Scoring

Assign each question a Curiosity Score from 0 to 100 using these factors:

- graph connection improvement

- label distinction improvement

- contradiction resolution potential

- reusability across cases

- practical feasibility

- expected information gain from the next search, interview, experiment, or analysis

Scores should be comparative, not decorative. Put the highest-value questions first.

## Output Format

Use this structure:

1. 목표 질문

2. 생성한 검색 쿼리

3. 사용한 주요 자료 요약

4. 현재 그래프 요약

5. 발견된 정보 부족 영역

6. 추천 질문 또는 탐구 주제

7. 우선순위

8. 사용자가 승인할 수 있는 후보 연결 목록

For each recommended question, include:

- 질문

- 왜 이 질문이 중요한가

- 어떤 근거에서 나왔는가

- 다음에 뭘 하면 되는가

- 예상되는 그래프 개선

- Curiosity Score

Only show detailed edge lists when the user asks for them or when they are necessary to explain the recommendation. Always provide links or source names for searched evidence when available.

## Recommendation Requirements

Recommend 3 to 5 items in Goal Question Mode.

Each item must be concrete enough to act on. Prefer:

- "어떤 데이터를 찾아야 하는가"

- "어떤 비교를 해야 하는가"

- "어떤 사례를 더 봐야 하는가"

- "어떤 실험이나 인터뷰 질문이 필요한가"

Avoid generic recommendations such as "더 조사하세요" or "시장 조사를 하세요".

## Optional Local MVP

This plugin includes two small local helper s:

- `s/goal_question_plan.py`: turns a goal question into 3 to 5 search queries and an evidence/graph plan.

- `s/graph_curiosity_mvp.py`: turns saved text or saved search summaries into an initial graph and question seeds.

Use it when the user wants a quick deterministic first pass over plain text files or saved search summaries. The does not replace deeper reasoning; it gives an initial graph, candidate gaps, and question seeds that you should refine for the user.

Goal question planning example:

[믕붕이(prior3567)]

활용예
https://m.dcinside.com/board/thesingularity/1310812

[GPT의가지볶음(decay4934)]

쪼아요

[믕붕이(prior3567)]

사람들 욕할까봐 심장 두근거렸는데 다행이다...

[익명(maple0696)]

@믕붕이 정보공유에 욕하는 게 이상한 애들이니까 겁목을 필요 없수다

[익명(mediate2590)]

[믕붕이(prior3567)]

[내일특이점옴(destroy4259)]

[믕붕이(prior3567)]

[포라이(macdoongs)]

[믕붕이(prior3567)]

[익명(maple0696)]

요거 토큰 많이 먹지? 페이블로 돌리긴 좀 에바인가?

[믕붕이(prior3567)]

어음...난 지피티만 써서 잘 모르겠음...5.5 기준 그렇게 많이 닳지는 않았음(주관적)

[익명(maple0696)]

@믕붕이 ㄱㅅ링