aihub-app-ideation

name	aihub-app-ideation
description	Generate detailed, solo-developer-focused app/web product ideas from AI-Hub Korea dataset metadata. Use when the user wants to brainstorm app ideas from a dataset, asks "이 데이터셋으로 뭘 만들 수 있어 / 앱 아이디어 뽑아줘", "what can I build with dataset X", or wants to judge feasibility and rank ideas. Always produces at least 5 ideas per dataset with a rich per-idea spec, metadata-only.

AI-Hub App Ideation Skill

Purpose

Turn AI-Hub Korea dataset metadata into concrete, buildable app/web product ideas — at least 5 per dataset — each with a detailed spec and an honest feasibility judgement. This skill is the deep, structured version of the ideate CLI command in this repo.

Use it when the user wants to explore "what could I build with this dataset?"

Hard Requirements

1. Minimum 5 ideas per dataset. Never return fewer. If the metadata is thin, still produce 5 and mark the weak ones no-go with the reason.
2. Metadata only. Reason strictly from title, inferred tags, file tree, file sizes, train/val split, label/source presence. Never claim dataset contents, label schemas, licenses, or quality you cannot see. State that every judgement is a metadata-only estimate.
3. Each idea uses the full per-idea template below. "More detailed than a one-liner" is the bar — no bare titles.
4. Honest feasibility. Score and verdict must reflect reality, including the solo-developer lens below. Do not inflate.
5. Always scan competitor apps. Every idea must include a real competitor check — never assume a niche is empty. Search the App Store / Google Play and the web for existing apps (Korean AND global/English), name 2–4 concrete ones with what they do and where they fall short, and let that gap drive the 차별화 무기. "No competitor found" is a valid result only after an actual search, and is itself a warning (small/non-existent market), not a green light.

Default Lens: Solo Developer Building Apps

Unless the user says otherwise, assume the builder is one person shipping B2C apps (mobile/web, self-serve distribution). Score ideas through this lens:

• Favor: image classification / text tasks that run cheaply or on-device; clear consumer hook; simple monetization (subscription / IAP); Korea-specific data as a moat against global apps.
• Down-rank (usually unrealistic for a solo app builder): anything needing B2B/B2G sales, on-prem hardware/sensor/factory/field access, or heavy regulation/liability (medical diagnosis, face recognition, financial advice). Examples to treat as low-priority for this persona: 전력설비, 공공기관 납품, 특허/지식재산 전문툴, 제조 불량검출 라인, 위성/교통 인프라.
• If an idea is attractive but unrealistic for a solo builder, say so explicitly rather than silently ranking it high.

When the user states a different persona (enterprise, team, research), adapt the lens accordingly and say which lens you used.

Workflow

1. Acquire dataset metadata

Prefer the repo's own pipeline so reasoning is grounded in real parsed metadata:

./run.sh inspect --datasetkey <KEY>     # parses + normalizes, writes summary JSON

Or reuse an existing summary at data/normalized/datasets/<KEY>.json. If neither is available (no aihubshell), work from whatever metadata the user provides, and note that keys/sizes are unverified.

Read from the summary: title, category_guess, modality_guess, tags, file_count, human_size, sample_file_paths, metadata_lines, parse_status, train/val and label/source presence.

2. Classify the dataset

State in one or two lines: domain, modality (or "unknown — verify by opening one source zip"), rough size class, and supervised-readiness (labels + train/val). Flag uncertainty honestly (e.g. "통합 데이터" → modality could be images or sensors).

3. Generate at least 5 ideas

Spread ideas across different angles — don't give five variants of one app. Cover a mix of: consumer utility, creator/content, education/learning, wellness/care, and a "data-as-moat" or niche angle. Each idea uses the template.

4. Competitor / market scan (required)

For each idea, do a real search before scoring — do not reason from memory alone:

• Search app stores and the web for existing apps, both Korean (점신, 포스텔러, 네이버/카카오 등) and global/English. Useful queries: " app", " 앱", " App Store / Google Play".
• If the user has a market-scouting repo (e.g. finding-cash-cow-android — Play Store scraping/analysis), reuse its data when it's in session scope.
• Record 2–4 concrete competitors per idea: name, platform, what they do, price/ monetization if visible, and the gap they leave. Note incumbents that make a head-on solo entry unrealistic (e.g. a category leader with huge revenue).
• Turn the gap into the idea's 차별화 무기, and let competitive density feed the opportunity_score (crowded + strong incumbents → lower; real unmet gap → higher).
• An honestly empty niche is a yellow flag (likely small market), not a win.

5. Score and judge each idea

Give three 1–10 scores and a verdict:

• opportunity_score — market pull / consumer demand
• feasibility_score — how realistically a solo dev ships it (infra, cost, skill)
• data_fit_score — how well the dataset actually supports the idea
• feasibility: go | maybe | no-go

6. Rank and recommend

Rank the ideas (a simple combined weighting: opportunity 0.45, feasibility 0.35, data-fit 0.20 — same as the repo's combined_score). Recommend the top 1–2 for a solo builder and say why, plus the single biggest risk for each.

Per-Idea Template (use all fields)

### <n>. <앱 이름> — <go|maybe|no-go>
- 한 줄 소개: <one-line pitch>
- 형태: app | web | both
- 타겟 사용자: <who, specifically>
- 핵심 기능: <3–5 features>
- 데이터 활용: 이 데이터셋의 무엇(라벨/모달리티/구조)을 어떤 모델 태스크로 쓰는지
- 모델 · 추론 비용: 온디바이스 가능 여부 / 학습 필요 여부 / 솔로 예산에서 현실적인지
- 솔로 실현성: 혼자 출시 가능한 이유 또는 막히는 지점
- MVP 흐름: 화면/단계 3–5개
- 수익 모델: 구독 / IAP / 광고 / 제휴
- 경쟁 앱: 실제 검색으로 찾은 기존 앱 2–4개 (국내+글로벌) — 이름·플랫폼·하는 일·빈틈. 없으면 "검색했으나 없음(=시장 작을 위험)"
- 차별화 무기: 위 경쟁 앱이 못 채운 빈틈을 어떻게 공략하는가 (특히 한국 특화 우위)
- 리스크: 2–3개 + 회피책
- 점수: 기회 N/10 · 실현 N/10 · 데이터적합 N/10

Multiple datasets

When given several datasets, ideate each (≥5 ideas each), then build a cross-dataset ranking: which dataset offers the strongest solo-app opportunity, using each dataset's best idea. The repo command for the aggregate view:

./run.sh ideate-rank

Using the automated pipeline

The repo can generate and persist this automatically (LM Studio / OpenAI / Anthropic providers):

./run.sh ideate --datasetkey <KEY> --ideas 5            # local LM Studio (default)
./run.sh ideate --datasetkey <KEY> --provider anthropic # Claude
./run.sh ideate-rank                                    # rank datasets

Artifacts: data/normalized/ideation/<KEY>.json and data/generated/ideation/<KEY>-<slug>.md. When you run ideation manually (no LLM key), follow the same schema and template so output stays consistent with the pipeline. The pipeline prompt also enforces the ≥5-idea and solo-lens rules.

Cautions

• License, approval scope, and policy fit are out of scope for metadata — always flag them as "verify separately".
• Mental-health / medical / safety ideas: position as journaling/logging/wellness, never diagnosis; add a crisis-resource / disclaimer note.
• Keep scores honest; a list where everything is go is a failed analysis.

Before Finishing

• Confirm every dataset got ≥5 ideas, each with the full template.
• Confirm each idea has a competitor scan from a real search (≥2 named apps or an explicit "searched, none found" note).
• Confirm the solo-app lens was applied (or the chosen lens was stated).
• End with a ranked recommendation and the top risk per recommended idea.