Skip to content

SC-Airu/groundwork-skill

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

15 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

🔭 Groundwork

Research the problem space before you build.
A landscape scan skill for Claude Code.

MIT License Claude Code Skill OMC Enhanced

한국어


Don't build blind. Groundwork runs 4 parallel research agents to scan the landscape — who has this problem, how they work around it, and what solutions exist — so you can make informed decisions before writing a single line of code.

Give your AI real-world context and it gives you smarter answers.

Quick Start

# Install
mkdir -p ~/.claude/skills/groundwork && curl -sL https://raw.githubusercontent.com/SC-Airu/groundwork-skill/main/SKILL.md -o ~/.claude/skills/groundwork/SKILL.md

# Use in Claude Code
/groundwork AI ad creative analysis for mobile games

What It Does

Give it a pain point. Get back 3 structured research files in ~2 minutes:

.omc/groundwork/{slug}/
├── triage.md      # Problem / Who / Why
├── context.md     # Workflow, affected roles, workarounds, adjacent problems, user voices
└── solutions.md   # Solution list, categories, frequency ranking, gaps, key insight
  • 4 parallel research agents — Context, Solutions, Behavior, JTBD run simultaneously (~2-3 min)
  • Gap analysis — Finds what no existing tool covers
  • Contradiction detection — Catches "marketed as X" vs "users say Y" discrepancies
  • Duplicate check — Won't overwrite existing research without asking
  • Facts only — No build/kill recommendations. You decide.
  • English search, localized output — Searches in English for broad coverage, saves in Korean (configurable)

How It Works

flowchart TB
    Input["🎯 Pain Point Input"] --> Triage

    subgraph Triage["Step 0: Triage"]
        direction LR
        Check{"Existing<br/>results?"}
        Parse["Parse What / Who / Why"]
        Ask["Ask max 2 questions<br/>(if unclear)"]
        Check -->|No| Parse
        Check -->|Yes| Confirm["Ask: overwrite<br/>or keep?"]
        Parse --> Ask
    end

    Triage --> Explore

    subgraph Explore["Step 1: Explore (parallel)"]
        direction LR
        A["🔍 Context<br/>workflow, workarounds<br/>user voices"]
        B["📦 Solutions<br/>keyword + curated<br/>lists merged"]
        C["👥 Behavior<br/>what people<br/>actually use"]
        D["🔀 JTBD<br/>alternative<br/>approaches"]
    end

    Explore --> Gap["🧩 Gap Analysis<br/>dedup + contradiction check"]
    Gap --> Save["💾 Save 3 files (Korean)"]
    Save --> Summary["📋 Summary to user"]
Loading

Research Agents

Each agent has a distinct search strategy and source set:

Agent Role Sources Strategy Limit
Context Workflow, workarounds, user voices Reddit, HN, forums, Stack Overflow Community-focused: finds direct quotes and real frustrations 10 searches
Solutions Existing tools and products GitHub, Product Hunt, G2, Capterra, AlternativeTo, "awesome-*" lists, "best X alternatives" articles Keyword + curated lists merged (70% overlap in testing when separate) 10 searches
Behavior What people actually use Reddit, forums, Stack Overflow, HN Searches "how do you handle..." and "what do you use for..." discussions — separates marketed claims from real usage 10 searches
JTBD Alternative approaches Cross-industry, cross-domain Jobs-to-be-Done lens: finds non-obvious competitors solving the same job differently 8 searches

All agents search in English regardless of input language, for maximum coverage. After all 4 complete, the orchestrator runs gap analysis inline: deduplicates solutions, cross-references workarounds, identifies structural gaps, and flags contradictions.

Example Output

Below is a real example from scanning "AI ad creative analysis for mobile games".

Note: Default output is in Korean. Shown here in English for readability. Output language is configurable.

triage.md — Problem definition
# Triage
- Problem: AI-powered analysis of mobile game ad creatives (video/image)
          to identify what drives performance
- Who: UA/marketing teams, creative teams, executives/PMs
- Why: Data-driven ad creative analysis is essential for efficient UA operations
       and creative optimization
context.md — Workflow & user voices
# Context: AI Ad Analyst

## Current Workflow
Bottleneck recurs across these UA workflows:
1. Post-campaign review — aggregate data from Meta/AppLovin/Google/TikTok via MMP
2. Creative fatigue detection — creative lifespan 5-10 days on high-spend channels
3. Creative briefing — derive new concepts from past performance (relies on memory)
4. A/B testing — platform algorithms cause uneven exposure within ad sets

## Who Is Affected
| Role | Responsibility | Skill Level |
|------|---------------|-------------|
| UA Manager | Campaign budgets, ROAS targets | Mid-advanced (data analysis required) |
| Creative Strategist | Translate performance data → creative briefs | Hybrid (data + creative) |
| Creative/Art Team | Mass asset production (46.2M assets in 2024) | Design/video expert, weak on data |
| Executives/PM | Creative ROI reporting, budget approval | Summary data consumer |

## Current Workarounds
1. Manual naming conventions + spreadsheets — ~20 hrs/week per team
2. Cross-platform manual switching between Meta/AppLovin/TikTok/MMP dashboards
3. Competitor spy tools — long-running creatives assumed profitable
4. Volume strategy — produce 20-40 new creatives/month, let algorithm find winners

## User Voices
> "Performance data is abundant, but actionable creative insight is scarce."
> — Segwise, 2026

> "Managing naming conventions was heinous."
> — Foxwell Digital
solutions.md — Solution landscape (key sections)
# Solution Landscape: AI Ad Analyst

## Solution List
| Name | Approach | Strengths | Weaknesses |
|------|----------|-----------|------------|
| Segwise | Multimodal AI tagging → IPM/CTR/ROAS | Mobile game native, playable ads | Newer entrant |
| AppsFlyer Creative Opt. | AI scene decomposition + MMP attribution | Direct attribution link | Requires AppsFlyer as MMP |
| VidMob | AI + human expert scoring, cross-channel | Industry incumbent, 300+ brands | Enterprise-only, expensive |
| Replai | Computer vision, frame-by-frame video tagging | $5B+ ad spend processed | Video-only, no playable ads |
| SensorTower | Competitor creative gallery, 14+ networks | Market leader | $25K+/yr, no own-creative AI |
| Motion | Visual-first creative analytics | Intuitive UI, $299-599/mo | Not game-specific |
| ... | (22 solutions total, 6 categories) | | |

## Categories
1. AI Creative Analysis (own creatives) — Segwise, AppsFlyer, VidMob, Replai ...
2. Competitor Intelligence (ad spy) — SensorTower, MobileAction, AppMagic ...
3. AI Creative Generation — Pencil, AdCreative.ai, Predis.ai
4. Automated Multivariate Testing — Marpipe, Smartly.io
5. Pre-Launch Consumer Testing — System1, Kantar, Behavio
6. Neuroscience/Attention — Realeyes, Dragonfly AI

## Key Gap
No tool solves the full pipeline:
| Gap | Description |
|-----|-------------|
| Element → Retention/LTV | No tool connects creative attributes to D1/D7/D30 retention |
| Playable Ad Analysis | Only Segwise attempts AI analysis of interactive ads |
| Integrated Workflow | Competitor intel + own analysis + briefing requires 3-5 separate tools |
| Game-specific + Affordable | Game-native tools are enterprise-priced; affordable tools are e-commerce focused |

## Contradictions
| Contradiction | Marketing | Reality |
|--------------|-----------|---------|
| AI creative analysis adoption | "Many tools available" | Most teams still use spreadsheets + naming conventions |
| VidMob accessibility | Most-cited incumbent | Enterprise-only, inaccessible to small/mid studios |
| Volume strategy efficiency | "Produce more to find winners" | 2% of creatives get 68% of spend — 98% structural waste |

## Key Insight
The market is in a "data-rich, insight-poor" structural paradox.
Most UA teams still rely on manual tagging + spreadsheets. AI creative
intelligence tools (Segwise, Replai, Reforged Labs) are emerging but
in early adoption. The biggest unsolved problems: creative element →
downstream metrics (retention/LTV), playable ad analysis, and an
integrated workflow from analysis → briefing → production.

Terminal Summary

After research completes, you get a brief summary:

## Groundwork 완료: ai-ad-analyst

### 컨텍스트
- 모바일 게임 UA 팀은 광고 소재의 성과 데이터는 풍부하지만, "왜 잘됐는지" 요소 수준
  인사이트가 부족해 수동 태깅+스프레드시트에 의존하는 구조적 병목 존재
- 주요 워크어라운드: 네이밍 컨벤션 + 피벗 테이블 (팀당 주 ~20시간), 물량 전략(월 20~40개 소재)

### 솔루션 현황
- 22개 솔루션, 6개 카테고리 (AI 자사 분석 / 경쟁사 인텔리전스 / AI 생성 / 다변량 테스트
  / 프리론칭 테스트 / 뉴로사이언스)
- 핵심 인사이트: AI 크리에이티브 인텔리전스 도구(Segwise, Replai, Reforged Labs)가 부상
  중이나 아직 초기 채택 단계. 대부분의 팀은 여전히 스프레드시트 기반
- 핵심 공백: ①소재 요소→리텐션/LTV 연결 ②플레이어블 광고 AI 분석 ③경쟁사 인텔리전스
  +자사 분석+브리핑을 잇는 통합 워크플로 ④중소 게임 스튜디오를 위한 게임 특화+합리적 가격 조합

### 파일
- .omc/groundwork/ai-ad-analyst/triage.md
- .omc/groundwork/ai-ad-analyst/context.md
- .omc/groundwork/ai-ad-analyst/solutions.md

Usage

# Korean input
/groundwork 게임 사운드 자동 배치 - AI가 영상 분석해 효과음 자동 삽입

# English input
/groundwork auto SFX placement for game ad videos in After Effects

# Detailed input (skips triage questions)
/groundwork Music Prompt Builder - a tool that generates Suno AI BGM prompts
  through simple clicks. Planners select game background, style, mood, tempo,
  instruments and get translated professional music terminology prompts.

Customization

Change output language

Edit the <Execution_Policy> section in SKILL.md:

- All saved files: written in Korean.

Change to your preferred language. Research agents always search in English.

Adjust search depth

Each agent has a Limit to N web searches max instruction. Defaults: 10 for most agents, 8 for JTBD.

  • Increase for deeper research
  • Decrease for speed
Use with downstream skills

Groundwork output is designed to feed into other skills:

Skill How
/plan Reads triage.md for problem context
CLAUDE.md Reference groundwork files for team context

Requirements

With vs Without OMC

With OMC Without OMC
Agent document-specialist (research-optimized prompt) general-purpose (automatic fallback)
Web search Yes Yes
Research quality Higher — specialist prompt tuned for landscape scans Good — same web search tools, no specialist framing
Speed ~2-3 min ~2-3 min

Groundwork works without OMC by falling back to Claude Code's built-in general-purpose agent. Install OMC for better research quality via the document-specialist agent's optimized system prompt.

Design Decisions

Decision Why
4 agents, not 6 Keyword + Curated merged (70% overlap in testing). Behavior kept separate — finds what people use vs what's marketed.
No Gap Check agent Orchestrator handles dedup + contradiction inline. No quality loss in testing.
English search Broader coverage than localized search. Output language is separate.
No depth modes Single mode. 4 agents is the sweet spot between speed and coverage.

Contributing

Issues and PRs welcome. This is a single-file skill (SKILL.md) — keep changes focused.

License

MIT

About

Lay the groundwork before you build. A lightweight research skill for Claude Code that scans the problem space — context, solutions, and gaps — in ~2 minutes.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors