gpt-research

name	gpt-research
description	Extracts project context into a structured research prompt for GPT-PRO and copies it to the clipboard. Use for requests that delegate research to GPT or extract project context as a structured prompt, such as "GPT에 물어봐", "GPT 리서치", "GPT-PRO", "리서치 위임", "컨텍스트 추출", "gpt research", "GPT에게 넘겨", "GPT한테 질문", "GPT 프롬프트", "research prompt", "컨텍스트 뽑아줘", "GPT용 프롬프트", "외부 리서치", "프롬프트 생성", "GPT 위임".
argument-hint	[module|arch|issue|custom] [target path or description]

user: "이 모듈에 대해 GPT에 물어보게 컨텍스트 뽑아줘 src/auth/" assistant: "module 모드로 src/auth/ 디렉토리의 소스, 의존성, 인터페이스를 추출하여 GPT-PRO 리서치 프롬프트를 생성하겠습니다." user: "프로젝트 아키텍처를 GPT-PRO한테 분석시키고 싶어" assistant: "arch 모드로 프로젝트 전체 구조, 의존성 그래프, 빌드 시스템을 요약하여 리서치 프롬프트를 클립보드에 복사하겠습니다." user: "이 에러 GPT한테 넘겨서 원인 분석 받자: TypeError: Cannot read properties of undefined" assistant: "issue 모드로 에러 관련 소스, 콜체인, git 히스토리를 추출하여 GPT-PRO 리서치 프롬프트를 구성하겠습니다." user: "GPT 리서치용 프롬프트 만들어줘, 내가 범위 지정할게" assistant: "custom 모드로 진행합니다. 어떤 파일이나 주제를 포함할지 알려주세요."

GPT-PRO Research Prompt Generator

Extract a specific part of the project (module, architecture, or issue) into a structured research prompt to delegate to GPT-PRO, and copy it to the clipboard (pbcopy).

Respond to the user in Korean.

The Four Modes

1. module — Module/File Analysis

Collect context centered on a specific module or file.

What to collect:

• Full contents of the target source files
• Imported/dependency files (1 level deep)
• Protocol/interface/type definitions
• Related test files
• Package/module declarations (Package.swift, package.json, Cargo.toml, etc.)

Detection strategy:

1. Glob all source files under the target path
2. Parse import statements in each file → Glob the dependency files
3. Grep for protocol/interface keywords
4. Match test files with the patterns *Test*, *Spec*, *_test*

2. arch — Architecture Analysis

Summarize the overall project structure and design.

What to collect:

• Directory tree (depth 3, excluding: node_modules, .git, build, .build, DerivedData, Pods, venv, pycache)
• Project docs such as CLAUDE.md, README.md, ARCHITECTURE.md
• Dependency files (Package.swift, package.json, Podfile, Cargo.toml, go.mod, requirements.txt, etc.)
• Build system config (Makefile, Tuist, *.xcodeproj configuration, etc.)
• Key config files (.env.example, tsconfig.json, .swiftlint.yml, etc.)

3. issue — Issue/Error Analysis

Collect the context around an error or bug.

What to collect:

• Source files referenced in the error message or stack trace
• Grep results for the error string (file + surrounding context)
• Call chain tracing (callers/callees)
• Related test files
• Recent git history (last 10 commits for the relevant files)
• Environment info (OS, runtime version, dependency versions)

4. custom — User-Specified

The user specifies the scope interactively.

Flow:

1. Ask the user which files/directories/topics to include
2. Collect file contents for the selected scope
3. Have the user enter the research question directly
4. Assemble the prompt, then copy to clipboard

Output Format

The generated prompt has a four-section structure:

# GPT-PRO Research Request

## Role
[모드에 따른 역할 지정]

## Context
[추출된 프로젝트 맥락 — 파일 경로 헤더 + 코드 블록]

## Research Request
[구체적 리서치 질문]

## Expected Output
[GPT-PRO 응답 envelope 포맷 지정]

GPT-PRO Response Envelope

In the Expected Output section, request the following structured response format from GPT-PRO:

Section	Required	Content
Summary	O	Key summary of the analysis in 3 lines or fewer
Findings	O	Per-mode analysis body (including file paths + line numbers)
Code Suggestions	△	Code change suggestions — file path, line, change type, code block
Action Items	O	P0-Critical / P1-Important / P2-Suggestion priority checklist
References	△	External reference material

This format is designed so that pasting the GPT-PRO response into Claude Code lets follow-up work begin immediately.

See references/output-templates.md for the detailed template.

Execution Flow

1. Determine mode
   ├─ Specified via argument → that mode
   ├─ No argument + error message present → issue
   ├─ No argument + file/directory specified → module
   └─ No argument + scope unclear → ask the user for the mode

2. Detect target
   ├─ module: path → Glob → source file list
   ├─ arch: project root → directory tree + core files
   ├─ issue: parse error message → locate related files
   └─ custom: wait for user input

3. Collect context
   ├─ Read files per each mode's strategy
   └─ Reference: references/context-extraction-guide.md

4. Validate size
   ├─ < 100K chars: proceed normally
   ├─ 100K~200K chars: show warning + suggest trimming
   └─ > 200K chars: hard limit → auto-trim or chunk
   └─ Reference: references/size-limits-and-chunking.md

5. Build prompt
   ├─ Assemble into the four-section structure
   ├─ Reference: references/output-templates.md
   └─ Reference: references/prompting-best-practices.md

6. Copy to clipboard
   └─ echo "..." | pbcopy

7. Report result
   ├─ List of included files
   ├─ Total char count / estimated token count
   ├─ Whether chunked (if applicable)
   └─ "클립보드에 복사되었습니다. GPT-PRO에 붙여넣기 하세요."

Size Management

Tier	Size	Handling
Small	< 30K chars	Use as-is
Medium	30K~100K chars	Use as-is
Large	100K~200K chars	Warning + suggest trimming
Oversized	> 200K chars	Auto-trim or chunk

Trimming priority (keep the highest first):

1. Core source code
2. Interface/protocol definitions
3. Error logs/stack traces
4. Config files
5. Test code
6. Documentation
7. Dependency files

See references/size-limits-and-chunking.md for the detailed strategy.

Codex Research Sidecar (Optional)

Only when the Codex skill is installed and the user explicitly asks for it, you may run /codex:rescue in parallel with GPT-PRO prompt generation.

When to Use

• The user explicitly asks, e.g. "codex도 같이 돌려줘", "codex로도 확인해줘"
• Or propose it as a fallback when GPT-PRO is unavailable

How to Run

1. After the existing prompt generation (pbcopy) completes
2. Dispatch the codex:codex-rescue subagent in the background (read-only, without --write):
   • Task: pass along the prompt's Expected Output summary as-is
3. Check progress with /codex:status, collect results with /codex:result
4. Report the Codex result to the user separately (do not mix it with the existing pbcopy prompt)

Guardrail: The essence of this skill is GPT-PRO handoff + pbcopy. Codex is a sidecar/fallback, not a full replacement. Do not run Codex automatically without an explicit user request.

Key Rules

• Read-only: Do not modify project files. Only read and search.
• pbcopy: Copy the result to the clipboard with pbcopy — this is the skill's delivery contract.
• Korean explanations: Write the prompt's explanatory parts in Korean, and keep code and technical terms in their original form.
• One-way: Scope ends at context extraction and prompt generation. Handling the GPT response is out of scope.

References

• references/output-templates.md — Output format templates
• references/context-extraction-guide.md — Per-mode context extraction strategies
• references/prompting-best-practices.md — GPT-PRO prompting best practices
• references/size-limits-and-chunking.md — Size management and chunking guide