research-clustering
Identifies research domains from text or keywords, clusters them into structured sub-areas, and generates an overview, keywords, and research strategy for each cluster. Use this skill for the initial
SKILL.md.claude/skills/research-clustering/SKILL.md
GitHubResearch Clustering — Domain Identification, Partitioning & Structuring
Takes text or keyword groups as input, identifies the academic/technical domains worth investigating, clusters (partitions) them, and outputs an overview, keywords, and research strategy for each cluster. This skill handles the "map-making" phase of a research project.
Auto Mode (--auto)
When $ARGUMENTS contains --auto, run the entire workflow non-interactively — skip ALL AskUserQuestion calls and use the following defaults:
| Parameter | Default Value |
|---|---|
| Research Type | Academic Paper Survey |
| Time Range | Last 4 years |
| Search Languages | English + Japanese |
| Output Granularity | Standard |
| Next Action (Step 7) | Done (自動終了) |
In --auto mode, the remaining text in $ARGUMENTS (after removing --auto) is used as the research theme input. For example: /research-clustering --auto LLM agent orchestration → theme is "LLM agent orchestration".
If $ARGUMENTS does NOT contain --auto, proceed with the normal interactive workflow below.
Workflow
Step 1: Parse Input
Analyze the user's input and extract research seeds.
Accepted input formats (support all):
- Markdown file (keyword list, research theme description, etc.)
- Conversational text (natural-language description of the research topic)
- Any combination of the above
Elements to extract:
- Core keywords (technical terms, proper nouns, abbreviations)
- Research purpose and context (why do they want to investigate this?)
- Target domains (academic, patent, business, technology trends, etc.)
- Known constraints (excluded fields, language, region, etc.)
Step 1.5: Preliminary Scan (Quick Scan)
Before the user hearing, run a lightweight web search on the input keywords to form a hypothesis about the research landscape. The goal is to improve the quality of the hearing — instead of asking vague questions like "what do you want to research?", present concrete options like "here's the field structure I see — which direction should we go deeper?"
Preliminary search (2–3 queries, keep it lightweight):
"{core keyword} survey"or"{core keyword} overview"to grasp the overall landscape"{core keyword} applications"or"{core keyword} use cases"to check application areas- If the input keywords are in Japanese, translate to English and search in both languages
Information to extract from the preliminary scan:
- Major sub-areas of the field (tentative cluster candidates)
- Recent trends and hot topics
- Potentially related adjacent fields
- Whether the field is paper-heavy, patent-heavy, or practice/case-study-heavy
Present the hypothesis: Summarize the preliminary scan results concisely and present them during the hearing as: "Based on a quick scan, the field appears to be structured as follows..." This lets the user compare against their own research intent and give more precise direction.
Step 2: User Hearing
--automode: Skip this entire step. Use the default values from the Auto Mode table above.
Confirm research parameters via AskUserQuestion, informed by the preliminary scan results. Reflect scan-derived hypotheses in default values and option descriptions to reduce user burden.
Hearing 1: Research Type
AskUserQuestion:
question: "What type of research would you like to conduct? (multiple selection)"
header: "Research Type"
multiSelect: true
options:
- label: "Academic Paper Survey"
description: "Focus on academic papers (arXiv, IEEE, ACM, etc.)"
- label: "Patent Search"
description: "Focus on patent literature (USPTO, EPO, JPO, etc.)"
- label: "Technology Trend Analysis"
description: "Tech blogs, conference talks, OSS projects, etc."
- label: "Business Case Study"
description: "Enterprise adoption cases, market reports, industry trends"
Hearing 2: Time Range
AskUserQuestion:
question: "What time range should the research cover?"
header: "Time Range"
multiSelect: false
options:
- label: "Last 4 years (default)"
description: "Results from 2022 to present"
- label: "Last 2 years"
description: "Focus on the latest trends only"
- label: "Last 7 years"
description: "Broader coverage"
- label: "Custom range"
description: "Specify a custom time range"
If "Custom range" is selected, ask a follow-up AskUserQuestion for the specific years.
Hearing 3: Search Languages
AskUserQuestion:
question: "Which languages should be used for web searches? (multiple selection)"
header: "Search Languages"
multiSelect: true
options:
- label: "English"
description: "Search in English (covers most academic and international sources)"
- label: "Japanese"
description: "Search in Japanese (useful for JPO patents, domestic cases, Japanese papers)"
- label: "Chinese"
description: "Search in Chinese (useful for Chinese patents, CNKI papers)"
- label: "Other"
description: "Specify additional languages"
If "Other" is selected, ask a follow-up AskUserQuestion for the specific languages. Pre-select defaults based on the input: if the input contains Japanese text, default to English + Japanese; otherwise default to English only.
Hearing 4: Output Granularity
AskUserQuestion:
question: "What level of detail do you want in the output?"
header: "Output Granularity"
multiSelect: false
options:
- label: "Standard (recommended)"
description: "Domain partitioning + overview, keywords, and research strategy per cluster"
- label: "Detailed"
description: "Standard + representative papers/patents/cases per cluster"
- label: "Overview only"
description: "Domain partitioning and keywords only — quick big-picture view"
Step 3: Domain Identification (Web Search)
Based on input keywords and confirmed parameters, use WebSearch to investigate related academic/technical domains.
Search strategy:
Initial search: Broad search using core keyword combinations
"{keyword1} {keyword2} survey"/"{keyword1} {keyword2} review""{keyword} taxonomy"/"{keyword} classification"- If keywords are in Japanese, translate to English and search in both languages
For academic paper surveys, prioritize review/survey papers
- Review and survey papers provide an efficient overview of the entire field, so search for these first
- Example queries:
"{topic} survey paper {year}","{topic} systematic review","{topic} literature review" - Use the taxonomy/classification from discovered survey papers as the starting point for domain partitioning
Expansion search: Discover related fields from initial results
- Identify adjacent domains from survey paper sections and classification schemes
- Discover new axes from frequently co-occurring keywords
For patent searches: Also investigate IPC/CPC classification codes for technology classification reference
Compile a comprehensive list of candidate academic/technical domains related to the input theme.
Step 4: Domain Clustering (Partitioning)
Partition the research targets into meaningful clusters based on information collected in Step 3.
Partitioning principles:
- Each cluster should be independently researchable
- Cluster boundaries should be clear (minimize overlap)
- Aim for 3–8 clusters (not too coarse, not too fine)
- Use up to 2 levels of hierarchy (major → minor) when the structure naturally calls for it
Partitioning perspectives (select based on research type):
- By method/approach (e.g., statistical methods / deep learning / rule-based)
- By application domain (e.g., healthcare / finance / manufacturing)
- By timeline (e.g., classical methods / recent developments / latest trends)
- By problem formulation (e.g., classification / regression / generation)
Step 5: Cluster Elaboration
For each cluster, produce:
- Cluster name: A concise, descriptive title
- Overview: 2–5 sentences explaining the scope of the cluster
- Keywords: Technical terms and concepts belonging to the cluster (5–15 items)
- Research strategy: Recommended approach for deeper investigation of this cluster
- Representative resources (only for "Detailed" granularity): A few survey papers, key papers/patents/cases
For academic paper surveys, list the survey/review papers found in Step 3 as "seed resources" for the relevant cluster. Also show how the survey paper's taxonomy maps to the cluster structure.
Step 6: Output File Generation
Select the output format based on the research scale:
Small to medium scale (5 or fewer clusters): Single file
Consolidate everything into one Markdown file.
Large scale (6 or more clusters): Directory structure
{output-dir}/
index.md # Overall overview + cluster list
cluster-01-xxx.md # Detailed info for each cluster
cluster-02-xxx.md
...
Output Template (single file / index.md)
# {Research Theme}
## Research Parameters
- **Research type**: {Academic Paper Survey / Patent Search / ...}
- **Time range**: {YYYY – YYYY}
- **Generated on**: {YYYY-MM-DD}
- **Input keywords**: {original keyword list}
## Big Picture
{3–5 sentences on the overall positioning of the research theme, current state of the field, and major trends}
## Reference Survey/Review Papers
{For academic paper surveys. List discovered survey papers. Note that their taxonomies informed the domain partitioning}
| Title | Year | Summary | Link |
|-------|------|---------|------|
| {title} | {year} | {one-line summary} | {url} |
## Domain Map
{Conceptual diagram showing inter-cluster relationships. Use ASCII art or Mermaid notation}
## Cluster Summary
| # | Cluster Name | Keyword Count | Summary |
|---|-------------|---------------|---------|
| 1 | {name} | {n} | {one-line summary} |
## Cluster Details
### Cluster 1: {Cluster Name}
**Overview**: {2–5 sentence description}
**Keywords**:
`keyword1`, `keyword2`, `keyword3`, ...
**Research Strategy**:
- {Recommended search queries and information sources}
- {Notable research groups or companies}
- {Recommended reading order from survey papers}
**Representative Resources** (detailed granularity only):
| Title | Type | Year | Summary |
|-------|------|------|---------|
| {title} | Paper/Patent/Case | {year} | {summary} |
---
{Repeat for each cluster}
Step 7: Output Confirmation and Next Actions
--automode: Skip this step entirely. Treat the result as "Done" and finish.
After output is complete, confirm the next action via AskUserQuestion.
AskUserQuestion:
question: "The domain map is complete. What would you like to do next?"
header: "Next Action"
multiSelect: false
options:
- label: "Done"
description: "Finalize the current output"
- label: "Adjust clusters"
description: "Change partitioning granularity or classification axes and regenerate"
- label: "Deep-dive into a specific cluster"
description: "Investigate papers/patents for a selected cluster in detail"
- label: "Generate research prompts"
description: "Generate research prompt files for each cluster (integrates with research-prompt-builder)"
If "Deep-dive into a specific cluster" is selected, show a follow-up AskUserQuestion with cluster names as options. For deep-dive research, also suggest integrating with the research-retrieval skill to generate detailed paper lists per cluster.
Output Location
MUST READ FIRST: Before deciding the output path, read docs/research/README.md (the single source of truth for the research directory layout) and .claude/rules/research.md.
Path resolution
Identify the domain (
<domain>,snake_case):- Infer from input file path if it lives under
docs/research/(runs|domains)/<domain>/... - Otherwise infer from the research theme or ask the user
- Infer from input file path if it lives under
If
docs/research/domains/<domain>/domain.yamlexists and definesoutput_paths.clustering, use it.Otherwise use the default path:
docs/research/runs/<domain>/clustering/<YYYYMMDD>/- Output
index.mdandcluster-NN-*.mdfiles inside this directory. <YYYYMMDD>is today's date (UTC or JST is acceptable, use whichever the pipeline uses).
- Output
Never write directly under
docs/research/domains/<domain>/— that layer is composed of symlinks pointing intoruns/.Never overwrite or modify existing files under
runs/<domain>/clustering/<old_date>/— clustering is append-only. Re-clustering creates a NEW dated directory.
After writing
After the new clustering run is written, update (or instruct the pipeline/user to update) the latest pointer:
ln -snf <YYYYMMDD> docs/research/runs/<domain>/clustering/latest
ln -snf ../../runs/<domain>/clustering/latest docs/research/domains/<domain>/clustering
If the domains/<domain>/ directory does not yet exist, create it (it is just a view layer).
Parallel Processing
Run multiple web search queries in parallel for efficiency. Use the Agent tool to spawn subagents for concurrent searches.
Integration with Other Skills
- research-prompt-builder: Can generate research prompts for each cluster
- research-retrieval: Can produce detailed paper lists per cluster
Language
- User interactions (AskUserQuestion, etc.) follow the project's response language setting
- Technical terms, paper titles, and proper nouns are kept in their original language
- Web searches are conducted in the languages selected during Hearing 3 (Search Languages)
- All user-facing output, reports, and summaries must be written in Japanese(すべてのユーザーへの出力は日本語にしてください)