github-research（GitHub 研究）

GitHub Research Skill

Trigger

Activate this skill when the user wants to:

• "Find repos for [topic]", "GitHub research on [topic]"
• "Analyze open-source code for [topic]"
• "Find implementations of [paper/technique]"
• "Which repos implement [algorithm]?"
• Uses /github-research <deep-research-output-dir> slash command

Overview

This skill systematically discovers, evaluates, and deeply analyzes GitHub repositories related to a research topic. It reads deep-research output (paper database, phase reports, code references) and produces an actionable integration blueprint for reusing open-source code.

Installation: ~/.claude/skills/github-research/ — scripts, references, and this skill definition. Output: ./github-research-output/{slug}/ relative to the current working directory. Input: A deep-research output directory (containing paper_db.jsonl, phase reports, code_repos.md, etc.)

6-Phase Pipeline

Phase 1: Intake     → Extract refs, URLs, keywords from deep-research output
Phase 2: Discovery  → Multi-source broad GitHub search (50-200 repos)
Phase 3: Filtering  → Score & rank → select top 15-30 repos
Phase 4: Deep Dive  → Clone & deeply analyze top 8-15 repos (code reading)
Phase 5: Analysis   → Per-repo reports + cross-repo comparison
Phase 6: Blueprint  → Integration/reuse plan for research topic

Output Directory Structure

github-research-output/{slug}/
├── repo_db.jsonl                     # Master repo database
├── phase1_intake/
│   ├── extracted_refs.jsonl          # URLs, keywords, paper-repo links
│   └── intake_summary.md
├── phase2_discovery/
│   ├── search_results/               # Raw JSONL from each search
│   └── discovery_log.md
├── phase3_filtering/
│   ├── ranked_repos.jsonl            # Scored & ranked subset
│   └── filtering_report.md
├── phase4_deep_dive/
│   ├── repos/                        # Cloned repos (shallow)
│   ├── analyses/                     # Per-repo analysis .md files
│   └── deep_dive_summary.md
├── phase5_analysis/
│   ├── comparison_matrix.md          # Cross-repo comparison
│   ├── technique_map.md              # Paper concept → code mapping
│   └── analysis_report.md
└── phase6_blueprint/
    ├── integration_plan.md           # How to combine repos
    ├── reuse_catalog.md              # Reusable components catalog
    ├── final_report.md               # Complete compiled report
    └── blueprint_summary.md

Scripts Reference

All scripts are Python 3, stdlib-only, located in ~/.claude/skills/github-research/scripts/.

Phase 1: Intake

Goal: Extract all relevant references, URLs, and keywords from the deep-research output.

Steps

1. Create output directory structure: SLUG=$(echo "$TOPIC" | tr '[:upper:]' '[:lower:]' | tr ' ' '-' | tr -cd 'a-z0-9-') mkdir -p github-research-output/$SLUG/{phase1_intake,phase2_discovery/search_results,phase3_filtering,phase4_deep_dive/{repos,analyses},phase5_analysis,phase6_blueprint}
2. Extract references from deep-research output: python ~/.claude/skills/github-research/scripts/extract_research_refs.py \ --research-dir <deep-research-output-dir> \ --output github-research-output/$SLUG/phase1_intake/extracted_refs.jsonl
3. Review extracted refs: Read the generated JSONL. Note:

- GitHub URLs found directly in reports - Paper titles and arxiv IDs (for Papers With Code lookup) - Research keywords and themes (for GitHub search queries)

1. Write intake summary: Create phase1_intake/intake_summary.md with:

- Number of direct GitHub URLs found - Number of papers with potential code links - Key research themes extracted - Planned search queries for Phase 2

Checkpoint

• extracted_refs.jsonl exists with entries
• intake_summary.md written
• Search strategy documented

Phase 2: Discovery

Goal: Cast a wide net to find 50-200 candidate repos from multiple sources.

Steps

1. Search by direct URLs: Any GitHub URLs from Phase 1 → fetch metadata: python ~/.claude/skills/github-research/scripts/repo_metadata.py \ --repos owner1/name1 owner2/name2... \ --output github-research-output/$SLUG/phase2_discovery/search_results/direct_urls.jsonl
2. Search Papers With Code: For each paper with an arxiv ID: python ~/.claude/skills/github-research/scripts/search_paperswithcode.py \ --arxiv-id 2401.12345 \ --output github-research-output/$SLUG/phase2_discovery/search_results/pwc_2401.12345.jsonl
3. Search GitHub by keywords (3-8 queries based on research themes): python ~/.claude/skills/github-research/scripts/search_github.py \ --query "multi-agent LLM coordination" \ --min-stars 10 --sort stars --max-results 50 \ --output github-research-output/$SLUG/phase2_discovery/search_results/gh_query1.jsonl
4. Search GitHub code (for specific implementations): python ~/.claude/skills/github-research/scripts/search_github_code.py \ --query "class MultiAgentOrchestrator" \ --language python --max-results 30 \ --output github-research-output/$SLUG/phase2_discovery/search_results/code_query1.jsonl
5. Fetch READMEs for repos that lack descriptions: python ~/.claude/skills/github-research/scripts/repo_readme_fetch.py \ --input <repos.jsonl> \ --output github-research-output/$SLUG/phase2_discovery/search_results/readmes.jsonl
6. Merge all results into master database: python ~/.claude/skills/github-research/scripts/repo_db.py merge \ --inputs github-research-output/$SLUG/phase2_discovery/search_results/*.jsonl \ --output github-research-output/$SLUG/repo_db.jsonl
7. Write discovery log: Create phase2_discovery/discovery_log.md with search queries used, results per source, total unique repos found.

Rate Limits

• GitHub search API: 30 requests/minute (authenticated)
• Papers With Code API: No strict limit but be respectful (1 req/sec)
• Add --delay 1.0 to batch operations when needed

Checkpoint

• repo_db.jsonl populated with 50-200 repos
• discovery_log.md with search details

Phase 3: Filtering

Goal: Score and rank repos, select top 15-30 for deeper analysis.

Steps

1. Enrich metadata for all repos: python ~/.claude/skills/github-research/scripts/repo_metadata.py \ --input github-research-output/$SLUG/repo_db.jsonl \ --output github-research-output/$SLUG/repo_db.jsonl \ --delay 0.5
2. Score repos (quality + activity scores): python ~/.claude/skills/github-research/scripts/repo_db.py score \ --input github-research-output/$SLUG/repo_db.jsonl \ --output github-research-output/$SLUG/repo_db.jsonl
3. LLM relevance scoring: Read through the top ~50 repos (by quality_score) and assign relevance_score (0.0-1.0) based on: python ~/.claude/skills/github-research/scripts/repo_db.py tag \ --input github-research-output/$SLUG/repo_db.jsonl \ --ids owner/name --tags "relevance:0.85"

- Direct relevance to research topic - Implementation completeness - Code quality signals (from README, description) - Update the relevance scores:

1. Compute composite scores and rank: python ~/.claude/skills/github-research/scripts/repo_db.py score \ --input github-research-output/$SLUG/repo_db.jsonl \ --output github-research-output/$SLUG/repo_db.jsonl python ~/.claude/skills/github-research/scripts/repo_db.py rank \ --input github-research-output/$SLUG/repo_db.jsonl \ --output github-research-output/$SLUG/phase3_filtering/ranked_repos.jsonl \ --by composite_score
2. Select top repos: Filter to top 15-30: python ~/.claude/skills/github-research/scripts/repo_db.py filter \ --input github-research-output/$SLUG/phase3_filtering/ranked_repos.jsonl \ --output github-research-output/$SLUG/phase3_filtering/ranked_repos.jsonl \ --max-repos 30 --not-archived
3. Write filtering report: Create phase3_filtering/filtering_report.md:

- Stats before/after filtering - Score distributions - Top 30 repos with scores and rationale

Scoring Formula

activity_score = sigmoid((days_since_push < 90) * 0.4 + has_recent_commits * 0.3 + open_issues_ratio * 0.3)
quality_score  = normalize(log(stars+1) * 0.3 + log(forks+1) * 0.2 + has_license * 0.15 + has_readme * 0.15 + not_archived * 0.2)
composite_score = relevance * 0.4 + quality * 0.35 + activity * 0.25

Checkpoint

• ranked_repos.jsonl with 15-30 repos
• filtering_report.md with scoring details

Phase 4: Deep Dive

Goal: Clone and deeply analyze the top 8-15 repos.

Steps

1. Select repos for deep dive: Take top 8-15 from ranked list.
2. Clone each repo (shallow): python ~/.claude/skills/github-research/scripts/clone_repo.py \ --repo owner/name \ --output-dir github-research-output/$SLUG/phase4_deep_dive/repos/
3. Analyze structure for each cloned repo: python ~/.claude/skills/github-research/scripts/analyze_repo_structure.py \ --repo-dir github-research-output/$SLUG/phase4_deep_dive/repos/name/ \ --output github-research-output/$SLUG/phase4_deep_dive/analyses/name_structure.json
4. Extract dependencies: python ~/.claude/skills/github-research/scripts/extract_dependencies.py \ --repo-dir github-research-output/$SLUG/phase4_deep_dive/repos/name/ \ --output github-research-output/$SLUG/phase4_deep_dive/analyses/name_deps.json
5. Find implementations: Search for key algorithms/concepts from research: python ~/.claude/skills/github-research/scripts/find_implementations.py \ --repo-dir github-research-output/$SLUG/phase4_deep_dive/repos/name/ \ --patterns "class Transformer" "def forward" "attention" \ --output github-research-output/$SLUG/phase4_deep_dive/analyses/name_impls.jsonl
6. Deep code reading: For each repo, READ the key source files identified by structure analysis. Write a per-repo analysis in phase4_deep_dive/analyses/{name}_analysis.md:

- Architecture overview - Key algorithms implemented - Code quality assessment - API / interface design - Dependencies and requirements - Strengths and limitations - Reusability assessment (how easy to extract components)

1. Write deep dive summary: phase4_deep_dive/deep_dive_summary.md

IMPORTANT: Actually Read Code

Do NOT just summarize READMEs. You must:

• Read the main source files (entry points, core modules)
• Understand the actual implementation approach
• Identify specific functions/classes that implement research concepts
• Note code patterns, design decisions, and trade-offs

Checkpoint

• Repos cloned in repos/
• Per-repo analysis files in analyses/
• deep_dive_summary.md written

Phase 5: Analysis

Goal: Cross-repo comparison and technique-to-code mapping.

Steps

1. Generate comparison matrix: python ~/.claude/skills/github-research/scripts/compare_repos.py \ --input github-research-output/$SLUG/phase4_deep_dive/analyses/ \ --output github-research-output/$SLUG/phase5_analysis/comparison.json
2. Write comparison matrix: Create phase5_analysis/comparison_matrix.md:

- Table comparing repos across dimensions (language, LOC, stars, framework, license, tests) - Dependency overlap analysis - Strengths/weaknesses per repo

1. Write technique map: Create phase5_analysis/technique_map.md:

- Map each paper concept / research technique → specific repo + file + function - Identify gaps (techniques with no implementation found) - Note alternative implementations of the same concept

1. Write analysis report: phase5_analysis/analysis_report.md:

- Executive summary of findings - Key insights from code analysis - Recommendations for which repos to use for which purposes

Checkpoint

• comparison_matrix.md with repo comparison table
• technique_map.md mapping concepts to code
• analysis_report.md with findings

Phase 6: Blueprint

Goal: Produce an actionable integration and reuse plan.

Steps

1. Write integration plan: phase6_blueprint/integration_plan.md:

- Recommended architecture for combining repos - Step-by-step integration approach - Dependency resolution strategy - Potential conflicts and how to resolve them

1. Write reuse catalog: phase6_blueprint/reuse_catalog.md:

- For each reusable component: source repo, file path, function/class, what it does, how to extract it - License compatibility matrix - Effort estimates (easy/medium/hard to integrate)

1. Compile final report: python ~/.claude/skills/github-research/scripts/compile_github_report.py \ --topic-dir github-research-output/$SLUG/
2. Write blueprint summary: phase6_blueprint/blueprint_summary.md:

- One-page executive summary - Top 5 repos and why - Recommended next steps

Checkpoint

• integration_plan.md complete
• reuse_catalog.md with component catalog
• final_report.md compiled
• blueprint_summary.md as executive summary

Quality Conventions

1. Repos are ranked by composite score: relevance × 0.4 + quality × 0.35 + activity × 0.25
2. Deep dive requires reading actual code, not just READMEs
3. Integration blueprint must map paper concepts → specific code files/functions
4. Incremental saves: Each phase writes to disk immediately
5. Checkpoint recovery: Can resume from any phase by checking what outputs exist
6. All scripts are stdlib-only Python — no pip installs needed
7. gh CLI is required for GitHub API access (must be authenticated)
8. Deduplication by repo_id (owner/name) across all searches
9. Rate limit awareness: Respect GitHub search API limits (30 req/min)

Error Handling

• If gh is not installed: warn user and provide installation instructions
• If a repo is archived/deleted: skip gracefully, note in log
• If clone fails: skip, note in log, continue with remaining repos
• If Papers With Code API is down: skip, rely on GitHub search only
• Always write partial progress to disk so work is not lost

References

• See references/phase-guide.md for detailed phase execution guidance
• Deep-research skill: ~/.claude/skills/deep-research/SKILL.md
• Paper database pattern: ~/.claude/skills/deep-research/scripts/paper_db.py