Deep Research

Deep Research Skill

Trigger

Activate this skill when the user wants to:

• "Research a topic", "literature review", "find papers about", "survey papers on"
• "Deep dive into [topic]", "what's the state of the art in [topic]"
• Uses /research <topic> slash command

Overview

This skill conducts systematic academic literature reviews in 6 phases, producing structured notes, a curated paper database, and a synthesized final report. Output is organized by phase for clarity.

Installation: ~/.claude/skills/deep-research/ — scripts, references, and this skill definition. Output: .//Users/lingzhi/Code/deep-research-output/{slug}/ relative to the current working directory.

CRITICAL: Strict Sequential Phase Execution

You MUST execute all 6 phases in strict order: 1 → 2 → 3 → 4 → 5 → 6. NEVER skip any phase.

This is the single most important rule of this skill. Violations include:

• ❌ Jumping from Phase 2 to Phase 5/6 (skipping Deep Dive and Code)
• ❌ Writing synthesis or report before completing Phase 3 deep reading
• ❌ Producing a final report based only on abstracts/titles from search results
• ❌ Combining or merging phases (e.g., doing "Phase 3-5 together")

Phase Gate Protocol

Before starting Phase N+1, you MUST verify that Phase N's required output files exist on disk. If they don't exist, you have NOT completed that phase.

After completing each phase, print a phase completion checkpoint:

✅ Phase N complete. Output: [list files written]. Proceeding to Phase N+1.

Why Every Phase Matters

• Phase 3 (Deep Dive) is where you actually READ papers — without it, your synthesis is superficial and based only on abstracts
• Phase 4 (Code & Tools) grounds the research in practical implementations — without it, you miss the open-source ecosystem
• Phase 5 (Synthesis) requires deep knowledge from Phase 3 — you cannot synthesize papers you haven't read
• Phase 6 (Report) assembles content from ALL prior phases — it should cite specific findings from Phase 3 notes

Paper Quality Policy

Peer-reviewed conference papers take priority over arXiv preprints. Many arXiv papers have not undergone peer review and may contain unverified claims.

Source Priority (highest to lowest)

1. Top AI conferences: NeurIPS, ICLR, ICML, ACL, EMNLP, NAACL, AAAI, IJCAI, CVPR, KDD, CoRL
2. Peer-reviewed journals: JMLR, TACL, Nature, Science, etc.
3. Workshop papers: NeurIPS/ICML workshops (lower bar but still reviewed)
4. arXiv preprints with high citations: Likely high-quality but unverified
5. Recent arXiv preprints: Use cautiously, note "preprint" status explicitly

When to Use arXiv Papers

• As supplementary evidence alongside peer-reviewed work
• For very recent results (< 3 months old) not yet at conferences
• When a peer-reviewed version doesn't exist yet — note (preprint) in citations
• For survey/review papers (these are useful even without peer review)

Search Tools (by priority)

1. paper_finder (primary — conference papers only)

Location: /Users/lingzhi/Code/documents/tool/paper_finder/paper_finder.py

Searches ai-paper-finder.info (HuggingFace Space) for published conference papers. Supports filtering by conference + year. Outputs JSONL with BibTeX.

python /Users/lingzhi/Code/documents/tool/paper_finder/paper_finder.py --mode scrape --config <config.yaml>
python /Users/lingzhi/Code/documents/tool/paper_finder/paper_finder.py --mode download --jsonl <results.jsonl>
python /Users/lingzhi/Code/documents/tool/paper_finder/paper_finder.py --list-venues

Config example:

searches:
  - query: "long horizon reasoning agent"
    num_results: 100
    venues:
      neurips: [2024, 2025]
      iclr: [2024, 2025, 2026]
      icml: [2024, 2025]
output:
  root: /Users/lingzhi/Code/deep-research-output/{slug}/phase1_frontier/search_results
  overwrite: true

2. search_semantic_scholar.py (supplementary — citation data + broader coverage)

Location: /Users/lingzhi/.claude/skills/deep-research/scripts/search_semantic_scholar.py Supports --peer-reviewed-only and --top-conferences filters. API key: /Users/lingzhi/Code/keys.md (field S2_API_Key)

3. search_arxiv.py (supplementary — latest preprints)

Location: /Users/lingzhi/.claude/skills/deep-research/scripts/search_arxiv.py For searching recent papers not yet published at conferences. Mark citations with (preprint).

Other Scripts

WebFetch Mode (no Bash)

1. Paper discovery: WebSearch + WebFetch to query Semantic Scholar/arXiv APIs
2. Paper reading: WebFetch on ar5iv HTML or Read tool on downloaded PDFs
3. Writing: Write tool for JSONL, notes, report files

6-Phase Workflow

Phase 1: Frontier

Search the latest conference proceedings and preprints to understand current trends.

1. Write phase1_frontier/paper_finder_config.yaml targeting latest 1-2 years
2. Run paper_finder scrape
3. WebSearch for latest accepted paper lists
4. Identify trending directions, key breakthroughs

→ Output: phase1_frontier/frontier.md, phase1_frontier/search_results/

Phase 2: Survey

Build a comprehensive landscape with broader time range. Target 35-80 papers after filtering.

1. Write phase2_survey/paper_finder_config.yaml covering 2023-2025
2. Run paper_finder + Semantic Scholar + arXiv
3. Merge all results: python /Users/lingzhi/.claude/skills/deep-research/scripts/paper_db.py merge
4. Filter to 35-80 most relevant: python /Users/lingzhi/.claude/skills/deep-research/scripts/paper_db.py filter --min-score 0.80 --max-papers 70
5. Cluster by theme, write survey notes

→ Output: phase2_survey/survey.md, phase2_survey/search_results/, paper_db.jsonl

Phase 3: Deep Dive ⚠️ DO NOT SKIP

This phase is MANDATORY. You must actually READ 8-15 full papers, not just their abstracts.

1. Select 8-15 papers from paper_db.jsonl with rationale → write phase3_deep_dive/selection.md
2. Download PDFs: python download_papers.py --jsonl paper_db.jsonl --output-dir phase3_deep_dive/papers/ --sort-by-citations --max-downloads 15
3. For EACH selected paper, read the full text (PDF via Read or HTML via WebFetch on ar5iv)
4. Write detailed structured notes per paper (see note-format.md template): problem, contributions, methodology, experiments, limitations, connections
5. Write ALL notes → phase3_deep_dive/deep_dive.md

Phase 3 Gate: deep_dive.md must contain detailed notes for ≥8 papers, each with methodology and experiment sections filled in. Abstract-only summaries do NOT count.

→ Output: phase3_deep_dive/selection.md, phase3_deep_dive/deep_dive.md, phase3_deep_dive/papers/

Phase 4: Code & Tools ⚠️ DO NOT SKIP