edwinhu/research

Academic Literature Search

Multi-source academic search with deduplication, DOI resolution, and journal filtering.

Always read ${CLAUDE_SKILL_DIR}/../google-scholar/domain-knowledge.local.md before presenting results.

IRON LAW: Always Use the Script

NEVER run the sources manually in sequence. ALWAYS use the research script. This is not negotiable.

uv run python3 "${CLAUDE_SKILL_DIR}/scripts/research.py" "<query>" [--n 50] [--min-citations N]

The script parallelizes all sources and DOI resolution automatically. Doing it manually serializes everything and triples wall time.

Sources

| Source | Tool | Strength | Default | |--------|------|----------|---------| | scholar lookup | Keyword/citation-ranked | Finance classics, foundational papers | ✅ | | consensus CLI | Empirical corpus, sorted by citations | Accounting/finance empirical literature | ✅ | | Paperpile bib | Personal library (My Library.bib) | Papers already in your collection | ✅ | | scholar search | NL semantic | Law reviews, conceptual literature | opt-in (--scholar-search) |

scholar search is opt-in because it shares rate limits with scholar lookup and 429s when run in parallel. Add --scholar-search when you specifically want semantic/NL results.

Output Schema

The script outputs a JSON array. Each paper has:

{
  "title": "...",
  "authors": ["..."],
  "year": 2023,
  "journal": "...",           // original journal label (may be SSRN)
  "journal_resolved": "...",  // CrossRef-resolved journal (present if SSRN label was resolved)
  "doi": "...",
  "citations": 150,
  "takeaway": "...",
  "url": "...",
  "sources": ["lookup", "consensus"]  // all sources that returned this paper
}

LLM Review Step (After Script)

After running the script, read ${CLAUDE_SKILL_DIR}/../google-scholar/domain-knowledge.local.md and cross-reference each paper's effective journal (use journal_resolved if present, else journal) against the trusted list:

• ★ = journal matches trusted list
• Papers in sources: ["lookup", "consensus"] (multiple sources) = higher confidence
• Papers from bib source = already in user's library (flag with 📚)

Presentation Format

★ [Title](url) — Authors (Year), *Journal*, N citations  [sources]
  > Takeaway: ...

📚 ★ [Title](url) — Authors (Year), *Journal*  [in your library]
  > Takeaway: ...

Trusted papers first (sorted by citations desc), then non-trusted in a collapsed table.

Red Flags

• About to run the sources manually in sequence → STOP. That serializes the work and triples wall time; run uv run python3 research.py "<query>".
• About to call mcp__consensus__search → STOP. It is rate-limited to 3 results; the script uses the CLI binary automatically.
• About to present results before reading domain-knowledge.local.md → STOP. The ★ trusted-journal signals come from that file; read it first, always.
• About to use the journal field when journal_resolved is present → STOP. The SSRN label hides the real venue; always prefer journal_resolved.

Common Patterns

# Standard search
uv run python3 "${CLAUDE_SKILL_DIR}/scripts/research.py" "mandatory disclosure"

# With citation floor
uv run python3 "${CLAUDE_SKILL_DIR}/scripts/research.py" "poison pill" --min-citations 50

# More results from Consensus
uv run python3 "${CLAUDE_SKILL_DIR}/scripts/research.py" "corporate governance" --n 100

# Disable streaming (wait for all sources, output pretty-printed JSON)
uv run python3 "${CLAUDE_SKILL_DIR}/scripts/research.py" "mandatory disclosure" --no-stream

Streaming Mode (default)

Without --stream, the script waits for all four sources before emitting anything — Consensus takes ~60s, so fast sources (bib <1s, Scholar ~10s) sit idle.

With --stream, the script emits one NDJSON line per event as it happens:

{"event": "source", "source": "bib", "papers": [...]}
{"event": "source", "source": "scholar-lookup", "papers": [...]}
{"event": "source", "source": "scholar-search", "papers": [...]}
{"event": "source", "source": "consensus", "papers": [...]}
{"event": "final", "papers": [...]}

• source events: raw papers from each source as it completes (may have duplicates across sources)
• final event: deduplicated + CrossRef-resolved unified set

Process source events as they arrive to present early results; use final for the complete deduplicated list. Pass --no-stream for batch mode (pretty-printed JSON after all sources complete).