chem-nlm-helper

# chem-nlm-helper — NotebookLM Helper for Chemistry Papers ## What this does This skill is a thin adapter between `paper-mentor` and NotebookLM. It handles: 1. Detecting whether NotebookLM CLI / MCP is installed 2. Creating notebooks 3. Adding sources (PDFs, URLs, abstracts) 4. Running structured queries (the 4-question summary) 5. Returning citation-grounded answers If NotebookLM is not installed, this skill guides the user through install before doing anything else. --- ## Pre-flight check **ALWAYS run before any NotebookLM call:** 1. Check for MCP tools: look for `mcp__notebooklm-mcp__*` tools in the available tool list 2. Check for CLI: run `which nlm` via Bash 3. Decide: ``` if has_mcp_tools: use_mode = "mcp" elif has_cli: use_mode = "cli" else: # NotebookLM not installed return install_guidance() ``` ### Install guidance (when both missing) Tell the user: > 「我看起來找不到 NotebookLM 工具。Phase 2 文獻陪讀需要它。 > > 你有兩個選擇： > A. 跑 `bash INSTALL.sh`（或 Windows 上 `INSTALL.ps1`）一鍵裝起來 > B. 跳過 NotebookLM，我幫你手動讀 abstract（會慢一點，但可以） > > 要選哪個？」 If they choose A, point them to: > 「在 paper-mentor/ 資料夾下跑 `bash tools/install_nlm.sh`，會自動裝好 nlm CLI 和 MCP server。完成後跑 `nlm login` 連 Google 帳號，然後重新進來就可以了。」 If they choose B, fall back to manual mode (skip Steps 2-4 below; do summaries by reading abstracts via WebFetch). --- ## Step 1: Create notebook ```python # MCP mode mcp__notebooklm-mcp__notebook_create( title="[student topic] — literature review", description="Auto-generated by paper-mentor for Phase 2" ) # Returns notebook_id # CLI mode nlm notebook create "[student topic] — literature review" ``` Save the notebook_id for subsequent calls. --- ## Step 2: Add sources For each paper from `chem-paper-search` results: | Source type | Best | Fallback | |-------------|------|----------| | Has openAccessPdf URL | Add as `url` | — | | Has Semantic Scholar page only | Add the page URL as `url` | abstract as `text` | | Has DOI only | Construct DOI URL → add as `url` | abstract as `text` | | Nothing | Add abstract as `text` source | — | ```python # MCP mode mcp__notebooklm-mcp__source_add( notebook_id=notebook_id, source_type="url", url="https://example.com/paper.pdf", wait=True ) # CLI mode nlm source add --notebook {id} --url {pdf_url} ``` **Batch processing**: add all sources before running queries (NotebookLM processing time can be 30s-2min per source). **Size limits**: - PDF >50MB → split or use abstract instead - Total sources per notebook → ≤50 (NotebookLM limit) --- ## Step 3: Run the 4-question summary For **each paper**, run a NotebookLM query restricted to that one source. Use these 4 questions (the chieh-ting-lin profile may customize wording): ``` 1. 用什麼方法解什麼問題？(具體技術 + 具體挑戰) 2. 結果改進多少？(要具體數字，PCE / 效率 / 穩定性) 3. 機制解釋夠扎實嗎？(只是現象學描述還是有 cross-characterization?) 4. 跟我們的研究方向是補充 / 競爭 / 啟發？ ``` ```python # MCP mode — exact parameter name varies by MCP server version. # Try in order: source_ids, source_filter, sources. Inspect the tool schema first # via ListMcpResourcesTool / look at notebook_query docstring before calling. mcp__notebooklm-mcp__notebook_query( notebook_id=notebook_id, query="用什麼方法解什麼問題？...", # parameter name to be confirmed at call time: source_ids=[paper_source_id] # or source_filter=, depending on MCP version ) # CLI mode (canonical) nlm chat {notebook_id} "用什麼方法解什麼問題？" --sources {source_id} ``` **Important**: if you cannot confirm the per-source restriction parameter, fall back to CLI mode (which is well-documented). Per-source restriction is required — without it, queries mix all papers in the notebook and the per-paper summary is meaningless. **Output structure** (per paper): ```markdown ### [Paper title] ([Year]) **Method**: [from Q1, with citations] **Result (numbers)**: [from Q2] **Mechanism rigor**: ⭐⭐⭐ (3/5) — [reasoning from Q3] **Relation**: 補充 / 競爭 / 啟發 — [from Q4, one-line why] ``` --- ## Step 4: Cross-paper synthesis (optional) If user wants a literature map (not just per-paper summaries), run synthesis queries on the WHOLE notebook (not source-restricted): ``` 1. 「這些論文最有共識的發現是什麼？」 2. 「最有爭議或矛盾的點是什麼？」 3. 「哪些方法是重複出現的標配？」 4. 「最近 2 年的趨勢跟更早的有什麼不同？」 ``` This produces the **gap report** input for `paper-mentor` Phase 2 Step 5. --- ## Step 5: Cleanup (optional) When done with the project: ```python # MCP mode mcp__notebooklm-mcp__notebook_delete(notebook_id=notebook_id, confirm=True) ``` Don't auto-delete — students may want to come back. Only delete on explicit request. --- ## Common issues | Issue | Fix | |-------|-----| | `nlm login` failed | Check Google account permissions for NotebookLM. Try again. | | PDF processing stuck | Wait up to 2 min, then check `nlm source list`. If stuck, re-add. | | Query returns no citations | Source may have failed processing. Re-check source status. | | MCP tool not responding | Restart Claude Code session; MCP server may have died. | | Quota exceeded | Free tier of NotebookLM has limits. Upgrade or use less. | --- ## Hard rules 1. **Never run queries without sources** — they will hallucinate. 2. **Always check `wait=True` succeeded** before querying — querying mid-processing returns garbage. 3. **Always pass `source_filter`** for per-paper questions (Step 3) — without it, NotebookLM mixes papers. 4. **Quote citations exactly** as NotebookLM returns them — do not paraphrase paper titles. 5. **If install was skipped, fall back gracefully** — don't crash.