find-notes

name: find-notes description: "Cross-cutting knowledge search across Drive desktop (local), Drive API, and Paperpile PDFs. Returns hits grouped by source with file path / URL / snippet. Use when the user says find-notes, /find-notes, メモを探して, あのメモどこ, 論文探して, papers about X, find that doc about X, Drive で X を探して, or wants to locate any note / doc / paper from across their knowledge base." metadata: short-description: "Search across Drive (local) + Drive API + Paperpile PDFs"

Find Notes

Cross-cutting knowledge search. Wraps scripts/drive/find_notes.py and formats the result for the user. The script searches multiple tiers:

• Tier 1 (fast): filename match via fd over both Drive desktop mounts
• Tier 2 (fast): full-text grep via rg over text-like files (.md / .txt / .csv / .json) in Drive
• Tier 3 (fast): Drive API fullText contains search (UTokyo token only at present)
• Tier 4 (slow, opt-in): Paperpile PDF full-text via pdftotext + match — Stream mode fetches each PDF on demand

This skill is intended for the life repo.

Repository Rules

Before running:

1. Read README.md for directory policy if writing anywhere.
2. Read Rules.md for the safety rules (do not leak secrets / PII / unpublished data into hub Notes etc.).

This skill is read-only (no Markdown writes).

When To Invoke

User-triggered. Typical cues:

• find-notes / /find-notes
• 「メモを探して」「あのメモどこ」「Drive で X を探して」
• 「論文探して」「X についての paper」「papers about X」
• 「学会メモ」「セミナーメモ」「読書メモ」と「キーワード」がセットの時

Do NOT auto-invoke on session start or Stop hook.

Workflow

1. Run date "+%Y-%m-%d %H:%M:%S %A %Z" once (per Rules.md「日時認識」) — not because the date matters to search results, but to keep the per-turn convention.

2. Identify the search query from the user's prompt. If unclear, ask a 1-line clarifying question (do not search ambiguously).

3. Run:

   conda activate life
   python scripts/drive/find_notes.py "<query>" --limit 10
   

   • Add --fulltext-pdf if (a) the user explicitly asks for PDF body content, OR (b) the fast tiers returned almost no results AND the query looks paper-shaped (English term + "about / paper / 論文" cue).
   • Add --year YYYY with --fulltext-pdf to restrict the slow scan to one Paperpile year folder.

4. Take the script's output and present it to the user as a Markdown summary grouped by source. Trim verbose paths but keep them clickable / pastable.

5. Offer 1-2 follow-up actions:

   • python scripts/drive/read_doc.py <URL> to read a specific Doc / Sheet
   • paper-close-reading skill for a Paperpile PDF (after copying to outputs/papers/tmp/)

Output Shape

Show the user a compact grouped list. Example:

**find-notes: "ferredoxin" — 18 hit(s) across 3 source(s)**

[paperpile]
- Frankenberg & Lagarias 2003 — Phycocyanobilin–Ferredoxin Oxidoreductase
- Hagiwara et al. 2006 — Crystal structure of PcyA
- Jagilinki et al. 2025 — Designing peptide fossils … ferredoxin fold
- (5 more)

[drive-api]
- Rockwell & Clark Lagarias 2017 — Ferredoxin-dependent bilin reductases
- Beale & Cornejo 1991 — Biosynthesis of phycobilins
- (3 more)

[drive-local-text]
- 2026 Wiki: "260518 Journal 波多野 … VIPP & ferredoxin discussion"

次のアクション:
- `python scripts/drive/read_doc.py <URL>` で Doc を開く
- `paper-close-reading` skill で PDF を精読

Safety

• The script reads only; never writes durable artifacts.
• Do NOT echo any secrets or tokens; the script does not print them.
• For Paperpile PDF full-text search, warn the user that Stream mode causes per-PDF download — the first run on a year folder may take minutes.

What This Skill Does NOT Do

• Does not open files in browser / external apps.
• Does not run paper-close-reading automatically (it leaves the choice to the user).
• Does not search outside Drive / Paperpile (no Notion / Confluence / etc.).
• Does not write to hubs or inbox automatically.