english-from-paper

name: english-from-paper description: "Generate English learning artifacts from a finished paper close-reading (-ja.md must already exist). Produces vocabulary / collocation / grammar study material, a quiz + answer key, an English-only podcast script, and a NotebookLM Audio Overview customize prompt. Outputs are co-located in portfolio/paper-close-readings/. The 2 NotebookLM input files (clean-original.html + -en-study.md) plus the prompt.txt are staged to Google Drive's paper-podcasts// for manual upload to NotebookLM (typically via fenrir remote desktop). Use when the user says english-from-paper, 英語学習素材, 論文から英語学習, NotebookLM script, X04 英語, or wants vocab / quiz / podcast assets generated from a finished paper close-reading." metadata: short-description: "精読済み論文 → vocab / quiz / 英語 podcast script + NotebookLM prompt 生成、NotebookLM 投入 file を Drive paper-podcasts// に配置"

English From Paper

Purpose

Reuse a finished paper close-reading (<slug>-ja.md) as the source for 4-modality English study material. One paper → vocab/quiz (reading + writing) → English podcast (listening) → speaking material once 2-3 papers are stocked.

NotebookLM generates the actual podcast m4a. This skill only prepares the input files + customize prompt for upload.

Expected Input

The paper close-reading must already exist:

portfolio/paper-close-readings/
  <slug>.md                       ← lightweight index (paper-close-reading skill output)
  <slug>-ja.md                    ← Japanese close-reading (paper-close-reading skill output)
  tmp/<slug>/clean-original.html  ← English cleaned HTML (HTML mode) — required for NotebookLM upload
  tmp/<slug>/Fig.*.png            ← figure assets

Note the per-slug tmp subdir convention from paper-close-reading: every tmp/... path below means tmp/<slug>/....

If clean-original.html is missing (PDF-mode papers), warn the user. Either skip Drive staging or upload the raw PDF to NotebookLM instead.

Read README.md and Rules.md before any write.

Outputs

Five new files in portfolio/paper-close-readings/:

1. <slug>-en-study.md — vocab / collocation / grammar / acronyms / speaking phrases
2. <slug>-en-quiz.md — quiz questions (no answers)
3. <slug>-en-quiz-ans.md — answer key + 1-line explanations
4. <slug>-podcast-script.md — English-only reference transcript (~12 min target)
5. <slug>-notebooklm-prompt.txt — Audio Overview customize prompt (paste into NotebookLM)

The index <slug>.md is updated to link to all 5 under a new ## English study section.

Output: <slug>-en-study.md

---
slug: <slug>
type: english-study
created: YYYY-MM-DD
paper_title: <title>
audience: biology researchers (broad)
---

# English Study — <paper title>

## 1. Key Vocabulary (20-30 items)

| Term | Definition (EN) | Example from paper | 日本語 |
|---|---|---|---|
| **thylakoid** | flattened membrane sac where light reactions occur | "...thylakoid membranes are partitioned..." | チラコイド |
| ... | | | |

## 2. Collocations & Phrases (10-15 items)

- **be partitioned among** — to be divided between groups / **〜の間に分配される、分割される**
  - Paper: "Thylakoids are partitioned among daughter cells..."
  - Reuse: "Resources are partitioned among competitors."
  - 解説: `among` は 3 者以上に分ける時、`between` は 2 者の時。"partition into N groups" (N つの群に分割) と "partition among N entities" (N 者の間で分配) の違いに注意。

## 3. Grammar / Sentence Patterns (5-10 examples)

### Pattern: Reduced relative clause (V-ed by ...) / **過去分詞による関係詞節の短縮**
- Paper: "Cells divided by binary fission inherit..."
- Expanded: "Cells that are divided by binary fission inherit..."
- 用法: `the X V-ed by Y` で `the X that is/are V-ed by Y` を短縮 (関係代名詞 + be 動詞 を省略)。
- 解説: 関係代名詞を省いて分詞化することで文を引き締める。Methods / Results / 図のキャプションで多用。日本語の連体修飾 (「Y によって V された X」) と同じ感覚で読める。

## 4. Acronyms

- **FtsZ** — Filamenting temperature-sensitive Z (bacterial tubulin homolog)

## 5. Phrases for Speaking About This Paper

- "The key contribution of this work is..."
- "What I find compelling is that..."
- "One limitation worth flagging is..."

Section 5 seeds the speaking-stage practice (used later in Claude voice sessions).

Output: <slug>-en-quiz.md and <slug>-en-quiz-ans.md

Fixed 20 questions total (8 cloze / 5 collocation / 7 comprehension — see Quality Bars), mixed format:

---
slug: <slug>
type: english-quiz
question_count: 20
---

# English Quiz — <paper title>

## A. Vocabulary cloze (8 questions)

1. The thylakoid membranes are ___ among daughter cells during binary fission. (Hint: divided into groups)

## B. Collocation choice (5 questions)

9. Choose the correct preposition: "The signal gives ___ to a downstream cascade."
   (a) rise (b) raise (c) arose (d) raised

## C. Comprehension (5 questions)

14. What does Figure 1 demonstrate about thylakoid distribution?

-en-quiz-ans.md mirrors the section structure with answer + 1-line explanation each.

Output: <slug>-podcast-script.md

English-only, ~12 minute target (1200-1800 words), 2 hosts. Reference transcript only — NotebookLM improvises from the prompt + notebook sources.

---
slug: <slug>
type: podcast-script
language: English
duration_target: 12 min
purpose: NotebookLM-generated podcast reference transcript
---

# Podcast Script (English) — <paper title>

This is a written reference. Audio is generated by NotebookLM with
`<slug>-notebooklm-prompt.txt` as the customize prompt and notebook
sources being `tmp/<slug>/clean-original.html` + `<slug>-en-study.md`.

## Segment 1: Introduction (~3 min)

**Host A**: Welcome back. Today's paper is from <Author et al., Year> in <Journal>. ...

**Host B**: ...

## Segment 2: Methods overview (~3 min)
...

## Segment 3: Key findings (~4 min)
...

## Segment 4: Discussion / implications (~2 min)
...

Output: <slug>-notebooklm-prompt.txt

Plain text, English (NotebookLM accepts English prompts).

Note on duration: NotebookLM treats the "X-minute" instruction as a hint, not a hard constraint. With a content-rich source, expect the actual output to run 1.5–2× the requested duration (i.e. a "12-minute" request typically lands at 15–25 min). This is acceptable for listening practice — the 12-min framing is kept because removing it gives even longer outputs, and the per-segment time budget (3/3/4/2) still shapes pacing within the longer episode. Confirmed 2026-06-07 with Kobayashi 2014 trial: 12-min request → 18-min .m4a output.

Generate a 12-minute Audio Overview of this notebook for biology researchers.

[Structure]
- Introduction (3 min): background, the question, why it matters
- Methods overview (3 min): emphasize technique names and experimental design
- Key findings (4 min): cite specific numbers, figure numbers, statistical results
- Discussion / implications (2 min): limitations and future directions

[Language]
- English only, throughout
- Expand abbreviations on first use (full form, then the abbreviation)
- Define discipline-specific jargon briefly when first introduced

[Audience]
- Working biology researchers (broad — not field-specific)
- Academic but conversational tone, natural two-host discussion

[References]
- Cite figures and tables by number (Figure 1, Table 2, etc.)
- Quote specific numerical values from the results section

[Paper]
- Title: <TITLE>
- Authors: <AUTHORS>
- Journal / Year: <JOURNAL>, <YEAR>
- Focus: <FOCUS_HINT — 1-2 concrete sentences on the central question, extracted from abstract>

The skill auto-fills <TITLE>, <AUTHORS>, <JOURNAL>, <YEAR>, <FOCUS_HINT> from the index <slug>.md and abstract. No <PLACEHOLDER> left in the final file.

Drive Staging

NotebookLM accepts: PDF / Google Doc / Google Slide / .txt / Website URL / YouTube URL / audio. It does not accept raw .html uploads. So the staging produces a .txt version of the paper body and copies it alongside the HTML (HTML kept as offline backup, txt is what gets uploaded).

DRIVE_ROOT="$HOME/Library/CloudStorage/GoogleDrive-38kta.lab@gmail.com/マイドライブ"
DST="$DRIVE_ROOT/paper-podcasts/<slug>"
mkdir -p "$DST"
cp portfolio/paper-close-readings/tmp/<slug>/clean-original.html  "$DST/"
python3 scripts/papers/html_to_txt.py \
  portfolio/paper-close-readings/tmp/<slug>/clean-original.html \
  "$DST/clean-original.txt"
cp portfolio/paper-close-readings/<slug>-en-study.md              "$DST/"
cp portfolio/paper-close-readings/<slug>-notebooklm-prompt.txt    "$DST/"

Tell the user (in the skill's final report):

• Files staged at paper-podcasts/<slug>/ in Drive
• On fenrir, open NotebookLM → New notebook → add sources in this priority:
  1. If the paper is on PMC (open access): "Add source → Website" with the PMC URL — gives the cleanest text + caption extraction. PMC URL pattern: https://www.ncbi.nlm.nih.gov/pmc/articles/PMCxxxxxxx/
  2. Otherwise: upload clean-original.txt as a text source.
  3. Always: upload <slug>-en-study.md as an additional source (NotebookLM accepts .md).
• Audio Overview → Customize → paste contents of <slug>-notebooklm-prompt.txt
• Save the resulting .m4a (NotebookLM exports AAC .m4a, not .mp3) back to the same Drive folder. Either keep NotebookLM's auto-generated descriptive title (e.g. Three_Unique_Enzymes_for_the_Protox_Bottleneck.m4a) or rename to <slug>-podcast.m4a — both are picked up by the per-paper folder, and the descriptive title is preserved in the index entry verbatim.
• Once m4a is in place, share the Drive URL so the index can be updated

Updating the Index

Add a new section to <slug>.md:

## English study

- 学習テキスト: [<slug>-en-study](./<slug>-en-study.html)
- Quiz: [Q](./<slug>-en-quiz.html) / [A](./<slug>-en-quiz-ans.html)
- Podcast script (EN): [<slug>-podcast-script](./<slug>-podcast-script.html)
- Podcast (m4a): TBD (NotebookLM 投入後に Drive URL を記入)
- NotebookLM prompt: `<slug>-notebooklm-prompt.txt`

After m4a is created and Drive URL provided, update the Podcast (m4a) line.

regen_paper_index.py Filter

scripts/automation/regen_paper_index.py must skip the english-study sibling files so they don't appear as spurious "paper" rows. This patch is already applied in the life repo (the SKIP set below). If a fresh checkout lacks it, re-apply before running render_all:

SKIP_SUFFIXES = (
    "-ja.md",
    "-en-study.md",
    "-en-quiz.md",
    "-en-quiz-ans.md",
    "-podcast-script.md",
)
# in main():
if md.name == "README.md" or any(md.name.endswith(s) for s in SKIP_SUFFIXES):
    continue

Required Workflow

1. Read <slug>.md and <slug>-ja.md. Confirm the close-reading exists. Capture title / authors / journal / year. Extract a 1-2 sentence central question for <FOCUS_HINT>.
2. Read tmp/<slug>/clean-original.html (HTML mode) or text-extracted PDF (PDF mode) for the canonical English source.
3. Generate the 5 output files in portfolio/paper-close-readings/.
4. Stage to Drive: clean-original.html + <slug>-en-study.md + <slug>-notebooklm-prompt.txt → paper-podcasts/<slug>/.
5. Update <slug>.md index with the ## English study section (m4a link as TBD).
6. Render to portal: bash scripts/render_all.sh. (4 markdown files become HTML; the .txt is not rendered.)
7. Report: file paths, Drive folder path, NotebookLM upload instructions, and a reminder to send back the m4a Drive URL.

Quality Bars

<slug>-en-study.md:

• 20-30 vocabulary items — include high-utility academic verbs and structural phrases, not just rare jargon
• Each vocab item has: paper example (verbatim), EN definition, reuse example, JA gloss
• Collocations entries have: EN meaning + JA 訳/対応表現, paper quote, reuse example, and a 解説 line in Japanese covering register (formal/informal), preposition choice, near-synonyms, and when to use vs avoid
• Grammar patterns entries have: paper quote, expanded form (when reduced), a 用法 line in Japanese (the abstract template), and a 解説 line in Japanese explaining when/why this pattern is used, the equivalent Japanese expression, and common pitfalls
• Collocations focus on what generalizes to writing other biology papers
• Grammar patterns must be transferable (not curiosities)

<slug>-en-quiz.md:

• Fixed total of 20 questions with section breakdown:
  • A. Vocabulary cloze — 8 questions
  • B. Collocation / preposition choice — 5 questions
  • C. Comprehension — 7 questions (~1–2 sentence answers)
• Frontmatter: question_count: 20
• Comprehension questions test the paper's contribution, not trivia
• All answers derivable from paper or study material — no external knowledge required
• Numbering is sequential across sections (1-8 in A, 9-13 in B, 14-20 in C)

<slug>-podcast-script.md:

• Two distinct hosts (A presents, B asks critically)
• Specific numbers, figure references, and quotes from results
• Limitations explicit in Segment 4

<slug>-notebooklm-prompt.txt:

• All <PLACEHOLDER> fields filled in
• <FOCUS_HINT> is 1-2 concrete sentences (not "this paper studies X")

What This Skill Does NOT Do

• Does NOT run NotebookLM — user does that manually on fenrir
• Does NOT generate the m4a audio
• Does NOT run Claude voice / speaking practice — separate session on user's side
• Does NOT update the Podcast (m4a) link with a Drive URL automatically — requires user to provide the URL after NotebookLM finishes
• Does NOT create new paper close-readings — <slug>-ja.md must already exist

Repository-Specific Convention For life

• Outputs live in portfolio/paper-close-readings/ (co-located with paper-close-reading artifacts)
• Drive staging path: ~/Library/CloudStorage/GoogleDrive-38kta.lab@gmail.com/マイドライブ/paper-podcasts/<slug>/
• Per-paper Drive folder collects: NotebookLM inputs (HTML + study + prompt) and the eventual m4a
• Audience for podcast prompt: biology researchers (broad) — wider than user's plant physiology focus so NotebookLM doesn't drift into one subfield

Auto-finalize

After producing the 5 files, run the shared finalize script:

bash scripts/agent_auto_finalize.sh \
  -m "docs: 📚 english-from-paper: <paper short title>" \
  portfolio/paper-close-readings/<slug>-en-study.md \
  portfolio/paper-close-readings/<slug>-en-quiz.md \
  portfolio/paper-close-readings/<slug>-en-quiz-ans.md \
  portfolio/paper-close-readings/<slug>-podcast-script.md \
  portfolio/paper-close-readings/<slug>-notebooklm-prompt.txt \
  portfolio/paper-close-readings/<slug>.md

AGENT_AUTO_COMMIT=1 on fenrir auto-commits; on Air the call is a no-op.