Find New Slides

Find Japanese NLP presentation slides about "$ARGUMENTS" that are not already in the awesome-japanese-nlp-slides list.

Instructions

Preamble — Establish the current date

Before anything else, run this once and remember the values — every step that mentions a year or today's date refers to them:

echo "TODAY=$(date +%F)"
echo "YEAR_NOW=$(date +%Y)"
echo "YEAR_PREV=$(($(date +%Y) - 1))"

Substitute these wherever this skill writes ${YEAR_NOW} or ${YEAR_PREV}. Do not hardcode years — the value of this list is that it stays current.

Step 0 — Handle empty input

If $ARGUMENTS is empty or blank, treat it as a sweep for the most recent Japanese NLP slides across every topic:

• Topic label for the output heading: "最近公開された日本語NLPスライド" (Japanese) / "Recent Japanese NLP slides" (English)
• Search queries for Step 4: drop the topic term and lean entirely on recency —
  • 自然言語処理 発表資料 ${YEAR_NOW} site:speakerdeck.com
  • LLM 勉強会 資料 ${YEAR_NOW} site:speakerdeck.com
  • 日本語 LLM スライド ${YEAR_NOW}
  • NLP 論文読み会 資料 ${YEAR_NOW} site:speakerdeck.com
  • 言語処理学会 ${YEAR_NOW} 発表資料
  • 自然言語処理 ${YEAR_NOW} site:docswell.com
• Target section: none — assign each survivor its own section in Step 7.

Then continue from Step 1 with those defaults.

Step 1 — Interpret the topic and pick the target section

The user's topic is: "$ARGUMENTS"

The material is Japanese slide decks, so the search queries have to be Japanese. An English-only query set finds almost nothing: the titles, the slide hosts' metadata and the surrounding blog posts are all Japanese. Translate the topic into the Japanese term first, then optionally add one English query.

The table below is the list's own 32 sections. Use it twice: pick the search phrases here in Step 1, and reuse the same row's cues to assign a section in Step 7. Pick 1–3 rows — the one the topic names plus any obvious neighbour.

If the topic matches no row, translate it literally into Japanese and append 資料 / スライド / 発表 to build the queries, then assign the closest section in Step 7.

Step 2 — Locate the existing data file

Same resolution as the search skill:

SLIDES_PATH="${CLAUDE_PLUGIN_ROOT}/data/slides.json"
[ -f "$SLIDES_PATH" ] || SLIDES_PATH="$(find "${HOME}/.claude/plugins" -type f -name slides.json 2>/dev/null | grep "awesome-japanese-nlp-slides/" | head -1)"
echo "SLIDES_PATH=$SLIDES_PATH"

Step 3 — Build the existing-URL set

The plugin's slides.json is generated from data/curated.json, so an installed copy can lag the repository by a few commits. When the run happens inside a checkout of the repo, merge curated.json in as well so freshly added decks are not reported as new.

EXISTING_URLS_FILE=$(mktemp -t awesome_ja_nlp_slides_urls.XXXXXX)

python3 << 'EOF'
import json
import os
import re

SLIDES_PATH = "SLIDES_PATH"           # from Step 2
OUTPUT_PATH = "EXISTING_URLS_FILE"    # from the mktemp above


def canon(url):
    """Normalise a slide URL the way Step 5 compares them."""
    url = (url or "").strip().lower()
    url = url.split("?")[0].split("#")[0].rstrip("/")
    url = re.sub(r"^http://", "https://", url)
    url = re.sub(r"^https://www\.speakerdeck\.com", "https://speakerdeck.com", url)
    url = re.sub(r"^https://(docswell|slideshare)\.", r"https://www.\1.", url)
    return url


urls = set()
with open(SLIDES_PATH) as f:
    for item in json.load(f):
        urls.add(canon(item["u"]))
count_plugin = len(urls)

# Walk up looking for the repo's data/curated.json (present only in a checkout).
path = os.path.abspath(os.path.dirname(SLIDES_PATH))
for _ in range(6):
    path = os.path.dirname(path)
    curated = os.path.join(path, "data", "curated.json")
    if os.path.exists(curated):
        with open(curated) as f:
            for section in json.load(f):
                for entry in section["entries"]:
                    urls.add(canon(entry["url"]))
        print(f"Supplemented {len(urls) - count_plugin} URLs from {curated}")
        break

with open(OUTPUT_PATH, "w") as f:
    f.write("\n".join(sorted(urls)))
print(f"Loaded {len(urls)} existing slide URLs ({count_plugin} from the plugin) -> {OUTPUT_PATH}")
EOF

Clean the temp file up when the skill finishes (rm -f "$EXISTING_URLS_FILE").

Step 4 — Discover candidates via WebSearch

Run 6–9 WebSearch queries built from the phrases picked in Step 1. Rules:

1. At least two thirds of the queries must be Japanese. This is the single biggest factor in whether anything is found.

2. Cover all three hosts, one site: per query. Speaker Deck holds most of the list, but Docswell and SlideShare are where a lot of corporate and academic material lives.

   • <japanese-phrase> site:speakerdeck.com
   • <japanese-phrase> site:docswell.com
   • <japanese-phrase> site:slideshare.net

   Never OR the hosts into one query (... docswell OR speakerdeck OR slideshare). Without the site: operator the engine reads the host names as topic words and returns articles about slide-sharing services instead of slides. One host per query, site: every time.

3. Add recency. Two or three queries should carry ${YEAR_NOW} or ${YEAR_PREV} — the list is weakest at its recent edge.

4. Search around the events too, since slides are announced in wrap-up blog posts before search engines index the deck itself:

   • <japanese-phrase> 勉強会 資料 ${YEAR_NOW}
   • <japanese-phrase> 登壇 資料
   • 言語処理学会 ${YEAR_NOW} <japanese-phrase> 発表資料

5. One English query is worth including for research-flavoured topics: <english-keyword> japanese site:speakerdeck.com.

Keep a narrow topic undiluted — this is the rule that decides whether a rare topic finds anything. Padding a specific term with generic ones (自然言語処理, NLP, LLM, 資料) lets the engine satisfy the query on the padding alone and silently drop the term you care about. Measured on the same topic:

So for a narrow topic, query the term by itself plus site:, and only then widen. The tell that the term was dropped is a result set whose titles do not contain it anywhere — when you see that, re-run with the bare term before concluding anything. The generic phrases in the Step 1 table are for broad topics that need narrowing, not for narrow ones that need protecting.

From every result — the result URLs and the URLs quoted inside wrap-up pages — extract anything matching a slide-deck shape:

Ignore user/profile pages (https://speakerdeck.com/<user> with no slug), player/embed URLs, PDF mirrors, YouTube, Qiita, Zenn and note.com — this list indexes slide decks on those three hosts only. Normalise each URL exactly as canon() in Step 3 does, and collect them into a candidate set tagged with its source.

Step 5 — Filter against the existing list

Drop every candidate whose normalised URL appears in $EXISTING_URLS_FILE, then deduplicate the remainder.

Then drop everything that is not actually about the topic. This is the filter that matters most, because of how the search engine behaves on a narrow term: asked for 手話 自然言語処理 資料 site:speakerdeck.com it quietly discards 手話 and returns ten popular general-NLP decks — real, well-known, genuinely NLP-related, and none of them about sign language. Half of them were already in this list; the other half would sail through every other check in this skill and be reported as "new 手話 slides".

So judge each candidate's title against the topic and drop the ones that only share the general field. A deck earns its place by covering the topic, not by being a good deck.

Report the count you dropped as "already listed: N". If everything was dropped, the topic is already well covered; if the survivors are all off-topic, the honest answer is that nothing was found. Say so in Step 8 — a short "nothing found" is worth more than a list of decks the reader has to re-check by hand, and it is the correct answer for a topic the Japanese NLP community has not presented on.

If more than 20 candidates survive, prioritise those that appeared in several result sets and those whose title clearly matches the topic.

Step 6 — Enrich the candidates via WebFetch

curated.json requires five fields per entry (title, url, author, date, source) and rejects a missing or malformed one, so every candidate has to be confirmed against its own page — never guess a date from a search snippet. Set the sixth, optional field, added, to $TODAY (from the Preamble) on every candidate this skill outputs — that is what puts it in the README's "🎉 latest additions" section for the 7 days after it lands in curated.json.

Cap at 10–15 candidates and issue up to 5 WebFetch calls in parallel (one message, multiple tool calls):

WebFetch url="<candidate url>" prompt="Extract as JSON: title (the deck's title exactly as displayed, in its original language), author (the presenter's display name exactly as shown on the page), published_date in YYYY-MM-DD form, slide_count, and topic_summary (one sentence, English). Also return is_nlp_related: true only if the deck is about natural language processing, LLMs, speech/text processing or search; deck_kind, one of technical_talk / lecture_or_reading / research_poster / company_or_event / personal_notes / unknown, judged from the title and the page's own description; and self_disclaimer: true if that description says the deck is a memo with no findings, an unverified AI-generated summary, or that the attempt did not work out. If any field is unavailable, set it to null."

Then drop a candidate when:

• the page 404s, redirects to a profile, or is private;
• is_nlp_related is false — a keyword can match a deck that only mentions NLP in passing;
• topic_summary shows the deck is not really about the topic, even though the title looked close. This is the second chance to catch what Step 5 is guarding against: is_nlp_related is true for most of this list, so it is nearly useless as a relevance test on its own;
• the deck is not in Japanese and not from the Japanese-speaking community. The list's scope is NLP/LLM material presented to a Japanese audience: a general LLM survey given in Japanese belongs, an English-only deck by an unrelated author does not;
• published_date is null. Put these in the "要確認" bucket in Step 8 instead of discarding them silently — the date usually just needs a human to look at the page.

Step 6b — Apply the quality bar

This list is curated: a deck earns a place by having something to say, not by existing. contributing.md holds the full bar; these are the parts a page's own metadata can decide.

Drop a candidate when:

• deck_kind is company_or_event — a sponsor-slot company introduction, a recruiting pitch, an event's opening deck. The tell is a title like <社名> Tech Meetup #8 or …勉強会_スポンサー公開資料, or a description that promotes the company rather than the talk.
• deck_kind is personal_notes — a 備忘録 of links and impressions, a competition write-up that only reports a placing.
• self_disclaimer is true — the deck says of itself that it has no useful findings, that it was generated without checking primary sources, or that the attempt simply failed. A write-up of something that did not work is welcome when it says why; the ones to drop stop at "試したがだめでした".
• the title is a bare fragment of a book or a seminar series — トピックモデル 3.2.3 ~ 3.2.4, B3勉強会 テキストマイニング. A self-contained chapter (【大規模言語モデル入門】1章) is fine; two subsections of equations with no surrounding argument are not.
• the subject is not language processing itself — AI-assisted coding productivity, a tool's setup steps.

Never use slide_count as the quality test. It is reported for the human reading the output, not for this filter. A research_poster is often a single page carrying an entire study, and a seven-slide lightning talk that derives an equation or reports a measurement is worth more than sixty slides of restated background. Judge deck_kind, topic_summary and the description; when they are not enough to tell, say so in Step 8's "要確認" bucket rather than guessing in either direction.

Normalise the two fields that scripts/slides.py cleans up on load, so the output can be pasted verbatim:

• title — collapse runs of whitespace, and strip a trailing .pdf / .pptx / .key if the presenter left one in.
• author — the platform display name as-is (Preferred Networks, Naoaki Okazaki, himkt). Do not romanise it, translate it, or replace it with an account ID.
• date — YYYY-MM-DD, the date the deck was published, not the date of the event.

Step 7 — Assign a section

Give every survivor exactly one of the 32 section names from the Step 1 table, using that row's cues. Rules:

• The name must match curated.json character for character — the generators key their translations off it, and a typo silently creates a duplicate section.
• Prefer the specific section over the generic one: a deck about running a pretraining job goes to 日本語LLMの開発（事前学習・コーパス構築）, not 大規模言語モデル（概論・動向）.
• A company case study whose substance is one technique goes to that technique's section; 産業応用・実務事例 is for decks whose subject is the rollout.
• A university lecture series goes to 大学講義・体系的な講義資料 whatever it covers.

Step 8 — Output

Language detection: if $ARGUMENTS contains any Japanese character, or is empty, write the prose in Japanese; otherwise in English. Titles, author names and section names are always quoted exactly as stored — never translated.

Lead with the paste-ready JSON, grouped by section. These are entry objects to append to the entries array of an existing section in data/curated.json — not new section objects.

## "$ARGUMENTS" の追加候補

未収録のスライド **N 件** を見つけました（検索した候補 M 件のうち、収録済み K 件・基準を満たさないもの Q 件を除外）。

*(検索クエリ: query1, query2, ...)*

### `検索・RAG` に追加

```json
{
  "title": "スライドのタイトル",
  "url": "https://speakerdeck.com/user/slug",
  "author": "発表者名",
  "date": "2025-06-14",
  "source": "speakerdeck",
  "added": "$TODAY"
},
```

- [スライドのタイトル](https://speakerdeck.com/user/slug) — 発表者名（2025-06）・全 42 枚 / どんな内容かの一文

### `評価・ベンチマーク` に追加

...

---

### 要確認

- [タイトル](url) — 発表者名 / ページを開いて公開日を確認してください
- [タイトル](url) — 発表者名（2025-06） / 中身が薄い可能性があります。掲載の基準を満たすか確認してください

### 反映手順

1. スライド本体に目を通し、`contributing.md` の「掲載の基準」を満たしているか確認する
2. 上の JSON を `data/curated.json` の該当セクションの `entries` に貼り付ける
3. `python3 scripts/slides.py` で検証する
4. `python3 scripts/generate_readme.py` と `python3 scripts/build_plugin_data.py` を実行する

Rules for the output:

• The JSON block is the deliverable — keep the six keys in the order title, url, author, date, source, added to match the file, and leave the trailing comma so the block splices into an existing array. added is always $TODAY, the day this skill ran, not the deck's own date.

• The bullet under each JSON block is for the human reading the result: presenter, YYYY-MM, slide count when known, and one sentence on the content drawn from the WebFetch summary. Never invent a description you did not read.

• Order the entries within a section newest first, matching how curated.json is kept.

• The "要確認" block takes both kinds of unresolved candidate: the ones missing a published_date, and the ones Step 6b could not clear or reject on the metadata alone. Say which of the two each one is, and never quietly promote a candidate you are unsure about into the JSON block — the JSON is what gets pasted into the file.

• Report the count Step 6b removed alongside the "already listed" count, e.g. 「収録済み K 件・基準を満たさないもの Q 件を除外」, so the reader can see the filter ran.

• Do not print the "要確認" block or the "反映手順" block when they are empty.

• If nothing survived, say so and offer next moves:

  ## "$ARGUMENTS" の追加候補
  
  未収録のスライドは見つかりませんでした（収録済み K 件・基準を満たさないもの Q 件を検索結果から除外）。
  
  - 別のキーワードで再試行: `<提案するクエリ>`
  - 収録済みの資料は `/awesome-japanese-nlp-slides:search $ARGUMENTS` で確認できます
  

Step 9 — Sources

Append the Sources: section that WebSearch requires, listing the result URLs you actually used, then clean up: rm -f "$EXISTING_URLS_FILE"