daily-search-trend - SKILL.md Agent Skill

name: daily-search-trend description: "Create Japanese daily search trend reports from PubMed, Europe PMC/bioRxiv preprint searches, Nature News, Science/AAAS News, and ナゾロジー（自然科学） for the previous calendar day, then render the Markdown to a Newsprint-themed HTML file. Use when the user asks for daily-search-trend, search trend, trend.md, trend.html, 今日の研究トレンド, 論文検索, 新着論文, 研究ニュース, 前日の論文, ideas/dailyへのtrend作成, or wants titles translated to Japanese and ranked strictly from portfolio context."

Daily Search Trend

Create a Markdown trend report in the current repository from recent papers and science/news items, then render the Markdown to standalone HTML.

Repo Context

Before running the trend:

Read README.md for the output directory policy.
Read Rules.md for safety, information-quality rules, and repo-specific digest constraints.
Read portfolio/ files. Use only portfolio/ to infer the user's research interests and score items.
If portfolio/ is missing or too sparse to judge interest, state that limitation in the output and use ★★★☆☆ as a neutral provisional score. Do not infer interest from memo.md, README.md, Rules.md, or requested keywords.

For the life repo, write the editable Markdown source under ideas/daily/md/ (git tracked) and the final HTML under /Users/kta/.local/share/life/_life/daily/ (non-git, served by fenrir caddy at /daily/). See GitHub issue #76 for the MD-source / HTML-artifact split.

Search Window

Use the previous calendar day as the default time window. If the skill is run on YYYY-MM-DD, collect items dated YYYY-MM-DD - 1 day.

Use day-level filters instead of strict 24-hour filters. For the life repo, name the output files after the target day:

Markdown source: ideas/daily/md/YYYY-MM-DD-trend.md (git)
HTML final output: /Users/kta/.local/share/life/_life/daily/YYYY-MM-DD-trend.html (non-git, served at http://fenrir:8080/daily/YYYY-MM-DD-trend.html)

Always use current dates from the runtime environment. Do not assume the model's knowledge is current.

Sources

Read references/sources.md when choosing sources or query tactics. Read references/output-template.md before writing the final Markdown.

Use the API helper first for paper/preprint search:

python3 <skill-dir>/scripts/fetch_papers.py --keywords "cyanobacteria,photosystem II,carbon fixation" --target-date YYYY-MM-DD

The helper queries PubMed through NCBI E-utilities and bioRxiv-like preprints through Europe PMC SRC:PPR. Read references/sources.md for API details and the selected news sources.

For Nature News, Science/AAAS News, and ナゾロジー RSS feeds, use the dedicated Python helper instead of shell curl. This is required so the skill works from non-interactive contexts (e.g., iPhone Claude Code) where per-curl permission prompts cannot be approved:

python3 <skill-dir>/scripts/fetch_news.py --target-date YYYY-MM-DD

Output is a single JSON manifest on stdout with one entry per RSS item filtered to the target day. Sources are configurable via --sources nature,science,nazology (default: all three). The helper handles browser-like User-Agent and includes a fallback IP path for environments where local DNS sinkholes nature.com / science.org. Do not invoke curl directly for these sources.

After writing Markdown, render HTML with:

python3 <skill-dir>/scripts/render_trend_html.py ideas/daily/md/YYYY-MM-DD-trend.md -o /Users/kta/.local/share/life/_life/daily/YYYY-MM-DD-trend.html

The renderer embeds assets/newsprint-trend.css, a Newsprint-inspired theme based on Typora's Newsprint theme.

IMPORTANT — /Users/kta/.local/share/life/_life/ write rule (life#77 follow-up):

claude binary は LaunchDaemon context で macOS TCC により /Users/kta/.local/share/life/_life/ への直接書き込みが拒否される (User=kta でも process binary 単位で TCC 判定、bash/python は FDA grant 済だが claude は version-dir based で grant が update に耐えない)。

NEVER use the Write tool to write a file under /Users/kta/.local/share/life/_life/... directly. 失敗するか、または silent fallback で別場所に書く事故が起きる。
/Users/kta/.local/share/life/_life/ への書き込みは 必ず python3 <render-script> を bash 経由で呼び出す (python3 は FDA grant 済)。本 skill では render_trend_html.py と refresh_daily_index.py がそれ。
MD は repo (ideas/daily/md/、/Users/kta/ 配下) に Write tool で書いて OK (TCC 非対象)。

Workflow

Determine paper/preprint keywords from the user request. If none are given, use the default research themes in references/sources.md.
Compute the exact target day from the runtime date or user-specified date.
Search PubMed by keyword with NCBI E-utilities, pinned to the exact target day with EDAT.
Search bioRxiv-relevant preprints by keyword through Europe PMC, using SRC:PPR and FIRST_IDATE pinned to the same exact target day, then filter toward bioRxiv when metadata allows.
Check Nature News, Science/AAAS News, and ナゾロジー（自然科学） by running scripts/fetch_news.py --target-date YYYY-MM-DD. The helper parses the official RSS feeds and filters to the target day. Do not invoke curl directly for these sources — use this script so the workflow runs without permission prompts on remote/headless contexts.
Do not check research-institution announcements by default. Check them only when the user explicitly asks for 研究機関発表 or names institutions and provides or implies a date range such as 1か月前まで.
- For RIKEN, build the year-specific press URL from the execution date or requested range, such as https://www.riken.jp/press/YYYY/index.html. If the requested range crosses years, check all relevant yearly pages.
Deduplicate by DOI, title, URL, and source.
Translate paper, preprint, and science-news titles into Japanese. For 研究機関発表, translate only titles that are not already Japanese; leave タイトル訳 blank for Japanese titles. Do not summarize content for this trend file.
Do not exclude paper/preprint or target-day news records only because they seem weakly related to the user's portfolio. Keep low-relevance records and mark them with a low interest score.
Score interest from 0 to 5 stars using only the portfolio/ context:
- ★★★★★: directly relevant to the user's core research or likely actionable soon
- ★★★★☆: strongly adjacent or useful for research planning
- ★★★☆☆: generally relevant background or trend
- ★★☆☆☆: weakly adjacent, mostly FYI
- ★☆☆☆☆: low relevance, included only for completeness
- ☆☆☆☆☆: no clear relevance to the portfolio, but retained to avoid hidden filtering
For papers and preprints, set カテゴリ to matched search-keyword tags such as #cyanobacteria #photosynthesis instead of generic 論文 or プレプリント.
For English science/news and non-Japanese research-institution announcement items, use 原文タイトル, タイトル訳, and 興味度. For Japanese science/news and Japanese research-institution announcement items, omit タイトル訳 and use only 原文タイトル and 興味度.
Write one Markdown file to ideas/daily/md/YYYY-MM-DD-trend.md, using the target day in the filename, unless the user specifies another topic slug.
Render /Users/kta/.local/share/life/_life/daily/YYYY-MM-DD-trend.html from that Markdown with scripts/render_trend_html.py. Treat the HTML as the final user-facing output and the Markdown as the source.
Include source URLs, target date, keywords, and portfolio files used.

Exact-Day Guardrail

Do not finalize the PubMed or Europe PMC sections from a rolling 24h window alone.

For PubMed, the final adoption step must use YYYY/MM/DD[EDAT] : YYYY/MM/DD[EDAT].
For Europe PMC preprints, the final adoption step must use FIRST_IDATE:[YYYY-MM-DD TO YYYY-MM-DD] for the same target day.
If a helper was first run with --hours 24 for rough collection, rerun or reconfirm with --target-date YYYY-MM-DD before writing the Markdown.
When a record's displayed pubdate looks different from the target day but the entry date matches the target day, it may still be retained if the exact-day query returned it.

Output Rules

Write repository notes and user-facing Markdown in Japanese unless the user asks otherwise.

Keep claims conservative. Do not overgeneralize from one paper, preprint, press release, or news post.

Separate papers/preprints and science/news site items by source.

Use this table shape for papers and preprints:

| 原文タイトル | タイトル訳 | 興味度 | カテゴリ |
|---|---|---|---|

Use this table shape for English science/news and non-Japanese research-institution announcement sources:

| 原文タイトル | タイトル訳 | 興味度 |
|---|---|---|

Use this table shape for Japanese science/news and Japanese research-institution announcement sources:

| 原文タイトル | 興味度 |
|---|---|

Put the source URL into 原文タイトル as a Markdown link: [Original Title](https://...).

For Nature News, use https://www.nature.com/nature.rss as the primary source and filter entries by dc:date. Use https://www.nature.com/news only as a human browser reference. If the RSS feed fails, fall back to a domain-constrained web search for site:nature.com/articles/d41586 plus the target date.

For Science/AAAS News, use https://www.science.org/rss/news_current.xml as the primary source and filter entries by dc:date or prism:coverDate. Use https://www.science.org/news only as a human browser reference. If the RSS feed fails, fall back to a domain-constrained web search for site:science.org/content/article plus the target date.

For 研究機関発表, leave タイトル訳 blank when the original title is already Japanese. Fill タイトル訳 only when the original title is not Japanese.

Do not include content summaries in YYYY-MM-DD-trend.md. Save deeper summaries for a separate future daily-research-digest skill.

Do not include empty source sections. If Nature News, Science/AAAS News, ナゾロジー（自然科学）, or a requested institution has no target-day results, omit that source section from the output.

Do not add 興味度の判断メモ or 制限・不確実性 sections to the daily trend output.

Always generate or update the HTML file after changing the Markdown trend file.

Refresh Portal Index

After rendering the HTML, regenerate the personal portal's daily index so the fenrir landing page (http://fenrir:8080/) reflects the new entry:

python3 scripts/refresh_daily_index.py

This rewrites /Users/kta/.local/share/life/_life/daily/index.json from the actual file listing there. It is fast (~ms) and idempotent. The index.json is non-git (lives next to the HTML artifacts), so do NOT include it in the auto-finalize call below.

Auto-finalize

After producing both the Markdown and HTML, run the shared finalize script. It is a no-op unless AGENT_AUTO_COMMIT=1 is exported in the shell. On fenrir this is the default; on Air / mini-lab it is unset, so this call has no effect.

bash scripts/agent_auto_finalize.sh \
  -m "docs: 📝 daily-search-trend: YYYY-MM-DD" \
  ideas/daily/md/YYYY-MM-DD-trend.md

Replace YYYY-MM-DD with the target day. Only the trend Markdown is committed — the HTML and index.json live under /Users/kta/.local/share/life/_life/daily/ (non-git, #76 Phase 3).