ai-hotspot-digest

Imported from yangzhifeio/ai-research-skills (ai-hotspot-digest/SKILL.md). Install upstream with npx skills add yangzhifeio/ai-research-skills --skill ai-hotspot-digest. Copyright stays with the author.

AI Hotspot Digest

Use this skill when the user wants AI news summaries mainly based on 量子位、机器之心、新智元. It supports one-off summaries and recurring digests.

This skill is optimized for:

• 前一天的 AI 热点日报
• 每周一产出的 AI 周热点
• 每月 1 号产出的 AI 月热点
• 同时覆盖日报、周报、月报的固定节奏

What this skill does

Given a run_date, collect relevant articles from the three target outlets, merge repeated coverage into one event, verify unstable claims where possible, and produce:

• daily: 前一自然日热点
• weekly: 上一个完整自然周热点
• monthly: 上一个完整自然月热点

If the user does not explicitly choose a mode, infer it like this:

• 每天都产出 daily
• 如果 run_date 是周一，再追加 weekly
• 如果 run_date 是每月 1 号，再追加 monthly
• 如果周一恰好也是每月 1 号，则同一次输出中同时包含日报、周报、月报

Unless the user asks for another style, keep the output concise, factual, and link-rich.

When reports are persisted across runs, store them in month-scoped archive documents rather than one ever-growing file. Use one archive file per calendar month.

Default inputs

Use sensible defaults and only ask follow-up questions when the choice would materially change the result.

• sources: 固定为量子位、机器之心、新智元
• run_date: 默认使用当前日期
• timezone: 默认使用用户时区
• output_language: 默认使用用户语言
• daily_count: 默认 10
• weekly_count: 默认 8
• monthly_count: 默认 10
• archive_dir: 默认 /Users/yangzhifei.9/Downloads/codex/research/ai-hotspot-digests
• archive_file_pattern: 默认 ai-hotspot-YYYY-MM.md

Useful optional inputs:

• focus: 例如模型发布、产品动态、融资并购、监管政策、开源项目、芯片与基础设施
• exclude: 明确不想看的方向
• audience: 例如投资视角、产品视角、技术视角、管理层视角

Time window rules

Use absolute dates in the final report.

Daily

• 默认窗口为 run_date 的前一自然日 00:00-23:59
• 例如 run_date = 2026-04-22，则日报窗口为 2026-04-21

Weekly

• 仅在周一默认追加
• 窗口为上一个完整自然周，即上周一到上周日
• 例如 run_date = 2026-04-27 (Monday)，则周报窗口为 2026-04-20 到 2026-04-26

Monthly

• 仅在每月 1 号默认追加
• 窗口为上一个完整自然月
• 例如 run_date = 2026-05-01，则月报窗口为 2026-04-01 到 2026-04-30

If the user gives explicit dates, use those dates instead of the defaults and say so clearly.

Monthly archive documents

When maintaining a persistent archive, use a separate Markdown document for each calendar month:

• File path: {archive_dir}/ai-hotspot-YYYY-MM.md
• Title: # YYYY-MM AI Hotspot Archive
• Month window: YYYY-MM-01 through the actual last day of that month
• Example: ai-hotspot-2026-06.md covers 2026-06-01 through 2026-06-30
• Example: ai-hotspot-2026-05.md covers 2026-05-01 through 2026-05-31

Archive routing rules:

• Daily reports go into the archive file for the daily report date, not the automation run date. For example, a run_date of 2026-06-03 creates the daily window 2026-06-02, so it belongs in ai-hotspot-2026-06.md.
• Weekly reports go into the archive file for the week start date. Keep the exact week window in the heading and metadata.
• Monthly reports go into the archive file for the month being summarized. For example, a monthly report generated on 2026-07-01 for June belongs in ai-hotspot-2026-06.md.
• Do not append a report to a file for a different month just because that file already exists.
• Before writing, check whether the same date or time window already exists in the target monthly file, and update in place instead of duplicating it.
• Keep entries reverse chronological within each archive section. New daily, weekly, or monthly entries must be inserted at the top of their section, immediately after the relevant ## Daily Reports, ## Weekly Reports, or ## Monthly Reports heading.
• Do not append new entries to the bottom of a section or the end of the file.
• If the target monthly file does not exist, create it from the archive wrapper in references/report-template.md.

Research workflow

1. Determine run_date, timezone, and the required windows.
2. Search each target outlet for articles published in the window. Prefer outlet article pages, archive pages, topic pages, or search results on the site itself. If site search is weak, use web search with the outlet name and date hints.
3. Collect candidate articles that describe materially new AI events:
   • foundation model releases or upgrades
   • major product launches
   • open-source releases
   • financing, M&A, partnerships
   • policy, regulation, standards
   • chips, compute, infrastructure
   • benchmark jumps or notable research with clear industry impact
4. Skip low-signal items unless they reveal new facts:
   • pure opinion columns
   • lightly rewritten reposts with no new information
   • repetitive same-day follow-ups that only change wording
5. Cluster articles by underlying event. One event should normally become one hotspot item even if all three outlets covered it.
6. For each hotspot, verify unstable facts when practical using primary sources such as official blogs, company announcements, papers, repos, or government notices. Keep the media reports as discovery sources.
7. Rank hotspots by importance, then write concise summaries that distinguish fact from inference.
8. For weekly and monthly modes, add a synthesis of major themes, not just a longer list.

Deduplication rules

• Merge cross-outlet coverage of the same event into one item.
• Merge follow-up articles into the same item unless the follow-up adds a materially new development.
• If an event evolves across several days but the new update changes the stakes, keep a new item and explain what changed.
• List all outlets that covered the event when known.

Ranking rules

Rank by a mix of these signals:

• impact: 对行业、产品、资本市场或技术路线的影响面
• novelty: 是否是新增事实，而不是旧闻翻炒
• evidence: 信息是否具体、可验证、是否有原始来源
• breadth: 是否被多个目标媒体同时报道
• follow-through: 是否带来可持续跟踪的后续影响

Do not rank purely by outlet excitement or headline style.

Source policy

• Always browse. Do not rely on memory for dates, numbers, or current statuses.
• Every hotspot item must include source links.
• Prefer at least one outlet source for every item.
• When practical, add one primary source link for verification.
• If the three outlets disagree on a detail, say what conflicts and avoid guessing.
• If a story appears in only one outlet but is clearly important and verifiable, it can still be included.

Output requirements

Use the canonical structure in references/report-template.md.

Daily section

The daily report should include:

• the exact date window
• a short top-line summary in 2 to 4 bullets
• Top 10 hotspot items ranked by importance
• a short "watch next" section for items likely to continue

For daily mode, treat 10 as the default target rather than a loose suggestion. Prefer exactly 10 deduplicated hotspot items when enough material exists across the three outlets.

Weekly section

The weekly report should include:

• the exact week window
• 3 to 5 weekly themes
• Top N hotspot items of the week
• one short paragraph on what changed versus the previous week if the signal is visible

Monthly section

The monthly report should include:

• the exact month window
• 3 to 6 monthly themes
• Top N hotspot items of the month
• a short judgment on emerging directions worth tracking next month

Writing style

• Keep summaries compact and high-signal.
• Prefer concrete nouns, dates, model names, company names, and product names.
• Do not inflate weak stories into strategic trends.
• Clearly separate 事实 from 判断.
• If something is unverified, label it explicitly.

Each hotspot item should answer:

• 发生了什么
• 为什么值得关注
• 有哪些关键信号
• 来源

Recurring use

If the user wants automation:

• prefer a thread automation unless the user explicitly wants a separate recurring job
• run on a daily schedule
• let the skill decide whether the current run_date also needs weekly or monthly sections
• keep the automation prompt focused on the task, not the schedule
• if the automation writes to disk, write to the correct ai-hotspot-YYYY-MM.md archive file instead of a single long archive

Quality bar

The report is good only if:

• the date window is explicit and correct
• duplicate stories were merged into one event
• weak stories did not crowd out major developments
• links are present for each item
• weekly and monthly sections synthesize patterns instead of repeating the daily list
• persisted reports are stored in the correct monthly archive file
• the tone stays analytical rather than sensational

Final checklist

Before returning:

• verify the actual date window used
• verify the outlet/source attribution
• verify major claims that could have changed quickly
• ensure the ranking reflects importance, not headline wording
• ensure the output is concise enough to skim
• if persisting, verify the target archive file month and avoid duplicate entries