Daily Briefing

Daily Briefing

Use this skill directly in Codex. Do not require Hermes, Telegram, or a separate chat bot.

Generate Today's Briefing

1. Run the bundled helper to fetch, normalize, deduplicate, and score the source pool:

skills/daily-briefing/scripts/run-daily-briefing.sh \
  --editorial-packet \
  --output-file "outputs/editorial_packet_$(date +%F).json"

Translate explicit user constraints into helper flags without asking them to learn CLI syntax:

• 最近 48 小时模型发布与更新 → --mode models --hours 48
• 最近 7 天 AI 视频产品更新 → --mode products --focus "AI视频" --hours 168
• 最近 24 小时应用层趋势 → --mode applications --hours 24
• 最近 7 天 AI 投融资 → --mode funding --hours 168
• 今日 AI 全产业链简报 → --mode industry --hours 24
• 过去 24 小时医疗 AI 新闻 → --mode hotspots --focus "医疗" --hours 24
• 今日完整早报 → keep the default --mode all; do not narrow the existing source pool.

Append the selected flags to the helper command before generating the editorial packet.

When invoked from an installed skill, use the absolute path to this skill's scripts/run-daily-briefing.sh.

2. Read sections.industry_chain, sections.cross_layer_connections, sections.model_releases, sections.product_updates, sections.application_trends, sections.ai_funding, sections.viral_ai_news, sections.official_social_updates, and sections.editorial_queue. Treat infrastructure, model/platform, product/function, application adoption, capital/commercial events, and potentially viral news as parallel intelligence tracks. Use the current Codex model to select the strongest items, write concrete Chinese explanations, organize sections, and produce varied X drafts. raw_summary contains source material for factual rewriting; do not discard it when the deterministic Chinese hint is generic.

3. When metadata.official_x_monitor.mode is codex_web_search, execute every query in sections.official_x_search_groups for the requested window. The groups are country-aware and put China first. Retain direct https://x.com/<handle>/status/<id> links, and add only posts published by the configured account. Also check sections.official_model_watchlist and sections.official_product_watchlist for vendors and products that do not expose a stable feed.

4. Execute the relevant batches in sections.industry_search_groups, starting with China. Use sections.industry_source_watchlist to return to company newsrooms, customer announcements, investors, exchanges, or regulatory filings. Discovery-only media and community sources may surface a lead, but they cannot confirm funding amounts, adoption figures, revenue, valuation, or infrastructure spending on their own.

5. Before writing, complete a freshness audit: confirm that every relevant X and industry search group was checked or name the failed group, and separately record first-party findings as model release, product/function update, application/adoption, funding/commercial event, ecosystem activity, or potential viral story. An account or page appearing in a watchlist is not evidence that it was searched. Do not finalize a briefing with an empty China official-source result until the CN groups have been checked explicitly.

6. Do not require a model launch for inclusion. Track major product releases, new features, application adoption, customer cases, user growth, pricing, revenue, partnerships, infrastructure, open-source projects, research breakthroughs, financing, policy or safety events, industry disputes, and fast-rising community discussions when they have clear AI relevance and credible links. Treat popularity as a distribution signal, not proof of truth.

7. Classify each item on two separate axes: industry_layer is only upstream, midstream, or downstream; event_types describes what happened, such as model_release, product_update, adoption, funding, acquisition, earnings, or policy. Capital is a cross-cutting event, never a fourth industry layer. A downstream application company remains downstream after raising money.

8. Save the final document as outputs/briefing_YYYY-MM-DD.md in the project directory. Then run the bundled validator and return a clickable local file link only after it passes:

python skills/daily-briefing/scripts/validate-briefing.py \
  "outputs/briefing_$(date +%F).md"

When invoked from an installed skill, use the absolute path to validate-briefing.py. If validation fails, rewrite the named entries and run it again; do not silently return a partial draft.

After validation passes, generate the companion interactive HTML by default:

python skills/daily-briefing/scripts/render-briefing-html.py \
  "outputs/briefing_$(date +%F).md"

The renderer preserves the final Codex-edited Markdown and creates outputs/briefing_YYYY-MM-DD.html. It is a standalone file with local search, section navigation, confidence and industry-layer filters, browser-local bookmarks, Markdown download, print/PDF support, theme switching, and copy buttons for code or X drafts. It must not require a web server or external CDN. Return clickable links to both Markdown and HTML. If the user explicitly asks for only one format, honor that request.

9. Keep every factual claim within the candidate packet or the first-party pages/posts checked above. Preserve original URLs and numbers. Do not invent missing dates, scores, funding amounts, customer counts, revenue, benchmark results, or source agreement.

10. Use clickable Markdown titles. Immediately below every retained news, model release, product update, funding event, official X post, and GitHub project, write a specific one- or two-sentence Chinese explanation before metadata or scores. It must identify who did what and include the most useful result, number, availability change, or impact supported by the source. A reader should understand the event without opening the link. Do not use generic filler such as 值得关注, 可作为行业观察, or 核心看点是工具链还不成熟 as the explanation. Do not add labels such as 中文总结, 摘要, or 一句话介绍 before explanations.

11. Use these sections in order when relevant:

相比昨天的新变化
AI 产业链全景
产业链联动
最新模型发布与更新
产品与工具更新
应用层趋势
可能爆火的 AI 新闻
AI 投融资与商业化
模型公司官方账号动态
今日必须看
适合发 X
B端/商业机会
持续跟踪
X 草稿
来源明细

12. Distinguish source confidence as 官方确认, 多源印证, or 单源信号. Funding and commercial figures marked 待官方复核 must not be written as confirmed. If a source fails, retain successful sections and name the failed source briefly.

13. Make X drafts structurally different: include at least a single post, a thread, and a visual or video script when the candidate pool supports them.

14. End every final Markdown briefing with a concise bilingual support section. Use the heading 支持这个项目 / Support the Project, invite readers in natural Chinese and English to Star the project if the briefing was useful, and include the canonical repository link exactly once: https://github.com/Rion-Wu-tech/ai-daily-briefing. Keep this section brief and place it after source notes so it does not compete with the briefing.

15. Before saving, audit every clickable item in the final Markdown. Reject the draft if any item is followed only by a source, tag, score, confidence label, or link. Also reject repeated template explanations that could describe several unrelated stories. Rewrite those entries from raw_summary and the original source facts. The bundled validator enforces the minimum structural and anti-filler checks; editorial factuality still requires Codex review.