industry-daily-brief-x-3

原始内容

---
name: data-ai-daily-brief
version: "5.0.0"
description: >
  Turn any industry into a daily intelligence briefing. An AI agent searches,
  filters, writes, and delivers structured daily briefs to 9 channels — with
  machine-checked formatting and a business review gate. Ships with a Data+AI
  profile out of the box; switch to any domain via config.
read_when:
  - daily brief
  - industry report
  - industry newsletter
  - intelligence brief
  - 日报
  - 行业日报
  - 情报简报
allowed-tools:
  - read_file
  - write_to_file
  - replace_in_file
  - execute_command
  - web_search
  - web_fetch
disable: false
---

# Industry Daily Brief

An AI-driven skill that generates a high-quality industry intelligence brief: it automatically searches, filters, writes, and delivers a structured daily report. It ships with a **Data+AI profile** as the working example, and can be switched to **any industry** through the configuration file.

> **How to read this document**
> The **Workflow**, **Confidence Tiers**, **Section Definitions**, **Format Contract**, and **Hard Rules** below are **industry-agnostic** — they are the engine. Everything inside a block marked **`[Default Profile: Data+AI]`** is an **example configuration** (vendor lists, search queries, focus areas) that you replace when targeting another domain. Do not treat the Data+AI specifics as part of the framework.

## Workflow

When the user requests a daily brief, execute the following steps in order.

### Step 1: Confirm Configuration [Deterministic]

1. Read the workspace `config.json` (if present).
2. If absent, initialize defaults via `scripts/init_config.py`.
3. Confirm the target date (default: today) and the output channels.

### Step 2: Collect & Filter Information [Deterministic + LLM]

Use `web_search` to gather information, applying the following priorities and filters.

#### Core Principles

**Relevance first, filter ruthlessly.** Every item must clearly answer: *does this affect the product roadmap, architecture, cost structure, governance, operational efficiency, or real-world adoption within the target industry?* If the answer is not a clear **yes**, exclude it.

**Less is more.** Never lower the admission bar just because a section has few items. The value of the brief is precision, not item count.

#### Three-Phase Search Strategy

**Phase 1 — Targeted first-hand source search (mandatory).**
For each Tier-1 vendor in the active profile, search its official channels (site blog, release notes, GitHub releases, press wires) one by one. Also run dedicated funding / M&A / earnings queries for the tracked companies.

**Phase 2 — Expanded discovery (supplementary coverage).**
Broaden with topic keyword searches across the profile's focus areas and the date range, to catch items the targeted search missed.

**Phase 3 — Source tracing (mandatory).**
For any item discovered via secondary media, use `web_fetch` or an additional `site:` search to trace it back to a first-hand source. Items with no traceable first-hand source are flagged **⚠️ unverified** or demoted to the Watchlist.

**Coverage requirement:** every Tier-1 vendor in the active profile must receive at least one targeted search.

#### Recency Window (red line)

**Weekdays (Tue–Fri):** strictly cover only information **first published within the last 24 hours** (08:00 the previous day → 08:00 today, target timezone).

**Monday special rule:** the window expands to **72 hours** (Friday 08:00 → Monday 08:00), covering Fri–Sun. Monday item cap is raised, and the title is marked as covering the weekend.

⚠️ **Recency red line — never admit any of the following:**
- Information whose original publish date falls outside the current window
- Information that appeared in a previous brief
- Stale announcements from days or weeks ago
- Pre-announced schedules (conferences/summits) that are not a *today* first disclosure

✅ **How to verify recency:**
1. Check the original page's publish date.
2. If it falls outside the window → exclude.
3. Before writing, list each candidate's publish date and confirm item by item.

> **Stale-news guard:** a "big news" item surfacing in search results must have its *original* publish date independently confirmed before admission — search ranking is not recency.

#### `[Default Profile: Data+AI]` — Focus, Vendors, Sources

> Replace this entire block when targeting another industry.

**Focus areas:** big data, data platforms, data infrastructure, data governance, data engineering, lakehouse architecture, query engines, stream/batch processing, vector search infrastructure, open-source data ecosystem. AI items are admitted **only when they clearly affect the data platform**.

**Strictly exclude:** pure-AI news (model releases, benchmarks, consumer AI), AI items with no direct data-platform link, secondhand financial/mass-media analysis, content farms / clickbait / unsourced rewrites.

**Tier-1 vendors:** AWS, Google Cloud, Microsoft Azure, Databricks, Snowflake, Alibaba Cloud, Tencent Cloud, Huawei Cloud, Volcengine.
**Tier-2 vendors:** Confluent, MongoDB, Elastic, ClickHouse, Cloudera, Starburst/Trino, dbt Labs, Fivetran, Airbyte, Dataiku, Palantir, Baidu AI Cloud, JD Cloud.
**Chip vendors (only if directly data-platform related):** NVIDIA, Intel, AMD.

**Open-source projects:** Iceberg, Hudi, Paimon, Delta Lake, Trino, Spark, Flink, Ray, Airflow, Kafka, dbt, ClickHouse, DuckDB, Milvus, Weaviate, Lance/LanceDB, StarRocks, Doris, SeaTunnel, Amoro.

**Funding/IR sources:** SiliconANGLE Big Data, DBTA, InfoQ, PR Newswire, Business Wire, SEC EDGAR, company IR pages, Crunchbase News, CB Insights, PitchBook News, TechCrunch Venture.

**Analyst firms:** Gartner, Forrester, IDC, a16z, Sequoia, Bessemer, Futurum Group, Constellation Research, Wikibon/SiliconANGLE Research; plus regional research institutes and policy/standards bodies relevant to the market.

**Source rules:** accept only first-hand sources (official sites/blogs/release notes, GitHub repos, original posts on X/LinkedIn/blogs, earnings-call transcripts, analyst reports, PR Newswire/Business Wire). Do **not** accept secondhand financial-media analysis (analyst reports excepted).

### Step 3: Write the Brief [LLM]

Produce a professional brief for practitioners in the target industry, based only on the filtered items.

#### Confidence Tiers

- **Level A — confirmed fact:** first-hand source (official site/blog, GitHub release, earnings filing, live event). → may go into A / B / C / D.
- **Level B — high-trust secondary:** reliable media (Reuters, Bloomberg, TechCrunch) but no first-hand doc yet. → may cautiously go into A / C with a "media report / no official filing" label; prefer the Watchlist.
- **Level C — indirect signal / unverified rumor:** social leaks, community chatter, unmerged-PR speculation. → Watchlist only.

#### Deduplication

Each item belongs to exactly one primary section. Priority: A > B > C > D > E. **Mandatory self-check:** after writing all sections, list every event/source/product name and check for cross-section duplicates. If one event appears in two or more sections, keep the highest-priority section and delete or reduce the rest to a one-line reference (≤15 chars).

#### Title & Opening

Title format: `# <Brand> Daily Brief | YYYY-MM-DD` (Monday marked as covering the weekend).

**Fixed opening structure:**
```markdown
**Top 3 takeaways:**
1. [a single trend judgment — direction not full event, 15–30 chars]
2. [a single trend judgment — direction not full event, 15–30 chars]
3. [a single trend judgment — direction not full event, 15–30 chars]

**Verdict:** [the single most important industry judgment of the day, 1–2 sentences, ≤120 chars, landing on direction / investment focus / market shift]
```

**Key distinction:** the "Top 3 takeaways" are *not* a summary of Top Signals nor three event headlines. They are three cross-event "directions worth taking away today."

**Hard constraints:**
- Each takeaway 15–30 chars; no full product names, version numbers, or specific figures; no bold.
- A good takeaway lets the reader grasp *what direction changed* and *so what*.

**Verdict constraints:** ≤120 chars; a directional judgment, not a repeat of the takeaways or of Section A event details.

**Self-check:** after writing the 3 takeaways, cover Section A and read only the takeaways. If the reader can reconstruct each Section A headline and key figure from them, the takeaways read too much like a summary — rewrite.

#### Sections

The five sections use **English titles** and are mandatory (a section may be empty but its header stays).

**A. Top Signals (3 items).** The day's most important events; first-hand source required.
```markdown
### 1. Event title
**来源：** [specific source](url)
**摘要：** 2–3 sentences
**为什么对数据平台重要：** ...
> 企微摘要：one-line semantic compression
```

**B. Product & Tech (0–6 items, less-is-more).** Strictly product & tech moves in the target industry. Each item: title, source, summary (1–2 sentences), impact judgment, WeChat summary.

**C. Views & Research (0–5 items).** Two kinds of high-value content: original views from key people (founders/CEOs/CTOs in interviews, talks, blogs, X, LinkedIn) and formal research from high-credibility institutions (Gartner/Forrester/IDC/Omdia/etc., regional research bodies, top brokerages), plus officially-released policy & standards relevant to the industry. Each item: name, source, core view, mapping-to-industry judgment, WeChat summary.

**D. Capital & Corporate (0–4 items, less-is-more).** Capital/company events directly relevant to the industry, with inline type tags: **【Funding】 / 【Earnings】 / 【IPO】 / 【M&A】**.
```markdown
### 1. 【Funding】Event title
**来源：** [specific source](url)
**核心数据：** amount / valuation / revenue / growth
**摘要：** 2–3 sentences
**对数据平台的影响：** ...
> 企微摘要：one-line semantic compression
```

**E. Watchlist (1–3 items).** Three kinds: **【Preview】** upcoming events, **【Demoted】** valuable items not meeting A–D bars, **【Tracking】** follow-ups whose impact is still unverified. Each item: title, source, why it's worth watching, what signal to wait for, WeChat summary.

#### WeChat-Summary Field (all sections)

After all detailed fields, every item must carry one line: `> 企微摘要：one-sentence semantic compression`.

Rules:
- Semantic compression of the whole item, **≤120 chars**.
- A standalone, self-contained complete sentence.
- No links, no source labels.
- **Markdown only** — never rendered in the HTML output.

> The `**来源：**` and `> 企微摘要：` field markers are kept in their original Chinese form because the WeChat-push and HTML-conversion scripts (`scripts/`) parse these exact strings. When localizing the brief to another language, keep these two markers as-is or update the scripts in lockstep.

#### Output Requirements

- **Rank by importance, differentiate by channel.** Wider search means more candidates — rank strictly by real industry impact > source authority > topic heat; never lower the bar because more sources appeared.
- **WeChat concise version:** keep section caps (A≤3, B≤6, C≤5, D≤4, E≤3); pick only the most important items.
- **HTML full version:** sections may relax by 2–3 items (A still ≤3, B≤8, C≤7, D≤6, E≤5).
- Professional, concise, restrained tone. Every item must have source, summary, and impact judgment. Total 10–14 items (Monday 14–20). Never fabricate data.

### Step 4: Generate Output Files [Deterministic + LLM]

1. **Markdown** `<Brand>_Daily_Brief_{date}.md`:
   - Every item carries a `> 企微摘要：...` line.
   - Contains all items (incl. HTML-extended ones); the WeChat-summary line marks which enter the concise push.
   - **Source links must use Markdown link syntax `[text](url)`** — never bare text `（https://...）`. Bare URLs render as plain text in HTML and break clickable links.
2. **HTML** `<Brand>_Daily_Brief_{date}.html`, styled per `assets/report-template.html`:
   - Each source is a clickable hyperlink.
   - Capital section uses colored type tags (Funding/Earnings/IPO/M&A).
   - **No WeChat-summary lines.**
   - Contains all items.

### Step 4.5: Format Pre-flight (machine self-check, mandatory) [Deterministic]

⚠️ **Run immediately after generating the MD. Must print three counts + all assertions. No output = fail = return to Step 3 and regenerate.** Do the counts with actual `grep -c`, never by eye.

```
📐 Format pre-flight
- item count (^### \d+\.)            = N1
- WeChat-summary count (^> 企微摘要：) = N2
- section count (^## [A-E]\.)         = N3
- assert N1 > 0 ?                     ✅ / ❌   (at least 1; N1=0 → fail)
- assert N1 == N2 ?                   ✅ / ❌   (if N2<N1, check for fullwidth/halfwidth colon variants, then return to Step 3)
- assert N3 == 5 ?                    ✅ / ❌
- section-title language (5 English) ✅ / ❌
- numbering: each section restarts at 1, no cross-section continuation ✅ / ❌
- item separators: a standalone `---` between every two items          ✅ / ❌
- source-link format: every `**来源：**` line is `[text](url)`, no bare `（http...）` ✅ / ❌
- "Top 3" format: `**Top 3 takeaways:**` + ordered list 1./2./3.       ✅ / ❌
- "Top 3" char count: each takeaway 15–30 chars                        ✅ / ❌
- "Top 3" contains product names / versions / figures / bold ?         ❌ clean / ⚠️ hit
- verdict char count = X (≤120)                                        ✅ / ❌
- WeChat-summary char count: each ≤120                                 ✅ / ❌
- field-name compliance (per section, four required fields, no invented field names) ✅ / ❌
- field order: source line first after title, WeChat-summary line last ✅ / ❌
- conclusion: ✅ pass → Step 5 / ❌ fail → regenerate
```

**Any ❌ → return to Step 3 and regenerate the affected content.** Never patch item-count gaps or backfill summaries during the Step 5 review. This gate exists to stop "format drift causing missing WeChat-push content" at the root.

### Step 5: Review & Fix (business layer, 7 checks) [LLM]

After generation and before push, run one full review. Fail → no push. **Precondition: Step 4.5 must pass.**

1. **Recency compliance** — verify each item's publish date against the window (forced date derivation + per-item verification table).
2. **Cross-section dedup** — one event in two+ sections → merge into highest-priority section.
3. **Section-admission compliance** — each item fits its section's admission bar.
4. **Source quality** — each item has a clear first-hand source link (C may allow high-trust secondary).
5. **Content quality** — takeaways are trend judgments (15–30 chars, not event summaries); verdict ≤120 chars, no event-detail repetition; each item follows fact → impact-judgment.
6. **Search coverage** — every Tier-1 vendor was targeted; thin content means expand search, not pad.
7. **Less-is-more** — item counts within range, no padding with low-quality items.

> Format integrity is backstopped by the Step 4.5 machine pre-flight. On any format problem, return to Step 3 to regenerate rather than patching in review.

### Step 6: Deliver (per config) [Deterministic]

Per `config.json`, push to any of **9 channels** (✅ verified / 📦 community-contributed, unverified):

**Regional:** ✅ **WeChat Work** (`scripts/send_wecom.py` — concise summary first <4096 bytes, then full HTML; 3-layer priority fill; no links in summary; duplicate-push lock), 📦 **DingTalk** (`send_dingtalk.py`), 📦 **Feishu/Lark** (`send_feishu.py`).
**Global:** 📦 **Slack** (`send_slack.py`), 📦 **Discord** (`send_discord.py`), 📦 **Telegram** (`send_telegram.py`), 📦 **Microsoft Teams** (`send_teams.py`).
**Universal:** 📦 **Email** (`send_email.py`, SMTP), ✅ **GitHub Pages** (`deploy_github.py`, auto-archives history).

> **Encoding note (Windows):** under a GBK locale, running the push script directly can raise `UnicodeEncodeError` on emoji output and crash a channel mid-print. Run with `PYTHONIOENCODING=utf-8 python -X utf8 <script>` or add `sys.stdout.reconfigure(encoding='utf-8')` at the top of the entry script. If some channels fail, re-push only the failed channel with `--force` to avoid duplicate sends.

## Hard Rules

> These cannot be violated. They override all other guidance.

1. **Stability first.** The system's only goal is zero format drift, zero missed push, zero manual repair. Do not add features, change structure, or "optimize" existing rules unless explicitly requested.
2. **Recency red line.** Items whose original publish date is outside the window (weekday 24h / Monday 72h) are never admitted, no exceptions.
3. **First-hand source required.** Every item traces to a first-hand source; no pure secondhand media analysis.
4. **Never fabricate.** All figures, dates, versions come from the source text; no guessing.
5. **Dedup self-check.** After all sections, run cross-section dedup; one event in at most one section.
6. **Step 4.5 pre-flight.** After generating MD, print the counts + all assertions; any ❌ → return to Step 3. Never patch in review.
7. **Review gate.** Step 5 must pass before push; better unsent than flawed.
8. **Less is more.** No section lowers its bar for item count; an empty section beats a padded one.
9. **Push authority tiers.** In automation mode, business-layer review issues may be self-corrected then pushed; a generation-stage crash (Step 4.5 ❌) must be regenerated, never self-patched; in manual mode, issues await user confirmation.
10. **Source links are Markdown.** Source lines must use `[text](url)`; bare `（http...）` is forbidden (it breaks HTML clickable links).

## Failure Handling

| Scenario | Action |
|----------|--------|
| **Generation-stage format crash (Step 4.5 ❌)** | Return to Step 3 and regenerate the affected section or whole doc; never patch item-count gaps in review |
| Search yields nothing | Report "no qualifying information today", generate an empty template (title + date only), do not push |
| A single source is unavailable | Skip it, continue other searches, note "⚠️ {source} unreachable" |
| All candidates fail review | Output the review detail, do not push, await user decision |
| Candidates severely insufficient (A–D total < 5) | Expand search first; if still short, demote borderline items to E (tagged 【Demoted】); if still short, report and pause for user decision |
| Push fails (webhook timeout / 403) | Retry once; if it still fails, save files to workspace and notify user to push manually |
| Config missing | Generate defaults via `scripts/init_config.py`, then continue |
| HTML template missing | Generate Markdown only, skip HTML, state so in output |
| File detection with non-ASCII (e.g. Chinese) filenames | Use `find`, not `ls` — shells like Git Bash mishandle quoting of non-ASCII paths and produce false "file missing" results |

## Customization

### Switch industry
Edit `config.json` `customization`: focus areas (default Data+AI), vendor priority list, open-source project list, output language and format. Then replace the `[Default Profile: Data+AI]` block in this SKILL.md with your domain's vendors, sources, and search queries.

### Add a delivery channel
Enable it in `config.json` `adapters` and set its config:

| Channel | Key | Type | Main env vars |
|---------|-----|------|---------------|
| WeChat Work | `wechatwork` | Webhook | `WECOM_WEBHOOK_URL` |
| DingTalk | `dingtalk` | Webhook | `DINGTALK_WEBHOOK_URL`, `DINGTALK_SECRET` |
| Feishu | `feishu` | Webhook | `FEISHU_WEBHOOK_URL`, `FEISHU_SECRET` |
| Slack | `slack` | Webhook | `SLACK_WEBHOOK_URL` |
| Discord | `discord` | Webhook | `DISCORD_WEBHOOK_URL` |
| Telegram | `telegram` | Bot API | `TELEGRAM_BOT_TOKEN`, `TELEGRAM_CHAT_ID` |
| Teams | `teams` | Webhook | `TEAMS_WEBHOOK_URL` |
| Email | `email` | SMTP | `SMTP_HOST`, `SMTP_USER`, `SMTP_PASSWORD` |
| GitHub | `github` | API | `GITHUB_TOKEN`, `GITHUB_USER` |

### Adjust the schedule
Edit `config.json` `cron`:
```json
{ "schedule": "0 8 * * 1-5", "timezone": "Asia/Shanghai" }
```

---

*Default profile: Data+AI infrastructure. Framework is industry-agnostic — configure for any domain with public news sources.*

name: data-ai-daily-brief version: "5.0.0" description: > Turn any industry into a daily intelligence briefing. An AI agent searches, filters, writes, and delivers structured daily briefs to 9 channels — with machine-checked formatting and a business review gate. Ships with a Data+AI profile out of the box; switch to any domain via config. read_when:

daily brief
industry report
industry newsletter
intelligence brief
日报
行业日报
情报简报 allowed-tools:
read_file
write_to_file
replace_in_file
execute_command
web_search
web_fetch disable: false

Industry Daily Brief

An AI-driven skill that generates a high-quality industry intelligence brief: it automatically searches, filters, writes, and delivers a structured daily report. It ships with a Data+AI profile as the working example, and can be switched to any industry through the configuration file.

How to read this document The Workflow, Confidence Tiers, Section Definitions, Format Contract, and Hard Rules below are industry-agnostic — they are the engine. Everything inside a block marked [Default Profile: Data+AI] is an example configuration (vendor lists, search queries, focus areas) that you replace when targeting another domain. Do not treat the Data+AI specifics as part of the framework.

Workflow

When the user requests a daily brief, execute the following steps in order.

Step 1: Confirm Configuration [Deterministic]

Read the workspace config.json (if present).
If absent, initialize defaults via scripts/init_config.py.
Confirm the target date (default: today) and the output channels.

Step 2: Collect & Filter Information [Deterministic + LLM]

Use web_search to gather information, applying the following priorities and filters.

Core Principles

Relevance first, filter ruthlessly. Every item must clearly answer: does this affect the product roadmap, architecture, cost structure, governance, operational efficiency, or real-world adoption within the target industry? If the answer is not a clear yes, exclude it.

Less is more. Never lower the admission bar just because a section has few items. The value of the brief is precision, not item count.

Three-Phase Search Strategy

Phase 1 — Targeted first-hand source search (mandatory). For each Tier-1 vendor in the active profile, search its official channels (site blog, release notes, GitHub releases, press wires) one by one. Also run dedicated funding / M&A / earnings queries for the tracked companies.

Phase 2 — Expanded discovery (supplementary coverage). Broaden with topic keyword searches across the profile's focus areas and the date range, to catch items the targeted search missed.

Phase 3 — Source tracing (mandatory). For any item discovered via secondary media, use web_fetch or an additional site: search to trace it back to a first-hand source. Items with no traceable first-hand source are flagged ⚠️ unverified or demoted to the Watchlist.

Coverage requirement: every Tier-1 vendor in the active profile must receive at least one targeted search.

Recency Window (red line)

Weekdays (Tue–Fri): strictly cover only information first published within the last 24 hours (08:00 the previous day → 08:00 today, target timezone).

Monday special rule: the window expands to 72 hours (Friday 08:00 → Monday 08:00), covering Fri–Sun. Monday item cap is raised, and the title is marked as covering the weekend.

⚠️ Recency red line — never admit any of the following:

Information whose original publish date falls outside the current window
Information that appeared in a previous brief
Stale announcements from days or weeks ago
Pre-announced schedules (conferences/summits) that are not a today first disclosure

✅ How to verify recency:

Check the original page's publish date.
If it falls outside the window → exclude.
Before writing, list each candidate's publish date and confirm item by item.

Stale-news guard: a "big news" item surfacing in search results must have its original publish date independently confirmed before admission — search ranking is not recency.

`[Default Profile: Data+AI]` — Focus, Vendors, Sources

Replace this entire block when targeting another industry.

Focus areas: big data, data platforms, data infrastructure, data governance, data engineering, lakehouse architecture, query engines, stream/batch processing, vector search infrastructure, open-source data ecosystem. AI items are admitted only when they clearly affect the data platform.

Strictly exclude: pure-AI news (model releases, benchmarks, consumer AI), AI items with no direct data-platform link, secondhand financial/mass-media analysis, content farms / clickbait / unsourced rewrites.

Tier-1 vendors: AWS, Google Cloud, Microsoft Azure, Databricks, Snowflake, Alibaba Cloud, Tencent Cloud, Huawei Cloud, Volcengine. Tier-2 vendors: Confluent, MongoDB, Elastic, ClickHouse, Cloudera, Starburst/Trino, dbt Labs, Fivetran, Airbyte, Dataiku, Palantir, Baidu AI Cloud, JD Cloud. Chip vendors (only if directly data-platform related): NVIDIA, Intel, AMD.

Open-source projects: Iceberg, Hudi, Paimon, Delta Lake, Trino, Spark, Flink, Ray, Airflow, Kafka, dbt, ClickHouse, DuckDB, Milvus, Weaviate, Lance/LanceDB, StarRocks, Doris, SeaTunnel, Amoro.

Funding/IR sources: SiliconANGLE Big Data, DBTA, InfoQ, PR Newswire, Business Wire, SEC EDGAR, company IR pages, Crunchbase News, CB Insights, PitchBook News, TechCrunch Venture.

Analyst firms: Gartner, Forrester, IDC, a16z, Sequoia, Bessemer, Futurum Group, Constellation Research, Wikibon/SiliconANGLE Research; plus regional research institutes and policy/standards bodies relevant to the market.

Source rules: accept only first-hand sources (official sites/blogs/release notes, GitHub repos, original posts on X/LinkedIn/blogs, earnings-call transcripts, analyst reports, PR Newswire/Business Wire). Do not accept secondhand financial-media analysis (analyst reports excepted).

Step 3: Write the Brief [LLM]

Produce a professional brief for practitioners in the target industry, based only on the filtered items.

Confidence Tiers

Level A — confirmed fact: first-hand source (official site/blog, GitHub release, earnings filing, live event). → may go into A / B / C / D.
Level B — high-trust secondary: reliable media (Reuters, Bloomberg, TechCrunch) but no first-hand doc yet. → may cautiously go into A / C with a "media report / no official filing" label; prefer the Watchlist.
Level C — indirect signal / unverified rumor: social leaks, community chatter, unmerged-PR speculation. → Watchlist only.

Deduplication

Each item belongs to exactly one primary section. Priority: A > B > C > D > E. Mandatory self-check: after writing all sections, list every event/source/product name and check for cross-section duplicates. If one event appears in two or more sections, keep the highest-priority section and delete or reduce the rest to a one-line reference (≤15 chars).

Title & Opening

Title format: # <Brand> Daily Brief | YYYY-MM-DD (Monday marked as covering the weekend).

Fixed opening structure:

**Top 3 takeaways:**
1. [a single trend judgment — direction not full event, 15–30 chars]
2. [a single trend judgment — direction not full event, 15–30 chars]
3. [a single trend judgment — direction not full event, 15–30 chars]

**Verdict:** [the single most important industry judgment of the day, 1–2 sentences, ≤120 chars, landing on direction / investment focus / market shift]

Key distinction: the "Top 3 takeaways" are not a summary of Top Signals nor three event headlines. They are three cross-event "directions worth taking away today."

Hard constraints:

Each takeaway 15–30 chars; no full product names, version numbers, or specific figures; no bold.
A good takeaway lets the reader grasp what direction changed and so what.

Verdict constraints: ≤120 chars; a directional judgment, not a repeat of the takeaways or of Section A event details.

Self-check: after writing the 3 takeaways, cover Section A and read only the takeaways. If the reader can reconstruct each Section A headline and key figure from them, the takeaways read too much like a summary — rewrite.

Sections

The five sections use English titles and are mandatory (a section may be empty but its header stays).

A. Top Signals (3 items). The day's most important events; first-hand source required.

### 1. Event title
**来源：** [specific source](url)
**摘要：** 2–3 sentences
**为什么对数据平台重要：** ...
> 企微摘要：one-line semantic compression

B. Product & Tech (0–6 items, less-is-more). Strictly product & tech moves in the target industry. Each item: title, source, summary (1–2 sentences), impact judgment, WeChat summary.

C. Views & Research (0–5 items). Two kinds of high-value content: original views from key people (founders/CEOs/CTOs in interviews, talks, blogs, X, LinkedIn) and formal research from high-credibility institutions (Gartner/Forrester/IDC/Omdia/etc., regional research bodies, top brokerages), plus officially-released policy & standards relevant to the industry. Each item: name, source, core view, mapping-to-industry judgment, WeChat summary.

D. Capital & Corporate (0–4 items, less-is-more). Capital/company events directly relevant to the industry, with inline type tags: 【Funding】 / 【Earnings】 / 【IPO】 / 【M&A】.

### 1. 【Funding】Event title
**来源：** [specific source](url)
**核心数据：** amount / valuation / revenue / growth
**摘要：** 2–3 sentences
**对数据平台的影响：** ...
> 企微摘要：one-line semantic compression

E. Watchlist (1–3 items). Three kinds: 【Preview】 upcoming events, 【Demoted】 valuable items not meeting A–D bars, 【Tracking】 follow-ups whose impact is still unverified. Each item: title, source, why it's worth watching, what signal to wait for, WeChat summary.

WeChat-Summary Field (all sections)

After all detailed fields, every item must carry one line: > 企微摘要：one-sentence semantic compression.

Rules:

Semantic compression of the whole item, ≤120 chars.
A standalone, self-contained complete sentence.
No links, no source labels.
Markdown only — never rendered in the HTML output.

The **来源：** and > 企微摘要： field markers are kept in their original Chinese form because the WeChat-push and HTML-conversion scripts (scripts/) parse these exact strings. When localizing the brief to another language, keep these two markers as-is or update the scripts in lockstep.

Output Requirements

Rank by importance, differentiate by channel. Wider search means more candidates — rank strictly by real industry impact > source authority > topic heat; never lower the bar because more sources appeared.
WeChat concise version: keep section caps (A≤3, B≤6, C≤5, D≤4, E≤3); pick only the most important items.
HTML full version: sections may relax by 2–3 items (A still ≤3, B≤8, C≤7, D≤6, E≤5).
Professional, concise, restrained tone. Every item must have source, summary, and impact judgment. Total 10–14 items (Monday 14–20). Never fabricate data.

Step 4: Generate Output Files [Deterministic + LLM]

Markdown <Brand>_Daily_Brief_{date}.md:
- Every item carries a > 企微摘要：... line.
- Contains all items (incl. HTML-extended ones); the WeChat-summary line marks which enter the concise push.
- Source links must use Markdown link syntax [text](url) — never bare text （https://...）. Bare URLs render as plain text in HTML and break clickable links.
HTML <Brand>_Daily_Brief_{date}.html, styled per assets/report-template.html:
- Each source is a clickable hyperlink.
- Capital section uses colored type tags (Funding/Earnings/IPO/M&A).
- No WeChat-summary lines.
- Contains all items.

Step 4.5: Format Pre-flight (machine self-check, mandatory) [Deterministic]

⚠️ Run immediately after generating the MD. Must print three counts + all assertions. No output = fail = return to Step 3 and regenerate. Do the counts with actual grep -c, never by eye.

📐 Format pre-flight
- item count (^### \d+\.)            = N1
- WeChat-summary count (^> 企微摘要：) = N2
- section count (^## [A-E]\.)         = N3
- assert N1 > 0 ?                     ✅ / ❌   (at least 1; N1=0 → fail)
- assert N1 == N2 ?                   ✅ / ❌   (if N2<N1, check for fullwidth/halfwidth colon variants, then return to Step 3)
- assert N3 == 5 ?                    ✅ / ❌
- section-title language (5 English) ✅ / ❌
- numbering: each section restarts at 1, no cross-section continuation ✅ / ❌
- item separators: a standalone `---` between every two items          ✅ / ❌
- source-link format: every `**来源：**` line is `[text](url)`, no bare `（http...）` ✅ / ❌
- "Top 3" format: `**Top 3 takeaways:**` + ordered list 1./2./3.       ✅ / ❌
- "Top 3" char count: each takeaway 15–30 chars                        ✅ / ❌
- "Top 3" contains product names / versions / figures / bold ?         ❌ clean / ⚠️ hit
- verdict char count = X (≤120)                                        ✅ / ❌
- WeChat-summary char count: each ≤120                                 ✅ / ❌
- field-name compliance (per section, four required fields, no invented field names) ✅ / ❌
- field order: source line first after title, WeChat-summary line last ✅ / ❌
- conclusion: ✅ pass → Step 5 / ❌ fail → regenerate

Any ❌ → return to Step 3 and regenerate the affected content. Never patch item-count gaps or backfill summaries during the Step 5 review. This gate exists to stop "format drift causing missing WeChat-push content" at the root.

Step 5: Review & Fix (business layer, 7 checks) [LLM]

After generation and before push, run one full review. Fail → no push. Precondition: Step 4.5 must pass.

Recency compliance — verify each item's publish date against the window (forced date derivation + per-item verification table).
Cross-section dedup — one event in two+ sections → merge into highest-priority section.
Section-admission compliance — each item fits its section's admission bar.
Source quality — each item has a clear first-hand source link (C may allow high-trust secondary).
Content quality — takeaways are trend judgments (15–30 chars, not event summaries); verdict ≤120 chars, no event-detail repetition; each item follows fact → impact-judgment.
Search coverage — every Tier-1 vendor was targeted; thin content means expand search, not pad.
Less-is-more — item counts within range, no padding with low-quality items.

Format integrity is backstopped by the Step 4.5 machine pre-flight. On any format problem, return to Step 3 to regenerate rather than patching in review.

Step 6: Deliver (per config) [Deterministic]

Per config.json, push to any of 9 channels (✅ verified / 📦 community-contributed, unverified):

Regional: ✅ WeChat Work (scripts/send_wecom.py — concise summary first <4096 bytes, then full HTML; 3-layer priority fill; no links in summary; duplicate-push lock), 📦 DingTalk (send_dingtalk.py), 📦 Feishu/Lark (send_feishu.py). Global: 📦 Slack (send_slack.py), 📦 Discord (send_discord.py), 📦 Telegram (send_telegram.py), 📦 Microsoft Teams (send_teams.py). Universal: 📦 Email (send_email.py, SMTP), ✅ GitHub Pages (deploy_github.py, auto-archives history).

Encoding note (Windows): under a GBK locale, running the push script directly can raise UnicodeEncodeError on emoji output and crash a channel mid-print. Run with PYTHONIOENCODING=utf-8 python -X utf8 <script> or add sys.stdout.reconfigure(encoding='utf-8') at the top of the entry script. If some channels fail, re-push only the failed channel with --force to avoid duplicate sends.

Hard Rules

These cannot be violated. They override all other guidance.

Stability first. The system's only goal is zero format drift, zero missed push, zero manual repair. Do not add features, change structure, or "optimize" existing rules unless explicitly requested.
Recency red line. Items whose original publish date is outside the window (weekday 24h / Monday 72h) are never admitted, no exceptions.
First-hand source required. Every item traces to a first-hand source; no pure secondhand media analysis.
Never fabricate. All figures, dates, versions come from the source text; no guessing.
Dedup self-check. After all sections, run cross-section dedup; one event in at most one section.
Step 4.5 pre-flight. After generating MD, print the counts + all assertions; any ❌ → return to Step 3. Never patch in review.
Review gate. Step 5 must pass before push; better unsent than flawed.
Less is more. No section lowers its bar for item count; an empty section beats a padded one.
Push authority tiers. In automation mode, business-layer review issues may be self-corrected then pushed; a generation-stage crash (Step 4.5 ❌) must be regenerated, never self-patched; in manual mode, issues await user confirmation.
Source links are Markdown. Source lines must use [text](url); bare （http...） is forbidden (it breaks HTML clickable links).

Failure Handling

Scenario	Action
Generation-stage format crash (Step 4.5 ❌)	Return to Step 3 and regenerate the affected section or whole doc; never patch item-count gaps in review
Search yields nothing	Report "no qualifying information today", generate an empty template (title + date only), do not push
A single source is unavailable	Skip it, continue other searches, note "⚠️ {source} unreachable"
All candidates fail review	Output the review detail, do not push, await user decision
Candidates severely insufficient (A–D total < 5)	Expand search first; if still short, demote borderline items to E (tagged 【Demoted】); if still short, report and pause for user decision
Push fails (webhook timeout / 403)	Retry once; if it still fails, save files to workspace and notify user to push manually
Config missing	Generate defaults via `scripts/init_config.py`, then continue
HTML template missing	Generate Markdown only, skip HTML, state so in output
File detection with non-ASCII (e.g. Chinese) filenames	Use `find`, not `ls` — shells like Git Bash mishandle quoting of non-ASCII paths and produce false "file missing" results

Customization

Switch industry

Edit config.json customization: focus areas (default Data+AI), vendor priority list, open-source project list, output language and format. Then replace the [Default Profile: Data+AI] block in this SKILL.md with your domain's vendors, sources, and search queries.

Add a delivery channel

Enable it in config.json adapters and set its config:

Channel	Key	Type	Main env vars
WeChat Work	`wechatwork`	Webhook	`WECOM_WEBHOOK_URL`
DingTalk	`dingtalk`	Webhook	`DINGTALK_WEBHOOK_URL`, `DINGTALK_SECRET`
Feishu	`feishu`	Webhook	`FEISHU_WEBHOOK_URL`, `FEISHU_SECRET`
Slack	`slack`	Webhook	`SLACK_WEBHOOK_URL`
Discord	`discord`	Webhook	`DISCORD_WEBHOOK_URL`
Telegram	`telegram`	Bot API	`TELEGRAM_BOT_TOKEN`, `TELEGRAM_CHAT_ID`
Teams	`teams`	Webhook	`TEAMS_WEBHOOK_URL`
Email	`email`	SMTP	`SMTP_HOST`, `SMTP_USER`, `SMTP_PASSWORD`
GitHub	`github`	API	`GITHUB_TOKEN`, `GITHUB_USER`

Adjust the schedule

Edit config.json cron:

{ "schedule": "0 8 * * 1-5", "timezone": "Asia/Shanghai" }

Default profile: Data+AI infrastructure. Framework is industry-agnostic — configure for any domain with public news sources.