ai-trend-brief

name: "ai-trend-brief" description: "Summarizes the latest AI product and agent developments into a concise daily brief. Invoke when user wants a 24-hour roundup of AI product launches, agent tooling updates, company announcements, builder ecosystem movement, or discussion signals from specified sources."

AI Trend Brief

Overview

Produce a one-page brief covering the most recent 24 hours of AI product and agent developments. Prioritize source-grounded summarization over broad commentary. The goal is to help the user scan what changed, why it matters, and which items are only community signal rather than confirmed fact.

Required operating rules

• Restrict the default scope to AI products, agents, agent tooling, application-layer AI workflows, model-powered product releases, developer tooling around agents, and notable ecosystem moves that directly affect builders or product teams.
• FLEXIBLE TIME WINDOW: Use a tiered approach when 24-hour results are insufficient:
  • Tier 1 (preferred): Most recent 24 hours — use when major announcements exist
  • Tier 2 (fallback): Most recent 7 days — use when 24-hour window yields < 3 items
  • Tier 3 (last resort): Most recent 30 days — use only when both above yield nothing, and ONLY include items that have ongoing community discussion or updates within 7 days
• Default time window is the most recent 24 hours unless the user explicitly changes it or conditions above trigger fallback
• Prioritize sources in this order:
  1. user-provided sources
  2. default company-official sources
  3. default agent and builder sources
  4. default media and community sources
  5. additional authoritative public sources discovered during verification
• Prefer direct source reading over secondary commentary whenever possible.
• Do not turn the brief into a generic media roundup. Select only items that have clear relevance to AI products or agent ecosystems.

Default Sources

When the user does not provide JSON sources, use these defaults from references/default-sources.md:

Priority A: company-official sources (Use first)

• OpenAI official blog and product pages
• Anthropic official blog and docs
• Google / Google DeepMind official blogs and product posts
• Microsoft official AI and product blogs
• Meta AI and Meta official product blogs
• xAI official site and official X account
• Mistral AI official blog and product pages
• Perplexity official blog and official announcements
• major product-company official X accounts when clearly official

Priority B: agent and builder sources (Use after Priority A or when the topic is tooling and builders)

• LangChain official announcements
• LlamaIndex official announcements
• Vercel AI / AI SDK announcements
• notable agent platform company blogs and release notes
• well-known builder ecosystem newsletters only when they contain concrete announcements

Priority C: media and community sources (Use for discovery and context, not as the sole basis for strong factual claims)

• Media: TechCrunch, The Information, Bloomberg, Financial Times, VentureBeat, other reputable technology media
• Community: Hacker News threads, relevant Reddit communities, X posts from non-official accounts

Priority handling: When the same event appears in multiple places:

1. prefer the official announcement for the factual description
2. use media for framing or extra context
3. use community only for signal strength or reaction, never as the core proof of a factual product claim

Input handling

Source input modes

Use these source modes in order:

1. User JSON sources if supplied
2. Built-in default source lists if user does not supply sources
3. Supplemental authoritative public sources only when needed to verify or contextualize an item

Accepted JSON schema for user-added sources

Expect this exact structure when the user adds sources:

{
  "sources": [
    {"type": "x", "value": "@account_name"},
    {"type": "website", "value": "https://example.com"},
    {"type": "reddit", "value": "https://www.reddit.com/r/..."},
    {"type": "hn", "value": "https://news.ycombinator.com/item?id=..."},
    {"type": "news", "value": "https://example.com/article"}
  ]
}

Allowed source types

Accept these type values: x, website, reddit, hn, news

If a source type is outside this list, ignore it and mention it in the skipped-source note.

Invalid or inaccessible sources

If a user-supplied source is inaccessible, malformed, ambiguous, or unavailable:

• skip it
• continue the task without blocking
• add it to a short Skipped sources line at the end
• do not fail the whole run because one source is unusable

Empty result handling (CRITICAL)

Tiered fallback rules — activate only when initial search yields insufficient results:

Step 1: If 24-hour search yields ≥ 3 qualifying items → Use Tier 1 (24 hours), brief complete

Step 2: If 24-hour search yields < 3 items but > 0 → Automatically expand to 7-day window, re-run search, mark items as "近期持续" if from 24h or "本周新增" if from 24h-7d

Step 3: If 24-hour search yields 0 items → Expand to 7-day search. If 7-day yields ≥ 3 items, include them with clear time labels. If still < 3, expand to 30-day search for items with fresh community discussion (HN/Reddit activity within 7 days)

Output rules:

• Always include a 时效说明 section at the top stating which tier was used and why
• Mark items clearly: [24h] for same-day, [7d] for 2-7 days ago, [30d+] for older items with recent discussion
• Do NOT mix tiers without labeling — reader must know the age of each item
• If all tiers yield nothing → output "无动态 — 近期暂无AI产品与Agent相关公开信息"

Retrieval Tools

Use these tools in the following order of priority when gathering information:

Tool Selection Strategy

Required Search Syntax

For time-restricted searches, ALWAYS include:

• after:YYYY-MM-DD — restricts results to after this date (today - 1 day)
• site:<domain> — restricts to specific domains for authoritative sources

Example queries for 24-hour window (2026-03-18):

site:openai.com OR site:anthropic.com OR site:deepmind.google.com after:2026-03-18
site:langchain.com OR site:llamaindex.com OR site:vercel.com/blog after:2026-03-18
AI agent product launch after:2026-03-18

Per-Source Search Commands

OpenAI:

WebSearch: site:openai.com OR site:chatgpt.com after:{date}

Anthropic:

WebSearch: site:anthropic.com OR site:claude.ai after:{date}

Google DeepMind:

WebSearch: site:deepmind.google OR site:blog.google after:{date}

Microsoft:

WebSearch: site:microsoft.com OR site:azure.microsoft.com after:{date}

Meta AI:

WebSearch: site:ai.meta.com OR site:meta.ai after:{date}

xAI:

WebSearch: site:x.ai OR @xaboratory after:{date}

Agent Builder Ecosystem (LangChain, LlamaIndex, Vercel):

WebSearch: site:langchain.com OR site:llamaindex.com OR site:vercel.com/blog after:{date}

Discovery Search (Fallback)

When to use: When site: restricted searches return < 3 results AND the tiered window search yields insufficient items.

Discovery mode — 2-step approach:

Step 1 — Broad discovery search:

WebSearch: AI agent product announcement {date_range}
WebSearch: OpenAI Anthropic Google AI release {month_year}

Step 2 — Source quality filter: After discovery, verify each item's source quality:

• Accept: Primary sources (openai.com, anthropic.com, deepmind.google, azure.microsoft.com, ai.meta.com, x.ai)
• Accept: Major tech media (techcrunch.com, theinformation.com, venturebeat.com, bloomberg.com)
• Conditional accept: Aggregation sites (toutiao.com, 36kr.com, etc.) ONLY if:
  • They cite direct source URLs in the article
  • You MUST fetch the cited original source for fact verification
  • The original source is from Tier 1 or Tier 2
• Reject: Personal blogs, unknown Chinese sites, social media reposts without original links

Verified content flow:

1. Discovery finds article about GPT-5.4 on toutiao.com
2. If article links to openai.com → fetch original → verify facts
3. If no original link → downgrade to community signal or exclude

Fetching Web Pages

After identifying relevant articles via WebSearch, use WebFetch to get the full content for:

1. Official blog posts for exact announcements
2. Product pages for feature details
3. Technical documentation for tool changes

Do NOT rely solely on search snippets. Always fetch the source page for fact verification when the information is critical or the snippet is ambiguous.

Source credibility rules

Credibility tiers

Treat source categories differently.

Tier 1: official / primary

• company blogs, official release notes, product pages, official documentation, official X accounts

Tier 2: strong secondary

• reputable technology media, direct interviews with named participants, conference or event recaps from known publishers

Tier 3: community signal only

• personal X accounts, Reddit posts, Hacker News threads, reposts, reactions, speculation, or summaries without direct evidence

Community signal rule

Content from Reddit, Hacker News, or non-official social discussion can surface leads, but it must not be written as established fact unless corroborated by a primary or strong secondary source. Mark these items explicitly as 社区信号.

Fact summarization rule

When describing an item, summarize only what the cited source actually supports.

• Do not exaggerate scope
• Do not upgrade a teaser into a launch
• Do not infer causality unless the source states it
• Do not append implications as if they were facts
• Do not collapse rumor, opinion, and announcement into the same statement
• If a source says a company "previewed" something, do not rewrite it as "released"
• If a source says users "reported" something, do not rewrite it as a confirmed product change

Use this distinction internally:

• Fact: directly supported by the cited source wording or direct evidence from the page
• Inference: a cautious interpretation based on multiple facts

Do not create an Unknown section unless the user explicitly asks for it.

Trend labeling rule

Use the label [趋势] only when the topic appears repeatedly across multiple sources within the last 24 hours. Because the default is frequency-based trend detection, [趋势] means repeated within the window, not necessarily structurally durable.

• Do not use [趋势] for a one-off mention
• Do not use [趋势] to imply long-term inevitability
• If an item is only from community discussion, write 社区信号 instead of [趋势] unless there is clear multi-source repetition

Output requirements

Language

Write in Chinese by default, while preserving English proper nouns, product names, repository names, company names, and technical terms when they are clearer in English.

Length

Target 500 to 900 Chinese characters for the main brief body. Stay compact.

Required output structure

Always use this structure unless the user explicitly requests another format:

# AI 产品与 Agent 简报
**时效说明：{Tier 1: 24小时 / Tier 2: 7天 / Tier 3: 30天+含近期讨论}**
**时效范围：YYYY-MM-DD HH:MM - YYYY-MM-DD HH:MM**

## 今日重点
- 2 to 4 bullets
- each bullet starts with either [趋势] or no label
- each bullet states the source-grounded fact first, then a short implication if warranted
- append compact source markers like [S1][S2]
- MUST include time tier label: [24h], [7d], or [30d+]

## Agent 产品动向
- 2 to 4 bullets focused on product releases, agent features, tooling changes, workflow builders, or shipping signals
- mark community-only items as 社区信号
- append source markers with time tier label

## 工具链动向
- 1 to 3 bullets on infra or developer tooling only if directly relevant to agent builders or product teams
- avoid drifting into pure research unless it has direct product impact
- append source markers with time tier label

## 社区信号
- 1 to 3 bullets maximum
- only include discussion items that are early but potentially useful
- must be labeled 社区信号
- do not overstate them
- append source markers with time tier label

## 信息源
- list S1, S2, ... with source title, link, AND publication timestamp

## Skipped sources
- include only if needed
- list skipped user-provided sources briefly

时效说明 (MANDATORY): Always include at the top to transparently communicate the data freshness to the reader.

Time tier labels: Every bullet MUST have a time tier prefix:

• [24h] — published within last 24 hours
• [7d] — published 2-7 days ago
• [30d+] — published >7 days ago but has recent community discussion (within 7 days)

Source markers

• Use bracketed markers like [S1], [S2]
• Every substantive bullet must have at least one marker
• Key bullets should have multiple markers when available

Selection criteria

TIMELINESS THRESHOLD (MANDATORY):

• An item is ONLY eligible if it was published or officially confirmed within the most recent 24 hours
• If a source article is about something that happened 2+ days ago, even if the article is new, do NOT include it unless the event itself was officially announced in the last 24 hours
• If no qualifying items exist in the 24-hour window, output a brief "无重大更新" notice and skip the standard sections

Include an item only if at least one of these is true:

• it announces or materially updates an AI product or agent capability (within 24 hours)
• it changes how builders can ship, deploy, evaluate, orchestrate, or monetize agent systems (within 24 hours)
• it reveals ecosystem movement that product teams should track soon (within 24 hours)
• it is repeated enough across the last 24 hours to be worth labeling [趋势]

Exclude items that are mainly:

• vague opinion without concrete event content
• generic research news without product or agent impact
• old stories resurfacing without a new development (even if coverage is new)
• low-value engagement bait
• duplicate bullets that say the same thing from different outlets

Workflow

1. Parse request — Check whether JSON sources were provided by the user
2. Calculate date ranges — Determine 3 date thresholds: {24h} = today-1, {7d} = today-7, {30d} = today-30
3. Execute Tier 1 search — Use WebSearch with site: restrictions and after:{24h} syntax on official + media sources
4. Count qualifying items — If ≥ 3 items found → proceed to Step 7
5. Execute site-restricted fallback — Expand after:{7d} → then after:{30d} with site restrictions
6. Discovery fallback (if Step 5 yields < 3 items):
   • Use broad discovery search WITHOUT site restrictions
   • Filter results by source quality rules above
   • For aggregation sites: fetch cited original source and verify facts
   • Mark items from unverified sources as 社区信号 or exclude
7. Fetch for verification — Use WebFetch on promising URLs to verify facts
8. Build brief — Assemble using required output structure with time tier labels
9. Output — Include 时效说明 header with tier used, include "无动态" only if all tiers fail