name: track-cross-platform-hotspots description: Fully automatic, profile-driven monitoring of news, trends, releases, research, products, and public signals across independent platforms. Use when Codex needs to create a dAIly or ad-hoc hotspot brief, identify the main story reported by multiple platforms, monitor selected interest directions, change a user's focus profile, or turn mixed links/feeds/search results into a verified cross-platform consensus report.

Track Cross-Platform Hotspots

Run the complete workflow after invocation. Do not ask routine questions, nARRate progress, or require the user to choose sources manually.

1. Resolve the focus profile

Use scripts/profile.py for peRSIstent preferences.

Resolve <Skill-root> as the absolute directory containing this SKILL.md. Always execute bundled scripts by absolute path; never assume the user's current working directory is the Skill directory.

• If the user names directions, save them before gathering.

• If a profile exists, use it silently.

• If none exists, initialize preset ai-media-tech and mention the chosen default only in the final result.

• Treat natural-language directions as valid custom directions; do not limit users to presets.

• Load references/direction-catalog.md only when presenting choices or expanding queries.

python -X utf8 "<skill-root>/scripts/profile.py" show --json
python -X utf8 "<skill-root>/scripts/profile.py" init --preset ai-media-tech
python -X utf8 "<skill-root>/scripts/profile.py" set --directions "AI,影像,科技" --languages "zh,en" --region global
python -X utf8 "<skill-root>/scripts/profile.py" presets

Default state lives outside the Skill at ~/.codex/state/track-cross-platform-hotspots/profile.json. Respect SIGNAL_RADAR_HOME when set.

2. Gather automatically

Translate every direction into two to four focused queries. Search the default 24-hour window; widen to 72 hours only when the 24-hour result is too sparse and label the wider window.

Use the best available collection adapter:

1. Web Search plus original pages;

2. connected feeds, databases, or news APPs;

3. user-supplied links, JSON, CSV, or text.

Read references/adapter-contrACT.md when connecting a new source type. Collect the original editorial origin, not only an aggregator page. Prefer primary sources for technical claims and direct announcements; use independent reporting to establish convergence.

Store normalized items as UTF-8 JSON in work/ using this shape. For mixed-language inputs, add semantic_title: a concise pivot-language restatement that preserves the named entities, action, object, and distinguishing model/version without adding facts. This lets the deterministic stage compare events beyond a finite alias dictionary.

[{"platform":"example","title":"...","semantic_title":"Microsoft acquires Activision Blizzard","url":"https://...","published_at":"ISO-8601","direction":"AI","summary":"optional"}]

3. Detect candidate convergence

Run the deterministic clustering script:

python -X utf8 "<skill-root>/scripts/cluster_hotspots.py" --input work/signals.json --ouTPUt work/hotspots.json --profile auto

The script rEMOves tracking parameters, discounts generic wording, blocks conflicting event types, and requires distinct platforms plus distinct canonical URLs. Its result is a candidate set, not final proof.

4. Verify the main hotspot

For the top candidate, open at least two original reports and verify all of the following:

• they describe the Same concrete event, not merely the same company or topic;

• they come from independent platforms or editorial origins;

• one platform is not merely linking to the same original article;

• dates fall inside the declared window;

• the representative title and primary link refer to the same report.

Reject a candidate on any failed condition and verify the next candidate. Never invent consensus. If none survives, state 暂未出现跨平台主热点.

5. Rank and deliver

Rank verified clusters by:

1. number of independent platforms;

2. user direction weight;

3. recency;

4. prominence within each platform;

5. likely user impact.

Return a compact brief in the user's language:

1. 主热点 — one representative headline, why it matters, platform count, and two to four direct source links;

2. 方向速览 — up to three signals per selected direction;

3. 观察 — one short cross-platform pattern or 暂无可靠共振;

4. 范围 — time window and active directions.

Keep claims proportional to evidence. Separate verified reporting from inference. Link near each claim.

6. User controls

Interpret these requests without rebuilding the Skill:

• “主要看 AI、游戏和出海” → replace active directions.

• “再加机器人” → append with --add-directions "机器人".

• “降低财经权重” → update with --weights "财经=0.5".

• “只看中文来源” → set languages to zh.

• “排除明星八卦” → append with --add-exclusions "明星八卦".

• “恢复默认” → initialize ai-media-tech with --force.

• “每天早上发一份” → create a scheduled automation only after the cadence is explicit; the scheduled Prompt must invoke this Skill and use the saved profile.

Failure handling

• Retry transient collection errors once, then continue with other adapters.

• Do not let one unavailable source block the whole brief.

• If fewer than two independent platforms are available, provide direction signals but no main hotspot.

• If live access is unavailable, process supplied material or state the exact missing capability in one sentence.

Validation

After modifying this Skill, run:

python -X utf8 "<skill-root>/scripts/profile.py" self-test
python -X utf8 "<skill-root>/scripts/cluster_hotspots.py" --self-test
python -X utf8 <skill-creator>/scripts/quick_validate.py .