topic-intelligence

Topic Intelligence

Use one public Skill for current-topic selection and evidence-backed content planning.

Before querying or reasoning, read references/quality-contract.md. Read references/selection-workflow.md when selecting or comparing topics and references/brief-workflow.md when the request includes a content brief.

Choose exactly one mode

Infer the smallest mode that satisfies the request. Do not make the user choose a mode name.

Mode 1: selection only

Run one bounded live feed query, compare qualified candidates, and stop after the requested shortlist or finalist. Do not add an unsolicited brief.

Mode 2: brief for a supplied current topic

• Exact Radar ID/current card: preserve its identity and do not run feed to select again. Use history <exact-feed-id> only when movement matters. If an ID alone lacks enough facts, ask for the current card or explain the gap.
• Topic name without stable identity: run at most one bounded feed with --q <supplied-topic-name> --lane raw solely to resolve that same event. Accept only a clear semantic match, preserve its exact id, and never replace it with a more convenient topic.

Mode 3: selection followed by brief

Run exactly one bounded selection feed, choose one finalist, preserve its exact feed id, and build the brief from the same evidence. Do not run a second feed. History is optional and finalist-only.

Public runtime contract

The bundled helper may call only public no-cost feed, sources, and history endpoints. The current host model performs comparison and editorial reasoning. Resolve the helper from the loaded Skill root:

scripts/topic_radar_client.py

Canonical commands:

python3 <skill-local-helper> --timeout 30 feed --q AI --lane all --locale zh --market CN --max-age-hours 24 --limit 12
python3 <skill-local-helper> --timeout 30 sources
python3 <skill-local-helper> --timeout 30 history <exact-feed-id>

Requirements:

• use python3, never python, python2, or direct execution;
• every feed explicitly passes --lane, --locale, and --market; never rely on API defaults;
• place helper-wide options before the subcommand;
• pass the history ID as the sole positional argument;
• run each helper call as one standalone direct command;
• never combine it with pipes, &&, ;, redirection, command substitution, backticks, a here-doc, jq, or another Python process;
• read JSON directly from stdout and never probe CLI syntax;
• never repeat a successful request to recover display-truncated output; use the returned JSON and mark unavailable fields unknown;
• never use a repository-root, sibling-repository, user-global, or copied helper;
• never use --base-url or an origin override in the official public workflow.

Ordinary selection starts with 12 candidates and never exceeds 24. Selection-only and selection-followed-by-brief modes each use exactly one successful feed. For an explicit large export only, paginate sequentially with identical filters and generated_at, non-overlapping offsets, and a stop at has_more=false. Do not merge pages from changed snapshots or retry completed pages.

Explicit lane, language, and market

Infer locale from the answer language and market from the requested audience or geography. With no explicit audience/market, use zh + CN for Chinese and en + GLOBAL_EN for English.

• current/today's research-worthy topics: lane=all plus an explicit time window;
• strongest and sufficiently evidenced: lane=curated;
• accelerating/heating up: lane=rising;
• overseas-first/regional lead: lane=market_lead;
• early opportunity: lane=early;
• monitor but not ready: lane=watch;
• evergreen: lane=evergreen;
• web page “全部”/raw event pool: lane=raw;
• supplied-name resolution: lane=raw plus the exact name as q.

raw is the page's “全部”; API all is the candidate-lane union. Do not simulate a lane with legacy signal/trend_stage, silently switch lanes, or force a topic when the requested lane is empty.

Live evidence gate

Current claims require a live Radar response obtained in this task, equivalent current Radar data from a native host connection, or a current Radar response explicitly supplied by the user. Never substitute model memory, sibling repositories, snapshots, databases, fixtures, caches, logs, reports, exports, or prior-task artifacts.

Inspect generated_at, partial, stale, snapshot_age_seconds, refreshing, source_status, and topic evidence. Missing values are unknown, not zero. For “current” or “today”, a supplied snapshot older than one hour or marked stale=true cannot support a current acceleration claim.

Radar fields and evidence links are observations of what Radar returned, not independent verification of external claims. Use links as research leads, prefer primary sources, and place unresolved claims in must_verify.

Selection invariants

• Preserve the user's domain in the first query and reject semantic substring collisions after retrieval.
• Treat duration, language, audience, tone, and format as editorial constraints, not source/platform filters.
• Read editorial_lane(s), category/tags, content_type, evidence_state, the four component scores, classification/cluster confidence, rejection reasons, and what_happened/why_worth_research.
• Explain Radar's trend_score, editorial_score, confidence_score, and market_fit_score tradeoffs; never calculate a fifth host score.
• Count independent evidence by unique independent_origin_id, not collector, source_id, or display-name count.
• Use history for a finalist when a real rising/acceleration claim matters.
• Treat market_lead as a regional-lead observation, not proof of a measured domestic information gap.

For real-time selection, reject marketing, routine_update, and evergreen_advice content unless explicitly requested. unverified is only a research lead; verified is not independent external fact-checking. classification_confidence < 60 is normally not a finalist. cluster_confidence < 60 and non-empty rejection_reasons require disclosure. By default, watch, raw, and unverified items cannot appear in recommendation or become a Mode 3 finalist. They may be labeled as research leads only. Unless the user explicitly asks to brief a research lead, report no qualified recommendation and stop before the brief when no eligible finalist remains.

Output and brief invariants

Keep these layers visibly separate:

1. Radar facts and evidence observations;
2. Radar deterministic editorial judgments;
3. Host editorial analysis;
4. recommendation;
5. unknowns / must_verify.

Use all five headings even for a compact shortlist; never merge Radar fields and host judgment into one candidate paragraph.

The first layer contains exact IDs, verbatim titles, timestamps, publishers, canonical links, source roles/tiers, independent origins, source markets, history, and factual limits. The second contains returned lanes, component scores, evidence/confidence states, why_worth_research, and rejection reasons. End both Radar layers before host comparison or recommendations. A Radar title or original title label requires a verbatim copy.

For a brief, preserve one exact Radar id and include freshness, evidence/source limitations, angle, audience payoff, hook, narrative beats, research questions, preferred sources, must_verify, avoid_claims, unknowns, risks, and a visual/material plan as relevant.

Radar's editorial_score and market_fit_score are deterministic estimates. Radar does not measure actual audience size, topic/content saturation, real user-demand volume, post-publication performance, future reach, or virality. Audience and distribution conclusions remain host editorial judgments.

Safety and cost boundary

• Never call anonymous/public /insight or another AI Workstation model-backed endpoint from the public Skill.
• Never embed a shared credential or ask the user to paste a private key.
• Optional Premium Topic Insight requires a native connection explicitly authenticated to the user's account and enforcing that user's quota; treat it as model analysis, not independent evidence.
• Never present any Radar score, target platform, or host judgment as a performance guarantee.