Topic Brief

Topic Brief — Thematic Observation Briefing Generator

Generate a public-news-based information briefing for any subject (region / industry / policy issue / institution) and produce a single self-contained HTML file ready to paste into the 微信公众号 / WeChat Official Account editor.

Bilingual skill. Chinese version: SKILL.zh.md. English is the single source of truth; the .zh.md is a synchronized translation — always edit the English first, then mirror the change into .zh.md in the same change-set, never edit only the Chinese.

Trigger phrases

User says something like:

• "做一份 5 月下半月的中东观察"
• "Generate a semiconductor industry briefing for the past month"
• "做一期 AI 立法主题观察，时间是 2026 年 Q2"
• "/topic-brief"
• "用 topic-brief 跑一份 ..."

Any descriptive request such as "make a briefing / observation / digest on XX" should activate this skill.

Scope exclusions

Before producing any briefing, check whether the subject falls into a refused category. The following are out of scope regardless of phrasing:

• Political issues — elections, parties, political figures' positions, government legitimacy debates
• Military issues — operations, force composition, equipment counts, conflict details, defense policy
• Religious issues — doctrines, denominational disputes, religious leaders' statements
• Entertainment celebrity gossip — personal life, scandals, fan disputes
• Other inherently controversial topics — culture wars, identity disputes, value-judgement debates

If the requested subject is one of the above, reply with exactly one line and stop:

Out of scope. (超出能力范围)

Do not run search, do not draft anything, do not negotiate the scope.

Response language

Reply in the language of the user's request. Chinese question → Chinese reply; English question → English reply; mixed input → follow the dominant language. The generated HTML itself follows the language of the source materials and the subject_name (Chinese subjects produce Chinese briefings, English subjects produce English briefings). Field labels in user-facing reports (e.g., 来源 / Source) should match the briefing's body language.

5-step workflow

Step 1 — Collect parameters (one user-facing question prompt, 4 questions)

If the user did not provide all parameters in the trigger, ask everything in one batched user-question prompt:

• The subject field lands in JSON as subject_name (e.g., "中东", "半导体", "AI 立法")

• The author field lands in JSON as author, rendered at the cover bottom-left. Three user options map to three JSON states:

  User choice	JSON action	Rendered effect
  A keep default	omit author field	shows "developed by Gen" (schema default)
  B leave blank	write "author": ""	blank
  Other custom (e.g., "张三 · 研究院")	write "author": "张三 · 研究院"	custom text

• Within the same session, if the user has already answered the author question, reuse the last value for subsequent runs (unless they change it).

If the trigger already specifies every parameter ("做 5.1–5.12 的中东观察，作者署名 '张三'"), skip the question prompt and proceed.

Step 2 — Parallel material gathering

Every search query must include a time filter. For Google / Bing: after:YYYY-MM-DD before:YYYY-MM-DD (using the period_start / period_end normalized in Step 1); for other engines, use the equivalent syntax. Without a time filter, the engine returns results by relevance — older "big events" with high SEO weight get pulled in and stale the briefing.

Time policy:

• Focus body may reference earlier background context for contrast (e.g., "last year a similar report by X said..."), but must explicitly tag the time point (e.g., "in 2025 Q3..."), so readers can tell background from in-period events at a glance
• Items in the 4 sub-sections must strictly fall in-window (event_date within [period_start, period_end]). Out-of-window material is either cut or moved to the focus body as labeled background

Search-engine query at least 4–6 times, covering these dimensions (adapt to subject type):

Fetch web body 1–2 times to deep-read the focus report's primary text (extract specific numbers, scenarios, policy implications).

Default authoritative whitelist (unless the user specified otherwise in Step 1):

• International institutions: IMF, World Bank, OECD, ADB, UNCTAD, IADB, AIIB, IEA, WTO
• Central banks: Federal Reserve, ECB, Bundesbank, individual country central banks
• Governments: European Commission, national statistical offices
• Mainstream financial media: Reuters, Bloomberg, FT, Economist, Nikkei Asia, Xinhua, Business Standard

Step 3 — Direction confirmation (mandatory user check-in) ⚠️

Stop after material gathering and confirm direction before writing:

Report to user:
- Time window: [period_start, period_end]
- Proposed focus report: <title> by <institution>, published <YYYY-MM-DD> (URL: ...)
- Proposed 4 sub-sections with candidate items (each tagged with event date):
  A. <section> — 3–5 items:
     · [YYYY-MM-DD] <headline> · <source institution>
     · [YYYY-MM-DD] <headline> · <source institution>
  B. ...
  C. ...
  D. ...

Ask:
- Confirm the direction?
- Swap the focus?
- Adjust the 4 sub-sections?
- Any candidate items falling outside the window (check the dates)?

Every candidate item must show its event date — this gate lets both the user and the model spot-check freshness, preventing stale items from leaking into the final draft.

If the user adjusts, revise and confirm once more. Only after sign-off proceed to Step 4.

Skipping this gate causes 5,000-character rewrites when the focus turns out wrong.

Step 4 — Compose the JSON

Follow the schema and discipline in prompts/system.md:

Length and structure:

• Focus body 1,500–2,500 characters (3–5 sections, 200–500 chars each)
• 4 sub-sections × 3–4 items each (headline ≤30 chars + body 100–300 chars)
• summary: focus blurb 150–250 chars + 4 region_items one sentence each
• Total: 3,000–5,000 characters

Top discipline — no fabricated numbers Every concrete number (percentage, currency amount, date, count) must be traceable to the materials gathered in Step 2. After writing, re-read the draft and ask "where did this number come from?" at every figure. If you cannot answer, fix it.

Time-window discipline:

• Every sub-section item must carry an event_date field (YYYY-MM-DD; YYYY-MM is acceptable when only the month is known)
• Every event_date must fall within [period_start, period_end]
• When the focus body references earlier background, tag the time point explicitly ("in 2025 Q3..." / "last year...")

JSON quote discipline: Chinese inline quotes must use paired " and ", never straight " (breaks JSON parsing).

Style reference: reference/ contains 4 historical samples (3 regional + 1 red-brand institutional). Mirror their phrasing, cadence, and tone.

Save path: output/seed/<subject>_<period_end>.json so seed data is traceable.

Step 5 — Render and report

python3 scripts/render.py output/seed/<subject>_<period_end>.json --out output --open

• Auto-runs fix_quotes to repair Chinese quotation pairs
• Renders HTML to output/<period_end>_<title>.html
• Opens it in the browser

Reporting checklist:

• File path
• Character count / items per sub-section
• Source URL list grouped by the 4 sub-sections
• Reminder: human review of every number and URL is required before publishing — AI output ≠ conclusion

Tool-call mapping (cross-LLM adaptation)

This skill describes tool actions in generic semantic terms so non-Claude terminals can map them to their own toolset:

If the running terminal lacks an equivalent for one of these verbs (e.g., no interactive prompt), the LLM should degrade gracefully — for instance, ask the parameter questions as one plain message and wait for the user's reply.

Failure handling

Quick command reference

# Enter the skill directory (path inside the market-research-skills monorepo)
cd /path/to/market-research-skills/skills/topic-brief

# Render (auto-fixes quotes + opens browser)
python3 scripts/render.py reference/region_middle_east.example.json --out output --open

# Quote-fix only
python3 -c "from lib.fix_quotes import fix_file; from pathlib import Path; fix_file(Path('output/seed/xxx.json'))"

# Install dependency (first time)
python3 -m pip install --user jinja2

Self-check checklist (run after Step 4)

• issue_title ≤ 24 chars (including punctuation) — longer titles wrap on the cover, fix it
• Exactly 4 sub-sections, each with 3–4 items
• summary.items has exactly 4 entries, aligned with the 4 sub-sections
• Every number is traceable to a source from Step 2
• Every item.source has a full URL
• Every item.event_date falls within [period_start, period_end]
• If the focus body references out-of-window background, the time point is explicitly tagged (e.g., "in 2025 Q3...")
• item.headline ≤ 30 chars
• Focus sections 3–5; total focus body 1,500–2,500 chars
• Chinese quotes use paired " ", no straight "
• subject_name field is filled
• period_start / period_end filled with ISO dates

Red lines

• Do not fabricate any number, date, or quote. If the source material does not contain it, rewrite without the number or cut the item entirely.
• Do not use a non-whitelisted source as the focus report. Pop-econ blogs, social-media posts, and unattributed paraphrases cannot anchor the focus section.
• Sub-section items must be strictly in-window. Any item whose event_date falls outside [period_start, period_end] must be cut or moved to the focus body as labeled background (with an explicit time tag).
• Every search must include a time filter. Searching without after:/before: lets the engine return SEO-weighted older content — this is the root cause of stale briefings.
• Do not skip Step 3 (direction confirmation). Writing 5,000 characters in the wrong direction wastes the user's review budget.
• Do not produce a briefing on a refused subject (politics / military / religion / celebrity gossip / inherently controversial topics). Reply Out of scope. (超出能力范围) and stop.
• Do not pad the briefing with non-substantive narration ("令人震惊" / "标志着" / "势必"). Keep it factual, dense, source-anchored.
• Do not omit human-review reminder in Step 5. AI output is draft material, not a publishable conclusion.