SuperAPI AI Directory

social-opinion-writer

General & other

View repository

Updated 10-16 08:43

About this skill

Social Opinion Writer

A social conditions and public opinion writing skill for democratic party and CPPCC information scenarios.

The goal is not to write a general argumentative essay, but to form a problem-and-suggestion draft closer to "submittable material": focused topic selection, reliable facts, tight logic, actionable suggestions, and truly written from the perspective of decision-making reference for Party committees and governments.

What to read first

Applicable input

  • Only a topic, for example "how to alleviate the loss of county high school teachers"
  • Only some scattered materials, transcriptions, chat minutes, or research notes
  • There is already a draft, but it needs online materials added and rewriting into social conditions and public opinion
  • Need to quickly form a democratic party submission draft around a hot issue

Default settings

  • Default length:1500-2000 characters; when it is truly important information, it may be relaxed to 3000 characters
  • Default genre: social conditions and public opinion information
  • Default reporting level:市级及以上; give priority to forming a decision-making reference draft for the municipal, provincial, or national level
  • Default structure:现状/前言 -> 主要问题 -> 对策建议
  • Default research method: automatically supplement public information online, prioritizing authoritative sources
  • Default number of problems:3-4 items
  • Default number of suggestions:3-4 items, corresponding one-to-one with the problems above
  • Default verification rounds:2 rounds; for strongly time-sensitive, policy-sensitive, or data-intensive topics, increase to 3 rounds
  • Default final delivery:markdown + 引用表 + humanized-markdown + docx
  • Default output directory:social-opinions/{slug}-{timestamp}/

If the user explicitly says "no internet" or "only based on the content I provide," skip external research, but the boundaries of evidence must be clearly stated.

Core principles

  1. Information work is essentially decision-reference intelligence work. Truthfulness, accuracy, and verifiability take priority over literary flair and elaboration.
  2. The reporting target is Party and government leaders, not corporate clients or ordinary readers. Problems and suggestions must be raised from the perspective of Party committee and government decision-making and policy implementation.
  3. Topic selection is by default handled as "starting at the municipal level and above." Specific matters that districts, streets, and communities can handle on their own should in principle not be elevated into formal submission drafts, unless they can be distilled into institutional bottlenecks at or above the municipal level.
  4. One draft addresses only one core issue, and the submission level must be single; do not mix recommendations at the municipal, provincial, and national levels in the same draft.
  5. Write deeply on a small incision; do not write grand, empty macro declarations. The foreword is only responsible for introducing the problem and does not carry lengthy background setup.
  6. Facts, data, cases, and policy statements must be traceable and must not be fabricated; any data without a source shall be treated as unusable.
  7. The title should state the problem directly; do not write slogan-style or literary-style titles, and do not quote leaders' speeches for a formulaic opening.
  8. Recommendations must be able to land on a responsible entity, mechanism action, or policy tool, and correspond item by item with the problems above, forming a closed loop.
  9. The language should read like a working document for decision-makers, not like an academic paper, news commentary, or publicity piece.
  10. The title should be concise, accurate, and to the point, prioritizing a summary of the core risk, key blockage, or governance direction, and avoiding generic titles such as "Recommendations on xxxx" or "Recommendations on further strengthening xxxx".
  11. The problem section cannot contain only individual cases and qualitative judgments; the full text must reflect a certain amount of verifiable data support. Data need not overshadow the main point, but it cannot be completely absent.
  12. Chinese body text shall uniformly use Chinese punctuation and Chinese quotation marks; straight quotation marks in a Chinese context " must be changed to “”, and English half-width quotation marks must not remain in the final draft.

Output files

Create a directory for each task:social-opinions/{slug}-{timestamp}/

  • 00-brief.md: task constraints, topic, object, length, whether online access is needed
  • 01-terms.md: glossary for this task, recording recommended wording, prohibited wording, and first-explanation method
  • 02-search-plan.md: search questions, keywords, preferred sources
  • 03-source-log.md: source list, original text URL, publication time, credibility, usable key points, for manual review
  • 04-evidence.md: fact matrix, distinguishing verified facts, judgments, and points to be confirmed
  • 05-outline.md: incision, title alternatives, structure
  • 06-draft-v1.md: first draft
  • 07-review-round-1.md: first-round self-check
  • 08-draft-v2.md: revised draft
  • 09-review-round-2.md: second-round self-check
  • 10-final.md: final draft
  • 11-references.md: citation table, summarizing all sources actually used in the final draft body, including original text URLs, corresponding one-to-one with body superscripts
  • 12-final-humanized.md: call humanizer-zh final draft after removing AI flavor
  • 13-final.docx: Word final draft exported in a format commonly used for mainland China official documents

If a third round of verification is done, append:

  • 10-review-round-3.md
  • 11-final.md
  • 12-references.md
  • 13-final-humanized.md
  • 14-final.docx

If the user only needs the result in the chat window, still save the final draft file as usual and provide the path in the reply.

Markdown formatting conventions

For the .md files output at each stage, by default use Markdown tables as little as possible, and prefer vertically arranged card-style structures, to facilitate review in the chat window, reading on mobile devices, and subsequent manual additions and deletions.

Default writing method:

  • For each stage, first give a brief title, then split it into several second-level or third-level sections by theme
  • Each record should be made into an independent card as much as possible, organized using ##、### or bold subheadings
  • Within cards, prefer short paragraphs and bullet points, avoiding wide tables with too many horizontal fields
  • Use tables only when "multiple objects must be compared horizontally and there are very few fields"; even then, keep them within 3-4 columns
  • When involving sources, terminology, evidence, problem-recommendation mapping, and self-check conclusions, prefer writing them in a vertical format of "one card per object"
  • Do not pursue flashy visual effects; the so-called "card-style" only means clear vertical sections, fixed fields, and easy quick scanning, not HTML cards

For a general card template, refer to:

## 卡片标题

- 字段一:内容
- 字段二:内容
- 字段三:
  这里可补一小段说明。
- 字段四:
  - 要点 A
  - 要点 B

Workflow

Step 1: Refine the task, write to 00-brief.md

Prioritize extracting directly from the current conversation; do not repeatedly ask follow-up questions just to fill in defaults. At minimum confirm:

  • What is the topic or issue
  • Submission level: municipal, provincial, national, or determined after judging the materials
  • Submission scenario: democratic parties, CPPCC, united front departments, internal reference, or a general-purpose draft
  • Whether internet access is required
  • Whether the latest policies, latest statistical data, or local cases are required
  • Target region: nationwide, a certain province, city, county, or unspecified
  • What materials the user already has
  • Length or style requirements

If the user gives only one broad topic, first automatically narrow it into 2-3 sub-angles that can be submitted, and choose the most specific one, the one easiest to obtain evidence for, and the one with the most actionable recommendations to continue writing. Only ask the user follow-up questions when the risks of several sub-angles differ greatly.

Additionally perform two topic-screening actions:

  • First judge whether the issue reaches the submission value of "starting at the municipal level or above." If it is merely a matter that can be handled at the district or street level, such as broken traffic lights, details of garbage sorting implementation, or a property dispute in a single community, do not directly draft it; first elevate it to a blockage at the level of institutions, policies, or coordination mechanisms; if elevation fails, clearly state that "it is not suitable as a public opinion and social conditions topic at this level."
  • First fix the manuscript level. If it is determined to be a municipal-level manuscript, develop it around municipal authority, cross-district common issues, and citywide policy implementation problems; do not write municipal-level blockages in half the piece and then propose national-level legislative recommendations in the latter half.

00-brief.md Prefer writing it as a one-page summary card, not a "large table with many fields." It is recommended to include at least:

  • Task topic
  • Proposed sub-angle
  • Submission level
  • Submission scenario
  • Whether internet access is required
  • Whether the latest policies/data are required
  • Target region
  • Existing materials
  • Length and style constraints
  • Current risks or points to be confirmed

Step 2: Unify terminology, write to 01-terms.md

When any of the following situations occurs, terminology must first be unified before proceeding to retrieval and drafting:

  • The topic involves policy proper nouns, institutional names, special actions, or reform formulations
  • It involves names of institutions, departments, democratic parties, or CPPCC systems
  • It involves statistical calibers, indicator names, or professional governance terminology
  • It contains English abbreviations, foreign concepts, or industry jargon
  • What the user provides is colloquial material, meeting minutes, or interview records, and the terms in them are obviously nonstandard

Processing order:

  1. When it involves democratic parties, the CPPCC, the united front, or expressions of duty performance, first read references/glossary-policy-org.md
  2. Then read the general terminology list as needed references/glossary-zh-en.md
  3. When it involves high-frequency livelihood or governance topics, additionally read references/policy-topic-map.md
  4. Extract the key terminology for this task from the user's materials and the proposed retrieval topic
  5. Form the session terminology list for this task and write it to 01-terms.md
  6. Mark the recommended wording, prohibited wording, whether first explanation is needed, and whether to retain English or abbreviations
  7. Subsequent retrieval, first draft, self-check, and final draft all take 01-terms.md as the standard, and there must be no drift back and forth

Execution principles:

  • For institutional names, policy names, and legal names, use the official full names first
  • Common abbreviations may be used alone only after the full name has been given at first occurrence
  • When an English abbreviation appears for the first time, first use 中文全称(英文缩写)
  • For words prone to ambiguity, do not frequently switch to synonymous expressions just to make the writing more varied
  • If policy-topic-map.md lists a relatively recent central document, use that name first; do not fall back to an older colloquial wording
  • Nonstandard expressions in the user's original materials may be retained as clues, but must not be copied directly into the final draft

01-terms.md It is preferable to write according to "one term, one card" rather than organizing by default into a large table. Each card should at least include:

  • Term
  • Recommended wording
  • Prohibited or discouraged wording
  • Whether explanation is needed at first occurrence
  • Notes

Step 3: Establish a search plan and write it into 02-search-plan.md

The search objectives should cover at least four types of information:

  1. Whether the problem truly exists and what the symptoms are
  2. Where current policies, systems, or implementation processes are stuck
  3. Whether there is authoritative data, typical cases, or local practice
  4. Which recommendations have already been widely made, how to avoid vague repetition, and whether similar policies have already been issued

Generate for each topic:

  • 5-10 Chinese search term combinations
  • Required sources: policies and regulations, competent departments, statistical definitions, public information from the CPPCC and political parties, and typical local cases
  • Time boundary: prioritize the recent 1-3 years; for highly time-sensitive issues, prioritize the recent 12 months
  • Evidence, data, and cases that are too far from the current time are generally not recommended for use unless they are still-effective laws and regulations, national standards, long-term foundational systems, or irreplaceable historical background materials
  • For each recommendation to be drafted, reserve a "check whether a policy has already been issued" search item to avoid rewriting existing policies as recommendations
  • If the topic involves sensitive scenarios such as accidents, extreme events, personal injury, minors, or public opinion controversies, search terms should prioritize neutral or official governance expressions and avoid directly using highly sensitive words such as "death" or "fatalities"; preferably rewrite them as search paths such as "personal injury", "safety accident", "serious consequences", "emergency incident", "aftermath handling", "risk hazards", and "safety governance"
  • If neutral search terms are already sufficient to locate authoritative sources, do not add more sensitive explicit expressions; retain the original wording only when checking the original title of a source document, the text of a regulation, or the original official notification

02-search-plan.md It is recommended to organize by "one search question, one card", and each card should clearly state:

  • Search objective
  • Keyword combination
  • Preferred sources
  • Time boundary
  • Check item for already-issued policies
  • Alternative wording for sensitive expressions

Step 4: Search for materials online and write them into 03-source-log.md

Source priority follows the order below:

  1. Party and government agencies, laws and regulations, policy documents, and official replies
  2. Official websites of the National Committee of the Chinese People's Political Consultative Conference, local CPPCC committees, and central or local organizations of democratic parties
  3. National Bureau of Statistics and local statistical bulletins, yearbooks, and public data platforms
  4. Reports from authoritative media and professional institutions
  5. High-quality academic papers or industry research

Execution requirements:

  • Collect at least 5 valid sources; for complex topics, strive for 8-12 items
  • For each source, record: title, original URL, publication date, credibility, and extracted key points
  • 03-source-log.md must retain original URLs that can be opened directly; prioritize recording the original publication page URL, and do not write only the site name, search page, aggregation page, or secondary repost link
  • If redirect links, short links, or portal repost pages exist, additionally record the original publishing unit's page URL; if the original text cannot be located, treat it as low credibility or unavailable
  • Optionally record the first access time, access notes, or archive links, but these must not replace the original URL
  • For the same fact, try as much as possible to achieve cross-verification with two sources
  • Separate "facts," "opinions," and "inferences"
  • Self-media, forums, and short-video comments can only serve as clues and cannot be directly used as evidence
  • Even if the content appears on platforms such as Sina, NetEase, or Tencent, its original source must still be judged; if it is essentially self-media reposting, has no original source, or is obviously AI-assembled content, it should still be treated as unavailable
  • Links, cases, statistics, and years provided by AI tools must be manually reviewed one by one; anything whose original source cannot be found or whose time is obviously outdated must be deleted
  • For materials whose publication time is obviously too early and for which updated wording or updated data already exists, in principle use the new rather than the old; if old materials really must be retained, it must be explained that they only serve as historical background or institutional evolution context and are not used as the core basis for judging the current situation
  • If sufficiently authoritative sources cannot be found, lower the strength of the conclusion; do not force the writing

03-source-log.md Recommended template:

## 来源 1:来源标题

- 原文 URL:
- 原始发布单位:
- 发布日期:
- 首次访问时间:
- 可信度:高 / 中 / 低
- 可用要点:
  - 要点 1
  - 要点 2
- 备注:
  - 是否原始发布页
  - 是否转载页或短链跳转页
  - 是否需要二次核验

Completion requirements:

  • 原文 URL Required, and it should be directly openable for manual review
  • 原始发布单位 Required, to facilitate judging the wording and authority
  • 可信度 At least mark as 高/中/低
  • 备注 Indicate whether it is an original publication page, repost page, or short-link redirect page, and whether the original link has been found
  • Unless the user explicitly requests an overview summary table, do not pile all sources into one large horizontal table

Step 5: Build an evidence matrix and write it into 04-evidence.md

Do not default to a three-column table. Prioritize organizing vertically in blocks:

  • Verified facts
  • Analytical judgments that can stand
  • Points that still have uncertainty

At the same time, add a section judging what "can be written or cannot be written":

  • Can be written: verified, broadly applicable, and relevant to decision-making
  • Write with caution: individualized, emotional, and lacking public evidentiary support
  • Do not write: unverifiable data, exaggerated statements, and political judgments that easily cause misreading

For each data point and case, add two more pieces of metadata:

  • Source: original publishing unit, title, publication date, link, or traceable identifier
  • How it is used: as an introduction to the current situation, support for the problem, or basis for recommendations

04-evidence.md It is recommended to write it into the corresponding group as "one evidence point, one card," for example:

## 已核实事实

### 证据卡 1
- 事实:
- 来源出处:
- 使用方式:
- 备注:

## 可以成立的分析判断

### 判断卡 1
- 判断:
- 依据:
- 使用边界:

## 仍有不确定性的点

### 待核卡 1
- 待确认点:
- 当前依据:
- 风险:
- 处理意见:

Step 6: Choose the angle and structure, and write it into 05-outline.md

At minimum output:

  • 2-3 alternative titles
  • One-sentence proposition: what change this draft is ultimately meant to promote
  • Submission level and corresponding responsible entities
  • Preface/current situation summary
  • 3-4 problem subheadings
  • Correspondence between problems and recommendations
  • Recommended framework
  • The data anchor or statistical basis corresponding to each problem; at minimum indicate which paragraphs will include data, to avoid the main text becoming only a pile of cases

Prioritize using this type of angle:

  • execution-level breakpoints
  • high-frequency pain points of the public
  • blockages in the implementation of new policies
  • new situations not covered by existing mechanisms
  • institutional gaps in which a certain group of people is overlooked

Avoid:

  • topics that are too broad, for example "comprehensively advancing the modernization of education"
  • pure value statements with no specific problems or mechanisms to act on
  • recommendations that do not correspond to the problems
  • having municipal-level implementation recommendations, provincial-level fiscal recommendations, and national legislative recommendations appear in the same outline at the same time

Before drafting recommendations:

  • first read by topic references/recommendation-playbooks.md
  • from which select 2-4 of the most matching recommendation anchors; do not write a single draft as a complete compendium of policy tools
  • retain only recommendations that directly correspond to the chain of problems in the preceding text
  • clearly establish 问题一 -> 建议一、问题二 -> 建议二 mapping; it is not allowed for the preceding text to state four problems while the later text has only three general recommendations

05-outline.md Do not default to making a "problems-recommendations" comparison table. Give priority to writing it as a vertical mapping card, for example:

## 问题卡 1

- 问题标题:
- 核心表现:
- 证据抓手:
- 对应建议方向:
- 对应责任主体:

## 问题卡 2

- 问题标题:
- 核心表现:
- 证据抓手:
- 对应建议方向:
- 对应责任主体:

Title drafting requirements:

  • Titles should give priority to summarizing "problem + risk/consequence" or "blockage + governance direction," rather than being written as vague topics
  • Try to avoid ordinary official-document-style titles such as "Recommendations on...", "Recommendations on further strengthening...", and "Report on the situation of..."
  • If the nature of the problem can be directly pointed out, give priority to directly stating the point; there is no need to add the word "recommendations" separately
  • Prompting words such as "urgently needs," "remains to be," "vigilance," "guard against," "break the deadlock," and "attention" may be used appropriately, but only on the premise that the judgment is supported by evidence and does not bluff
  • Titles should be controlled as much as possible within 22 characters; in special circumstances, no more than 25 characters
  • Titles should match the reporting level; do not have the main text discuss municipal-level implementation problems while the title is elevated to the national strategy level

Title style reference:

  • AI手机智能体亟待加强安全、隐私与生态协同治理(problem judgment sentence)
  • 警惕私域直播围猎银发群体(risk warning sentence)
  • 谨防”沉默企业”成为阻碍市场经济发展的”灰犀牛”(risk metaphor)
  • 税收政策复杂及服务短板制约民营企业发展有待破局(blockage description)
  • 关注跨国生物医药企业在华战略调整带来的系统性风险(attention + characterization of risk)

Avoid this type of title:

  • 关于加强 AI 手机智能体治理的建议
  • 关于私域直播问题的建议
  • 关于进一步促进民营企业发展的建议

Step 7: Write the first draft, save it to 06-draft-v1.md

By default, draft according to the following template:

# 标题

【现状与前言】
用简明权威数据、主管部门通报或典型事件把问题引出来,篇幅控制在全文的 `10%-20%`。只说明“为什么值得报”,不写长背景,不引用领导人讲话,不做穿靴戴帽。

【主要问题】
一、问题小标题
用事实、数据、案例说明问题表现,并点出机制堵点或执行断点。

二、问题小标题
继续用可核实证据展开,不写空泛判断。

三、问题小标题
如确有必要可写到四个,但不要为了凑数扩写。

【对策建议】
一、对应问题一的建议
二、对应问题二的建议
三、对应问题三的建议
建议部分可占全文约 `40%`。每条建议尽量写清责任主体、动作机制、政策抓手和预期效果。

Writing requirements:

  • Titles should be controlled as much as possible within 22 characters; in special circumstances, no more than 25 characters
  • Titles should be concise and accurate, giving priority to writing them as "problem judgment sentences" or "risk warning sentences," and using ordinary official-document tone less
  • The foreword must be brief and serve only to introduce the topic
  • The main text should use short sentences as much as possible and fewer formulaic expressions
  • Do not write a literature review or provide lengthy background buildup
  • For the problems section, give priority to writing 3-4 items; each problem must have a subheading and evidence support
  • The problems section must not be composed entirely of individual cases, media cases, or abstract judgments; the full text must reflect at least 2 pieces of verifiable data-based information, preferably placed in the foreword or problems section
  • Data and cases may be used together; it is not required that every problem be dominated by data; however, as long as usable statistics, administrative notices, monitoring results, or industry quantitative measures exist in public materials, they should be appropriately included
  • Focus on facts and mechanisms, not emotional drive
  • It is recommended to sort by "most important, most implementable" and to strictly correspond one-to-one with the issues.
  • It is recommended to write in the form of "who does it, what is done, through what mechanism it is done, and how the effect is measured."
  • In-text citations in the main text shall use superscript numbers: whenever specific data, statistics, typical events, or policy statements are involved, a superscript number shall be marked at the end of the cited content and before the punctuation [1]、[2]... (half-width square brackets + Arabic numerals). Multiple citations from the same source share the same number, and the numbers are arranged in order of first appearance and correspond one-to-one with the citation table 11-references.md . The main text no longer uses long parenthetical citations, does not use footnotes, and does not write URLs directly into the main text; source details are uniformly collected in the citation table.
  • The Chinese main text uniformly uses full-width Chinese punctuation; quotation marks in Chinese context are uniformly “”, book titles are uniformly 《》, and parentheses, enumeration commas, commas, and periods preferably use Chinese forms, without retaining English half-width double quotation marks.
  • It is recommended to use a layered writing style:(一)/一、 Under major items, embedded 一是……二是……三是…… may be used to list several specific measures. This is a standard official writing style and does not constitute prohibited multi-level nesting. What is prohibited is American-style hierarchical lists (such as 1.1.1、A-a-i) or parenthetical nesting of more than three levels.
  • Do not use the "large ABC, small abc" style multi-level nested structure; keeping it within two levels is sufficient.
  • If the user provides ready-made content, retain the valuable first-hand observations in it, but convert colloquial material into submission language.

Step 8: First round of self-check, write into 07-review-round-1.md

In the first round, only look at "whether it can stand"; do not look at wording. At minimum, check:

  • Whether the title accurately corresponds to the issue
  • Whether the title avoids ordinary generic titles such as "Suggestions on..."
  • Whether one issue is discussed per submission
  • Whether the submission level is clear and consistent throughout
  • Whether facts, figures, policy names, and institution names are accurate
  • Whether every issue has source support, and whether the sources can be traced back
  • Whether the issue section reflects the necessary data support, rather than relying only on individual cases and qualitative judgment
  • Whether individual phenomena have been written as universal laws
  • Whether the suggestions target the issues above, rather than starting a new line of discussion

If there are hard flaws, directly rewrite the relevant paragraphs, then generate 08-draft-v2.md。

07-review-round-1.md It is recommended to write in the form of "check item cards + conclusions," not a large checklist table. Each item may be written as:

  • Check item
  • Whether it passes
  • Problems found
  • Revision action

Step 9: Second round of self-check, write into 09-review-round-2.md

In the second round, look at "whether it resembles high-quality public sentiment and public opinion." At minimum, check:

  • Whether it achieves a small entry point, real problems, and a short path
  • Whether the title is concise, accurate, and to the point, and whether it can directly reflect problem awareness
  • Whether the preface is controlled to be short enough, without empty formalities
  • Whether the number of issues is controlled at 3-4 , and each issue has a subheading
  • Whether responsibility is generalized to empty subjects such as "society" and "the masses"
  • Whether the suggestions are executable, assignable, and evaluable, and correspond one-to-one with the issues
  • Whether the suggestions repeat existing policies, and whether there is a situation of "already issued but still proposed as a suggestion"
  • Whether the language resembles official work writing, rather than an academic paper, commentary, or publicity article
  • Whether Chinese punctuation is consistent, and whether English double quotation marks, half-width parentheses, or half-width punctuation still remain in a Chinese-language context
  • Whether there is novelty, and whether it obviously looks like an 'old topic rewritten'

After revising, output the final draft.

09-review-round-2.md Continue the card-style writing, and record separately by card under 'structural quality', 'suggestion quality', 'language quality', and 'compliance'.

Step 10: Conduct a third round of verification when necessary

A third round must be added when the following situations occur:

  • It involves the latest policies, the latest statistical data, or recent public opinion
  • It involves high-attention livelihood issues such as education, healthcare, employment, housing, and the business environment
  • It is highly local, and the user requests specificity to a certain place
  • The topic evidence is not solid enough, but a usable first draft still needs to be produced

The third round focuses on checking:

  • Whether the time is outdated
  • Whether the argument is excessive
  • Whether the suggestions exceed the current policy authority
  • Whether the data, cases, and links contain AI hallucinations or pass off secondhand retellings as firsthand sources
  • Whether there is a more prudent way of expression

Step 11: Output the final draft and citation table

Output three parts by default:

  1. 定稿正文
  2. 引用表
  3. 核验摘要

定稿正文 Write into 10-final.md, keeping only [n] superscript markers in the main text, without URLs.

引用表 Write separately into 11-references.md, and present it separately from the final draft main text in the chat output. Requirements:

  • Include only the sources actually cited and marked in the final draft main text, with numbering strictly consistent with the superscript markers in the main text, no more and no less
  • Sources are based on 03-source-log.md , and no new entries may be invented; sources not cited in the main text do not enter the citation table
  • Each item must include a directly openable original URL, preferably the original publishing page
  • Arrange vertically as 'one source, one card', not piled into a large horizontal table

引用表 Recommended template:

## 引用表

[1] 《来源标题》,原始发布单位,YYYY-MM-DD
- 原文 URL:https://...
- 引用位置:前言第 1 段;问题二
- 可信度:高

[2] 《来源标题》,原始发布单位,YYYY-MM-DD
- 原文 URL:https://...
- 引用位置:问题一;建议一
- 可信度:中

核验摘要 Keep it brief, explaining only:

  • What types of public sources this draft is based on
  • Which points have already been verified
  • Which points remain prudent judgments

核验摘要 Also prefer short cards or short lists, and do not add another overview table.

Add a closing action before output:

  • For final.md , do a Chinese punctuation normalization check; if a file is placed in the workspace, execute python3 scripts/normalize_cn_punctuation.py --input <final.md> --in-place
  • Focus on checking English double quotation marks, half-width parentheses, and half-width punctuation in a Chinese-language context, to avoid bringing uncleaned English punctuation into the chat output and formal draft

If the user only wants the formal draft, only the final draft main text may be shown, but internal files must still be retained.

Step 12: Call humanizer-zh to remove the AI flavor

Before generating docx , by default first call the humanizer-zh skill to perform a 'remove AI flavor' pass on the final md final draft, and save it as final-humanized.md。

Execution requirements:

  1. The input file uses the final draft final.md
  2. When calling humanizer-zh , clearly limit the goal to removing AI cliches, promotional tone, stacked connectives, mechanical three-part structure, and obvious model-like traces
  3. It is not allowed to change formal submission materials into colloquial essays, and it is not allowed to introduce first person, slang, humor, self-expression, emotional judgments, or internet-savvy expressions
  4. After removing the AI flavor, immediately review 01-terms.md、04-evidence.md and the final draft, and confirm that terminology, facts, data, policy names, and submission tone have not drifted
  5. In the main text, [n] Superscript markers are citation markers and are not AI traces; when removing AI flavor, they must be retained as-is and must not be deleted, rewritten, or moved.
  6. After removing AI flavor, perform Chinese punctuation normalization once more, especially checking whether Chinese quotation marks have been restored to English straight quotation marks.
  7. If humanizer-zh modifications would weaken formality, accuracy, or the tone of decision reference, the relevant changes should be reverted, and it is preferable to retain a more restrained work-document style.

The goal of this round of processing is not to "write more like a personal blog," but to make the text less mechanically generated while still reading like a work document that can be submitted.

Step 13: Generate docx final draft

Unless the user explicitly says "no Word needed" or "keep only markdown," after finalization the default is to additionally generate a docx file.

Execution order:

  1. Read references/docx-format-cn-official.md
  2. First call humanizer-zh to generate final-humanized.md
  3. For final-humanized.md perform Chinese punctuation normalization:python3 scripts/normalize_cn_punctuation.py --input <final-humanized.md> --in-place
  4. With final-humanized.md as input, run python3 scripts/generate_docx.py --input <final-humanized.md> --references <references.md> --output <final.docx>; in the body text, [n] superscript markers will be rendered as true superscripts in Word, and a "Appendix: Citation Sources" section will be automatically appended at the end
  5. In the reply, provide the docx file path; if fonts fall back due to missing fonts in the local environment, explain this as well.

Layout requirements follow "GB/T 9704-2012 page setup baseline + common typesetting habits for mainland China government materials." Note: social conditions and public opinion information is not equivalent to a formal Party and government official document with a masthead, document number, and record; therefore, by default only the body layout suitable for submitted materials is applied, and no red header, document number, record, or seal position is added without authorization.

Default layout:

  • A4 portrait
  • Margins: top 37mm, bottom 35mm, left 28mm, right 26mm
  • Title: centered,二号, prefer 方正小标宋简体, fall back to 宋体
  • Body:三号, prefer 仿宋_GB2312
  • Level 1 heading:一、 format,三号黑体
  • Level 2 heading:(一) format,三号楷体_GB2312
  • Body first-line indent 2 characters
  • Line spacing: fixed value 28pt
  • Space before and after paragraphs:0
  • Leave 1 blank lines below the title; by default, no extra blank lines between body paragraphs

If the user provides their own unit template or explicit layout requirements, the user's template takes priority; do not override the user's existing template with the default layout.

Online research constraints

  • When the user requests "latest," "current," "this year," or "recent," you must verify online and must not write from memory.
  • Prioritize original policy texts, original statistical tables, and government department notices; do not look first at secondhand interpretations.
  • When factual conflicts are found, prioritize higher-level, newer, and more clearly defined sources.
  • Do not package content that lacks publicly available evidence as established fact.
  • Bases and data too far from the current time should generally not be used unless necessary; prioritize materials from the last 1-3 years, and for strongly time-sensitive topics prioritize materials from the last 12 months
  • Only when citing laws and regulations, national standards, foundational institutional arrangements, or necessary historical evolution that remain in effect should older materials be retained; and old materials must not serve as the primary evidence for current judgments.
  • In the retrieval phase, apply "de-sensitized wording" by default: first search using neutral, governance-oriented, and official expressions, then decide whether it is necessary to look at the original title or source text.
  • Avoid using words such as "death" or "the deceased" as preferred search terms; give priority to alternative expressions such as "personal injury," "safety incident," "accident response," "risk prevention," and "aftermath work."
  • However, source records must remain accurate: if an official document, bulletin, or original report title itself contains the relevant words,03-source-log.md it should still be recorded faithfully according to the original title, and must not be rewritten without authorization.

Default rewriting strategy

The user provides only a topic

Research first, then write. Do not directly fabricate an article that "looks like" one when evidence is insufficient.

The user provides source materials

First extract verifiable facts and firsthand observations, standardize terminology and naming, then supplement with public information, and finally reorganize it into the standard public opinion and social conditions structure.

The user provides a draft

First diagnose the problems, then supplement the evidence, then rewrite. Do not merely polish the original draft mechanically.

Finalization standards

Only when all of the following conditions are met may it be judged final:

  1. The topic is clear and singular
  2. Factual and policy statements can withstand review
  3. Problems and recommendations correspond one to one
  4. Recommendations are realistically actionable
  5. Proper nouns, policy terminology, and institutional abbreviations are consistent throughout
  6. The writing is concise, with academic tone, propaganda tone, and empty clichés basically eliminated
  7. The reporting level is clear, and district or street affairs are not forcibly written as decision-making issues at or above the municipal level
  8. The full text contains no unsourced data, fabricated cases, or fabricated links
  9. The superscripts in the main text and the 11-references.md citation table correspond one to one, and each item in the citation table contains a directly openable original URL

If this cannot be achieved, continue revising, and do not pass off content that is obviously still a "draft" as final.

Example triggers

Example 1 Input: Please write a public opinion and social conditions report around "the difficulty of mutual recognition of examination and test results in primary-level hospitals," and supplement it with the latest policies and cases.

Example 2 Input: I will give you a set of research notes, and you help me organize them into public opinion and social conditions information submitted by a democratic party.

Example 3 Input: With "the outflow of teachers from ordinary county-level high schools" as the topic, automatically search for materials, produce a draft, and after repeated self-checking, give me the final version.