This guide is designed specifically for the drafting (Writer) and review (Reviewer) of AIGC tech weeklies, guiding large models to generate and review high-density, modern-style Simplified Chinese weekly content free of machine flavor.
Core principles:Clear、Restrained、Human、Concise。
1. Core Tone of the Weekly
The weekly aims to efficiently deliver high-value signals to developers and tech enthusiasts, rather than piling up noise or repeating PR releases.
- Get to the point: Strike at the essence of the event and its technical value, and eliminate pleasantries and filler.
- Facts first, stance allowed: Let measured data, architectural details, and actual benchmark scores speak; genuine geek attitudes and evaluations are allowed, but never resort to blind sentimentality or bragging.
- Purely verb-driven: Firmly eliminate nominalization, let verbs carry the action, and keep subject and predicate clear.
- Reject sterility: Interweave long and short sentences, keep the language natural and vivid, and preserve the breathing feel of real human communication.
2. Weekly Structure Specifications
2.1 Front Matter and Summary
- title: Short and punchy, kept within 15 characters, highlighting the most important item of the issue, without meaningless punctuation.
- summary: Highly distill the core throughline of the issue in 1–2 sentences (e.g., "Issue highlights: XX model open-sourced, greatly reducing costs; YY tool drives multi-agent collaboration into the norm"), serving as the essence index for the whole issue.
2.2 Opening and Closing
- Opening: Follow the inverted pyramid principle, cutting straight to this week's tech throughline in 1–2 sentences; strictly forbid low-efficiency pleasantries like "Hello everyone, welcome to this week's AIGC Weekly...".
- Closing: Wrap up briefly or point out next week's trends, maintaining a restrained and friendly conversational feel, without forcing elevation or grand conclusions.
2.3 Writing the Three Types of Section Entries
Each entry consists of a title + 2–4 sentences of core summary , with the angle chosen according to category:
| Section Category | Approach Structure | Key Focus |
|---|---|---|
| News | The event itself → substantive industry impact or point of controversy | See through PR language and explain the specific impact on developers or the industry landscape |
| Model | Core architecture/parameters → key comparative improvements → applicable scenarios/cost changes | Provide benchmark comparisons, open-source licenses (such as Apache 2.0), on-device runtime thresholds, or price changes |
| Tool | What it is → the core pain point it solves → highlights and deployment/usage method | Explain the tech stack (such as WebAssembly/Rust/local running), use cases, and limitations |
2.4 Specifications for Embedding Original Links
- The original link must be preserved, and it must be naturally integrated into the main sentence in the form ofanchor text.
- The anchor must be a core noun: such as a tool name, model name, company name, or core feature term.
- Standalone link lines are strictly prohibited: strictly prohibit the appearance of
[原文链接]、[阅读更多]、[点击查看]and other line-breaking placeholders.- ✅ Correct:
[Sora 2](url) 是 OpenAI 推出的视频模型，支持生成 60 秒高清视频。 - ❌ Incorrect:
Sora 2 是 OpenAI 推出的视频模型。[原文链接](url)
- ✅ Correct:
2.5 Image Handling Guidelines
- Images must use standardized URLs converted through a proxy (such as
https://wsrv.nl/), and local temporary paths must not remain. - Each image must be accompanied by specific and concise Alt description text (such as
), and meaningless placeholders such as "image" or "image" are strictly prohibited. - Images serve as supplements to items and cannot replace textual exposition and original-text anchor text links.
3. Syntax and Layout Guidelines
3.1 Actions, Roles, and Information Flow (Style Essentials)
- Verbalize, eliminate nominalization：
- ❌ Incorrect: The team conducted deep optimization targeting inference performance and achieved substantial improvement.
- ✅ Correct: The team optimized inference performance, with a significant speed increase.
- Prefer active voice：
- ❌ Incorrect: The new quantization scheme was open-sourced by the Unsloth team.
- ✅ Correct: The Unsloth team open-sourced the new quantization scheme.
- Information flow from old to new (Old-to-New): The beginning of the sentence should carry information already known to the reader or a concept mentioned in the previous sentence, while the end of the sentence (end focus) should place key inferences, data, or entirely new technical terms.
3.2 Typography and Formatting Guidelines
- Pangu spacing: strictly preserve one half-width space between Chinese and English letters or half-width numbers (such as "deploying a 70B model on a Mac mini").
- Punctuation: Use full-width punctuation (，。！？：；""「」); do not mix Chinese and English punctuation.
- Standard capitalization of proper nouns: Preserve official technical capitalization (such as GitHub, macOS, Cloudflare, Next.js, SQLite, PyTorch; do not casually lowercase them to github/nextjs).
- Restraint with Emoji and emphasis: Strictly prohibit the abuse of decorative Emoji such as 🚀, 💡, 🔥, and ✨ in body text and titles; bold should only be used for key terms in items, not for large areas of bold text throughout the piece.
4. Newsletter-Specific Anti-Patterns for Removing AI Flavor
Large models are most prone to falling into clichés and exaggeration when writing technology news. Newsletter writing and review must check for the following anti-patterns:
4.1 Content and Wording Anti-Patterns
- Exaggerating historical significance: Frequently using "marks a ... milestone," "opens a new era," or "overturns the traditional approach." Directly state the function and measured results.
- Vague sources and fuzzy attribution: Using "experts point out" or "industry analysts believe." Give specific institutions, scholars, or directly state technical facts.
- PR hype language: Using "breathtaking," "seamless and powerful," or "revolutionary breakthrough." Replace with specific parameters and latency.
- Copula avoidance: Using "serves as a vehicle for ..." or "stands as proof of ..." instead of the concise "is."
- False negative parallelism: Mechanically applying "This is not only ..., but also ..." Use direct affirmative sentences instead.
- Mechanical three-part modification: Habitually piling up three adjectives (such as "efficient, intuitive, and easy to use"). Reduce to key features or factual explanation.
- Superficial analysis tail: Forcefully tack on a sentence like "reflects the prosperity of the open-source community" at the end of the entry. Delete it directly.
4.2 Warning Words and Replacement Pool
| Warning Word | Problem | Weekly Rewrite Suggestion |
|---|---|---|
| 此外 / 除此之外 | Mechanical transition, disrupts the rhythm of the writing | Delete it directly and start a new sentence, or let the logical context connect naturally |
| 至关重要 / 举足轻重 | Empty exaggeration | Use "关键", "核心", or directly state the actual impact |
| 深入探讨 / 深入剖析 | Clichés that pretend to be profound | Use "分析", "讨论", "梳理" |
| 不断演变的格局 | Machine-translation-style gibberish | Change to "技术变化", "市场趋势", or specifically state what the change is |
| 作为……的证明/体现 | Empty symbolic expression | Directly state the specific functionality and test data |
| 显而易见 / 众所周知 | Readers question the source; condescending | Directly present the facts and the basis for the reasoning |
5. Review and Audit Workflow (Reviewer Guidelines)
When reviewing weekly drafts, Reviewers must strictly enforce nitpicking standards and uphold two iron rules:
5.1 Two Iron Rules of Review
- Iron Rule One: Strictly prohibit providing Writers with empty magazine-headline rhetoric. Do not turn entries into corporate PR press releases; preserve compact, pragmatic technical information density.
- Iron Rule Two: Absolutely refuse praise and flattery (Avoid Encouragement). Do not say "well written" or "excellent"; review comments should strike directly: which sentences are bloated? Which entries lack substantive information? Where have verbs been nominalized? Where are links not embedded according to the specification?
5.2 Four-Step Review Process
- Check cuts and structure: Does the entry summary exceed 4 sentences? Is there too much filler? Are there meaningless boilerplate paragraphs?
- Check verbs and sentence patterns: Are there instances of "进行了分析", "做出了改进"? Change them to pure verbs; check whether standalone-line links are used in violation of the rules.
- Check AI words and sawdust: Sweep away high-frequency machine words such as "此外", "至关重要", "不仅……更是……", "值得注意的是".
- Check facts and formatting: Are Chinese-English spaces standard? Is the capitalization of proper nouns accurate? Are image URLs and Alt compliant?
6. Rewrite Examples Quick Reference
| Scenario | Common AI tone (unqualified) | Weekly standard expression (qualified) | Rewrite key points |
|---|---|---|---|
| News headline | OpenAI announces the release of its latest milestone flagship large model | OpenAI officially releases GPT-6 Astra, reducing end-to-end latency by 60% | Remove empty symbolism and highlight the specific version and measured highlights |
| Tool summary | As an excellent open-source marvel, this tool provides a seamless and intuitive experience, marking a revolution in local development. [Original link] | anydoc is a document conversion library developed in Rust that supports offline conversion of 14 formats, including Word, EPUB, and PDF, into clean Markdown. | Embed anchor text that explains the language architecture, supported formats, and offline characteristics |
| Model item | This update not only improves reasoning ability but also lays the foundation for solving complex math problems, and experts believe it shows excellent potential. | Kimi K3 has officially opened the 2.8T parameter weights, adopting an asymmetric MoE architecture that cuts deployment costs to one quarter of the original while maintaining flagship accuracy. | Eliminate negative parallelism and vague sources, and give the parameter count, architecture, and specific cost changes |
| Opening introduction | Hello everyone! In this vibrant September, the AI field is thriving once again. Let us delve together into this week's exciting developments. | This week saw a dense burst of open-weight models: GLM-5.3 topped the open-source coding leaderboard, and Qwen3.8-Flash was also open-sourced; on the hardware side, OpenAI announced the first batch of measured data for its in-house chip. | Remove all pleasantries and boilerplate, and provide a one-sentence panoramic overview of this week's core highlights |
7. Delivery self-check checklist
Before the weekly issue is delivered as final or judged PASS, go through it item by item:
- Structure: Does the Front matter include title, date, summary, and tags, with complete formatting?
- Opening: Does the opening get straight to the point within 1–2 sentences, with no empty pleasantries?
- Links: Are the original article links for all items naturally embedded in the main keywords? Are there no standalone “[original article link]” lines?
- Summary: Is each item summary kept to 2–4 sentences, with sufficient information density?
- Action: Have nominalizations such as “carry out …… optimization/analysis/release” been eliminated?
- De-AI flavor: Have AI warning words such as “in addition,” “crucial,” “not only … but also …,” and “marks” been excluded?
- Formatting: Are half-width spaces retained between Chinese and English? Are proper nouns capitalized accurately? Are all punctuation marks full-width? Are there no extra Emoji?
- Images: Are image Alts concise and specific? Are all URLs valid web addresses with no temporary paths?
Reference sources
- Microsoft Writing Style Guide (https://learn.microsoft.com/en-us/style-guide/)
- op7418/humanizer-zh (https://github.com/op7418/humanizer-zh)
- Thomas Ptacek: How To Write With An LLM (https://sockpuppet.org/blog/2026/09/17/how-to-write-with-an-llm/)
- Joseph M. Williams & Joseph Bizup: Style: Lessons in Clarity and Grace