When analyzing technical articles related to agents, a common issue emerges: traditional translation tools and methods often fail to deliver accurate and fluent Chinese translations. Literal, word-for-word approaches frequently produce "translationese"—awkward, rigid outputs with unnatural sentence structures and poorly handled technical terms, rendering the content challenging for readers.
To tackle this problem, an exploration of various AI tools led to the creation of an AI-driven, multi-round review and refinement workflow on Dify. Techniques such as segmented translation followed by merging, reflective refinement, and a preference for "rewriting" over direct "translation" were integrated into the process. This workflow enables high-quality conversion of English technical articles—spanning AI, programming, product development, business, and agent-related topics—into Chinese.
The resulting process ensures that rewritten articles retain their original meaning while achieving fluency, naturalness, and alignment with Chinese linguistic conventions and professional technical standards. Designed to serve both technical professionals and general readers, this approach broadens accessibility to complex content.
This article outlines the practical development of this workflow, covering its design, Prompt Engineering strategies, model selection, and additional steps like typesetting and publishing. The objective is to showcase how AI can boost content creation efficiency and quality, offering a solution for readers who find English technical materials difficult to access. Ultimately, this workflow supports the production and dissemination of high-quality Chinese technical blogs, encouraging learning and knowledge-sharing across a wider community.
Why Multi-Round Review and Refinement?
"translation" or "rewriting" often falls short of high-quality content standards. Initial drafts, even when rewritten, may exhibit flaws such as:
Language Issues: Unnatural phrasing, imprecise terminology, and sentence structures misaligned with Chinese norms.
Content Issues: Omissions, distortions, unclear logic, or mishandled technical terms.
Style Issues: Inconsistent tone or misalignment with the preferences of target audiences, whether experts or novices.
To address these challenges, a multi-round review and refinement process was developed by Gino. Multiple LLMs assess and enhance the rewritten text from diverse angles, producing polished, high-quality Chinese output. Drawing inspiration from software development’s "code review" practices, this method leverages varied perspectives to identify and resolve issues, elevating overall quality.
Workflow Overview
The workflow comprises the following stages:
Content Crawling: The FireCrawl tool extracts article content from a provided URL, filtering out irrelevant elements like navigation bars and advertisements.
Initial Rewriting: An LLM processes the extracted English text into a preliminary Chinese rewrite, adhering to basic grammar and expression norms.
Multi-Round Review:
Parallel Review: Three distinct LLMs evaluate the draft simultaneously, focusing on language fluency and authenticity, content accuracy and logic, and style consistency and reader suitability, providing targeted revision suggestions.
Reflective Improvement: Multi-round, multi-dimensional reviews catch issues a single LLM might miss, comprehensively enhancing rewrite quality.
Comprehensive Refinement: Another LLM synthesizes the original text, initial draft, and feedback from the three reviewing LLMs to produce an improved rewritten version.
Final Polish: The enhanced article undergoes a final refinement and consistency check to ensure top-tier output quality.
Get the Workflow:
Prompt Engineering Breakdown
Prompt Engineering forms the backbone of this workflow. Below is an overview of the design and key elements for each stage:
1. Initial Rewriting
Role Assigned: The LLM acts as a senior language expert optimizing expression.
Objectives:
Fully grasp the original text’s meaning (emphasizing context, precise term translation, cultural adaptation, and information completion).
Generate natural, idiomatic Chinese (reordering phrases, splitting long sentences, refining word choice, matching tone, and standardizing punctuation).
Ensure complete, accurate information transfer (covering all points and highlighting key ideas).
Exclude irrelevant content (handled primarily by FireCrawl).
Polish and proofread for excellence.
Output: A fluent, accurate Chinese rewrite preserving essential Markdown formatting (images, links, code blocks, headings, lists, etc.).
2. Multi-Round Review
LLM 1: Language Fluency and Authenticity
Role Assigned: Senior Chinese language expert and technical editor.
Focus:
Fluency (sentence smoothness, transitions, grammar).
Authenticity (word choice, collocations, sentence structure).
Naturalness (style, tone, emotional resonance).
Constraints: Avoid structural changes; prioritize expression and sentence-level adjustments.
Suggestions: Deliver specific, actionable feedback (location, issue, fix).
LLM 2: Content Accuracy and Logic
Role Assigned: Senior academic editor, technical content reviewer, and fact-checker.
Focus:
Accuracy (information consistency, term precision, fact verification).
Logic (structure, paragraph coherence, concept alignment).
Completeness (background, argumentation, coverage).
Constraints: Preserve structure; emphasize expression, accuracy, logic, and completeness.
Suggestions: Provide clear, practical improvement ideas (location, issue, fix).
LLM 3: Style Consistency and Reader Fit
Role Assigned: Senior copywriter, technical communicator, and UX researcher.
Focus:
Consistency (style, tone, phrasing, rhythm).
Reader Fit (audience profile, language level, depth, expression, emotional appeal).
Impact (engagement, clarity, persuasiveness, memorability, call-to-action).
Constraints: Maintain structure; focus on expression and audience alignment.
Suggestions: Offer precise, actionable revisions (location, issue, fix).
3. Comprehensive Refinement
Role Assigned: Senior technical editor and language expert.
Objective: Integrate the original text, initial draft, and three LLM reviews into an improved Chinese rewrite.
Rules:
Evaluate and selectively apply review feedback.
Retain structure while enhancing expression.
Balance technical precision with accessibility (accurate terms, readable style).
Output: An enhanced rewrite with optional revision notes.
4. Final Polish
Role Assigned: Senior language expert, technical editor, and proofreader.
Objective: Refine and ensure consistency for a flawless final text aligned with reader expectations.
Focus:
Language polish (vocabulary, syntax, tone, rhythm).
Content proofreading (facts, logic, clarity).
Consistency (terms, style, punctuation, formatting, proper nouns).
Note: Limit to minor adjustments; flag major issues in notes without altering text.
Output: Polished article with optional revision notes.
Model Selection
The following LLMs were employed in this workflow:
Initial Rewriting, LLM 3 Review, Comprehensive Refinement, Final Polish: Google Gemini 2.0 Flash—chosen for its long-context handling (suitable for lengthy articles), speed, and cost-efficiency.
LLM 1 Review: Qwen-max-latest—selected for its strong Chinese comprehension and generation, especially with complex, lengthy texts.
LLM 2 Review: OpenAI o3-mini—picked for its superior performance in translation and Chinese language understanding.
Test Run
Testing reveals that this workflow markedly improves the quality of Chinese adaptations of English technical articles. The output is fluent, authentic, and natural, with accurate content, clear logic, consistent style, and wide reader appeal.
Post-Processing: Typesetting, Cover, and Publishing
To ensure professionalism and visual appeal, the following steps were executed:
Typesetting Adjustments: Using the Cursor editor, the Markdown article was refined—adjusting headings, spacing, and code block formatting.
Cover Creation: Cover image prompts were generated with Cursor, based on the article’s content and site style.
Flux Pro (or similar AI tools) produced a thematic, stylish cover image.
Publishing: The polished article and cover were uploaded to blog.
Lessons Learned and Future Outlook
This workflow underscores AI’s transformative potential in content creation. Built on Dify, it enhances the quality of Chinese adaptations of English technical articles, eliminates "translationese," and delivers reader-friendly content.
