A rigorous 10-stage Chinese-Vietnamese translation workflow that combines terminology management, multi-model reflection, blind back-translation, and human review.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install back-translate-reflect
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install back-translate-reflect using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is 回译反思法 · 越南语翻译实战版?
回译反思法 is a specialized translation quality engine for Chinese-Vietnamese localization, based on the 10-stage workflow developed by 悦迅翻译 and 回译派. It combines terminology extraction, web verification, glossary standardization, batch translation, reflective editing, final revision, blind back-translation QA, defect repair, delivery gates, and compliance review.
This Openclaw Skills workflow is designed to make AI more valuable to translators rather than replace professional judgment. Its Vietnamese-focused reinforcement layers target pronoun stance, classifiers, post-nominal modifiers, passive voice, aspect markers, sentence-final particles, cultural terms, Sino-Vietnamese vocabulary, tones, and regional usage.
回译反思法 · 越南语翻译实战版 Use Cases
- Chinese-to-Vietnamese translation for business, technical, legal, marketing, and educational content.
- Vietnamese-to-Chinese translation requiring terminology consistency and natural target-language phrasing.
- Quality assurance for translations where conventional back-translation may hide Vietnamese word-order or pragmatic errors.
- Projects involving Vietnamese forms of address, regional vocabulary, classifiers, aspect markers, or culturally loaded terms.
- Multi-agent translation pipelines that need explicit separation between drafting, critique, repair, and approval.
- High-risk documents involving contracts, amounts, names, honorifics, legal language, or delivery-critical formatting.
- Translator workflows using Openclaw Skills to coordinate specialist models, glossaries, external research, and human review.
How 回译反思法 · 越南语翻译实战版 Works
- Extract terminology: Identify names, locations, organizations, brands, technical terms, abbreviations, and cultural vocabulary in structured JSON.
- Verify terminology: Search authoritative Vietnamese and Chinese sources, including Vietnamese media, official organization websites, reference works, and cross-checked sources.
- Standardize the glossary: Consolidate approved equivalents into a reusable glossary such as
{glossary:[{src,vi,zh,note}]}. - Create the draft translation: Use a professional translation model to produce the initial Vietnamese or Chinese translation while following the glossary and language-specific rules.
- Run reflective editing: A different reasoning-focused model reviews and improves the draft against the Chinese-Vietnamese checklist.
- Revise the final text: Integrate the draft and edited version into a deliverable, optionally producing literal, natural, and stylistic variants.
- Perform blind back-translation QA: After final revision, independently translate the target text back into the source language without polishing it. The back-translation model must differ from the drafting and reflection models.
- Diagnose and repair defects: Compare the original, final translation, and blind back-translation to identify semantic, pragmatic, terminology, and register defects, then apply targeted repairs with a separate model.
- Apply the delivery gate: Conduct human-led review of critical issues, including names, amounts, contracts, forms of address, and unresolved confidence concerns.
- Complete compliance review: Check legal language, formatting, regional Vietnamese usage, industry conventions, and all items marked for user confirmation before delivery.
回译反思法 · 越南语翻译实战版 Setup
- Copy the skill instructions into the prompt, agent skill directory, or orchestration layer used by your Openclaw Skills environment.
- Provide the source text, translation direction, target audience, domain, preferred Vietnamese region, formatting requirements, and any existing terminology.
- Configure separate model assignments for drafting, reflective editing, blind back-translation, and defect repair. Do not allow the same model to perform both initial translation and blind back-translation.
- Enable web access for terminology verification when authoritative equivalents or current proper names are required.
- Add the Vietnamese-specific checklist to the review prompt, covering pronouns, classifiers, modifier placement, passive constructions, aspect, sentence particles, cultural terms, Sino-Vietnamese vocabulary, tones, and dialect.
- Treat blind back-translation as a post-finalization gate only. It must not be run against the raw draft or used as a substitute for final revision.
- If only one model is available, label the result as not cross-model validated and require stronger human review.
- No package installation command or mandatory runtime dependency is specified in the source SKILL.md; deployment consists of configuring the workflow, prompts, model routing, research access, and human approval process.
Minimum operational inputs
- Source text and target language.
- Domain, audience, tone, and regional preference.
- Existing glossary or terminology constraints.
- Model aliases for translation, reflection, back-translation, and repair.
- Human reviewer and compliance requirements.
回译反思法 · 越南语翻译实战版 Data Schema & Taxonomy
The skill defines logical translation artifacts rather than a mandatory filesystem layout. Implementations can store each artifact as JSON, Markdown, or records in an orchestration system.
| Artifact | Core fields | Purpose |
|---|---|---|
| Source document | src, language, domain, audience, formatting rules |
Canonical input for every downstream stage |
| Term extraction | terms[] with src, type, hint |
Captures names, entities, abbreviations, technical terms, and cultural terms |
| Verification record | Source term, approved equivalent, source link, confidence, notes | Documents external evidence and translation decisions |
| Glossary | glossary[] with src, vi, zh, note |
Enforces consistent terminology across the translation |
| Draft translation | Source segments, target text, model identifier, stage | Records the initial batch translation produced by Model A |
| Reflection review | Issues, rationale, revised text, checklist dimensions, model identifier | Captures Model B analysis and Vietnamese-language improvements |
| Final translation | Final target text, variants, unresolved items, revision metadata | Represents the text eligible for blind QA |
| Blind back-translation | Back-translated source text, segment comparisons, Model C identifier | Detects semantic loss after final revision |
| Repair report | Defect type, severity, evidence, corrected segment, repair model | Tracks targeted fixes without letting the QA model silently rewrite content |
| Delivery and compliance gate | Human decisions, legal checks, regional checks, formatting status, confirmation items | Records approval, escalation, and release readiness |
Metadata taxonomy
- Language direction: Chinese-to-Vietnamese or Vietnamese-to-Chinese.
- Term type: Person, place, organization, brand, technical term, abbreviation, or cultural term.
- QA dimension: Pronoun stance, classifier, modifier order, passive voice, aspect, modal or sentence particle, cultural load, Sino-Vietnamese vocabulary, tone, or dialect.
- Severity: Informational, minor, major, critical, or human confirmation required.
- Model role: Translator, reflection editor, blind QA, repair, or compliance assistant.
- Confidence: Verified, probable, provisional, disputed, or unresolved.
- Regional preference: Northern Vietnamese, Southern Vietnamese, Central usage, or neutral formal Vietnamese.
The source does not prescribe generated filenames or directories, so an implementation should preserve stage numbers, model identifiers, source references, glossary versions, and review timestamps in its own storage layer.
回译反思法 · 越南语翻译实战版 Advanced Features
- Three-agent architecture: Separates the translation officer, quality officer, and final controller so production, criticism, and approval are not performed by one undifferentiated agent.
- Four-agent deployment mapping: Maps planning and translation to the translation officer, blind QA and diagnosis to the quality officer, and human-plus-compliance checks to the final controller.
- Mandatory model diversity: Enforces different models for initial translation, reflective editing, and blind back-translation to reduce self-confirming errors.
- NVIDIA relay auto-routing: Supports routing stages 4, 5, and 7 to distinct model aliases; explicit model selection is recommended because automatic routing may accidentally reuse a model.
- DUAL-REFLECT-style pre-translation signal: Optionally runs an intermediate draft back-translation and process assessment before reflective editing, while retaining the final blind back-translation gate.
- Vietnamese specialist checklist: Directly targets errors that semantic back-translation can conceal, including honorifics, pronouns, classifiers, post-nominal modifiers, particles, cultural terms, and dialect.
- Mother-tongue self-check: Prompts a reviewer or model to judge whether the target text sounds natural when the original source is assumed to be unavailable.
- Reference-free quality estimation: Supports target-language scoring with a CometKiwi-like model without routing the text through another back-translation cycle.
- Controlled literal back-translation: Uses low-temperature, word-order-aware instructions to reduce polishing and preserve detectable defects.
- Human-in-the-loop delivery gate: Keeps final responsibility with a human reviewer for contracts, financial amounts, names, forms of address, legal clauses, and ambiguous findings.
- Transparent fallback handling: Requires a visible warning when cross-model validation is unavailable instead of silently presenting single-model output as fully verified.
- Compliance and regional checks: Reviews Vietnamese legal terminology, formatting, and Northern-versus-Southern usage, with unresolved items explicitly surfaced to the user.
SKILL.md
Loading
Related Openclaw Skills
Contract Management
A template-driven contract management suite that routes legal requests to the right workflow, validates every output, and delivers traceable contract insights and documents.
Trade Alert — Crypto Trading Notifications
Trade Alert monitors real-time cryptocurrency markets and sends actionable price, risk, stop-loss, and take-profit notifications through Telegram.
Novel Writing SOP
A file-based AI writing workflow for planning, drafting, reviewing, and scaling original Chinese web novels.
LinkFox AI Text Generation
An asynchronous multimodal AI skill for generating, translating, and analyzing text from prompts, images, and videos.
Completion Report Generator
An automated Openclaw Skills report generator that aggregates task outcomes into daily tenant reports, fairness metrics, and 30-day welcome-back summaries.
Seedance 2.5 First-and-Last-Frame Video
Generate controlled Seedance 2.5 videos by defining the opening and closing frames, then automate media upload, task tracking, and result downloads.
Featured*
