Claude Code（Codex）中文提示词与英文提示词差异（RFC征求意见稿、ADR架构决策记录文档）Claude提示词、Codex提示词

文章目录对 Codex / Claude Code / GPT 系 Coding Agent 来说一、训练数据分布问题二、中文不会“看不懂”中文工程术语容易“漂移”三、最容易出问题的是“抽象描述”四、真正推荐的方法中文思考 英文工程定义五、最佳实践非常推荐1. 中文思考2. 英文工程文档六、尤其是这些东西强烈建议英文Agent 名称状态流转Artifact 类型Prompt七、什么时候中文反而更好数据层必须支持中文八、你现在最推荐的方式用英文写用中文提示词示例对 Codex / Claude Code / GPT 系 Coding Agent 来说英文 System Design 文档通常效果更稳定。原因不是“中文不行”而是一、训练数据分布问题这些 coding agent 的高质量训练数据大部分来自GitHubRFC (Request for Comments 一种技术文档标准格式最初用于互联网标准制定)ADR(Architecture Decision Record 架构决策记录文档)OpenAPIArchitecture DocsTech SpecsPRDIssuesStackOverflow而这些绝大部分是英文。所以 agent 对这种英文结构Responsibilities: Outputs: Constraints: Success Criteria:理解非常强。它会自动联想到模块边界接口设计工程分层workflowdependency二、中文不会“看不懂”实际上现在 GPT-5 / Claude 4 已经完全能理解中文工程文档。所以中文完全能做项目。尤其产品描述PRDworkflowprompt用户需求都没问题。但问题在于中文工程术语容易“漂移”比如你写机会发现Agent 可能理解成recommendationlead generationtrend analysisopportunity mining但英文Opportunity Discovery含义会稳定很多。三、最容易出问题的是“抽象描述”例如中文后面做一些评估LLM 很容易自由发挥。但英文Evaluate technical feasibility and market viability边界会清晰很多。四、真正推荐的方法不是“全英文”。而是中文思考 英文工程定义这是现在很多 AI 工程团队真实做法。例如你自己用中文整理我希望系统从社媒发现用户痛点然后转成The system discovers user pain points from social discussions.再比如中文后续进行价值评估转Evaluate market value, technical feasibility, and competitive landscape.五、最佳实践非常推荐你以后所有 AI 工程项目建议采用1. 中文思考用于brainstorming产品构思workflow 推演agent 设计因为中文对你认知负担最低。2. 英文工程文档用于READMEArchitecturePRDCodex instructionsPrompt templatesAgent contractsAPI schema因为LLM 对英文工程语义最稳定。六、尤其是这些东西强烈建议英文以下内容最好永远英文Agent 名称用Collector Agent Ranking Agent Evaluation Agent不要收集代理 排序代理因为英文 agent naming 更稳定。状态流转pending running completed failed retrying不要中文。Artifact 类型Opportunity Report PRD Architecture Doc Task BreakdownPromptPrompt 强烈建议英文。因为英文 prompt更短token 更少语义边界更稳定七、什么时候中文反而更好如果你的项目面向中文用户分析中文社媒尤其 Rednote提取中文语义中文情绪分析那数据层必须支持中文包括中文 embedding中文 chunking中文 normalization因为中文互联网语义和英文互联网差异极大。八、你现在最推荐的方式你现在应该用英文写项目定义agent contractarchitectureworkflowartifact schema用中文思考产品推演灵感记录这是目前 AI 工程效率最高的方式。提示词示例#Project VisionThis project is a multi-agent opportunity discovery and evaluation system.The system continuously collects discussions,complaints,feature requests,and trend signals from multiple social platforms,then uses a series of AI agents to:1.Discover emerging user needs2.Cluster and rank opportunities3.Evaluate market and technical feasibility4.Generate structured business and engineering artifacts Thelong-term goal is to build an AI-nativeopportunity intelligence pipelinethat transforms raw internet discussions into actionable product and engineering outputs.---#Core ConceptThis is NOT just a chatbot or autonomous agent demo.The system is designed as an Artifact Pipeline:Social Signals ↓ Need Discovery ↓ ClusteringRanking ↓ Evaluation ↓ Planning ↓ Artifact Generation The final output of the system is structured artifacts,not conversations.---#Information SourcesThe MVP should support:-Reddit Future versions may support:-Hacker News-GitHub Issues-Twitter/X-Product Hunt-Discord-Telegram-YouTube comments-App Store reviews-Rednote(Xiaohongshu)---#MVP ScopeThe MVP should remain intentionally small and focused.DO NOT over-engineer the system.Avoid:-distributed systems-microservices-autonomous self-improving agents-complex memory systems-multi-region infrastructure-advanced scheduling systems The MVP goal is to validate the workflow and artifact pipeline.---#MVP ArchitectureThe MVP should contain5core agents.##1.Collector Agent Responsibilities:-collect social posts-normalize data-clean content-deduplicate records Outputs:-raw signal dataset---##2.Signal Ranking Agent Responsibilities:-embedding generation-clustering-trend detection-pain point extraction-signal ranking Outputs:-ranked opportunity candidates---##3.Evaluation Agent Responsibilities:-market analysis-technical feasibility-competitor analysis-risk analysis-confidence scoring Outputs:-evaluated opportunities---##4.Planner Agent Responsibilities:-PRD generation-architecture planning-task breakdown generation-implementation suggestions Outputs:-structured engineering plans---##5.Document Agent Responsibilities:-markdown generation-JSON export-report formatting-artifact persistence Outputs:-final artifacts---#MVP ArtifactsThe MVP should generate4primary artifacts.##1.Opportunity Report Contains:-identified problem-supporting evidence-user pain analysis-market signals-confidence score---##2.PRD Contains:-product goals-user stories-functional requirements-non-functional requirements---##3.Architecture Document Contains:-system components-workflow design-infrastructure suggestions-storage design---##4.Task Breakdown Contains:-implementation tasks-priorities-estimated complexity-milestone suggestions---#Suggested TechStack(MVP)Backend:-Python-FastAPI AI Orchestration:-LangGraph LLM:-OpenAI/Anthropic APIs Vector Storage:-pgvector or Chroma Database:-PostgreSQL Async Tasks:-asyncio initially Frontend:-simple dashboard-minimal UI Observability:-basic logging only Deployment:-Docker Compose---#Recommended Repository Structuremonorepo preferred Example:/apps/api/worker/frontend/packages/agents/prompts/shared/docs/artifacts---#Production VisionFuture production versions may include:## Workflow Infrastructure-Temporal-durable execution-retries-checkpointing ## Advanced Agent Runtime-long-running workflows-human-in-the-loop-memory systems-agent versioning ## Observability-Langfuse-OpenTelemetry-tracing-token/cost tracking ## Scalability-Kafka-Redis streams-distributed workers ## Advanced Evaluation-evaluator agents-reinforcement loops-historical opportunity tracking ## Artifact Expansion Additional artifacts may include:-Jira tickets-GitHub issues-startup briefs-investor reports-architecture diagrams-sprint plans---#Engineering PrinciplesThe system should prioritize:1.Simplicity over complexity2.Deterministic workflows over autonomy3.Structured artifacts over chat responses4.Clear state transitions5.Human-readable outputs6.Observable workflows7.Modular agents---#Important Constraints-Keep the MVP implementation small-Avoid premature optimization-Prefer explicit workflows overmagic-Buildforclarity and maintainability-Every agent must have clear inputs and outputs-Every workflow stage should produce structured artifacts---#Success CriteriaforMVPThe MVP is considered successfulifit can:1.Collect real social signals2.Detect at least one meaningful opportunity3.Generate useful structured artifacts4.Demonstrate a complete end-to-end workflow5.Produce outputs that humans can review and act on The goal is NOT full autonomy.The goal is a reliable AI-assisted opportunity pipeline.