Skip to content

English Terminology Draft

This file is the working glossary for the English edition of ai-programming-book.

Its purpose is not to force artificial one-to-one translation. Its purpose is to keep the book's conceptual vocabulary stable across chapters while still allowing natural English writing.

How to use this glossary

  • Prefer the recommended rendering unless the local chapter context clearly requires a narrower or broader wording.
  • Do not mechanically translate Chinese surface forms if the English technical community already has a more natural term.
  • Keep the same core concept stable across chapters, especially for high-frequency architectural ideas.
  • If a chapter introduces a new major concept, add it here after the translation is stable enough to reuse.

Style principles for terminology

  • Favor natural technical English over literal translation.
  • Favor conceptual consistency over local rhetorical variation.
  • Favor reader comprehension over preserving Chinese phrasing.
  • Avoid turning principle-level concepts into product-marketing language.

1. Book-level framing

Chinese Recommended English Avoid / Use with Caution Notes
AI 编程 AI coding AI programming Use AI coding as the book's default umbrella term because it matches the manuscript's framing and current practitioner language more closely. AI programming can appear in a subtitle or where broader formality is needed.
第一性原理 first principles fundamental principles Prefer first principles for the title and the book's core framing.
原理书 principles book theory book principles book keeps the engineering tone better than theory book.
判断力 engineering judgment judgment ability Often better rendered as engineering judgment, architectural judgment, or decision-making judgment depending on the chapter.
问题链 chain of questions problem chain Prefer chain of questions when describing chapter-to-chapter progression. In some contexts question chain is acceptable, but problem chain sounds unnatural in English.
运行真相 operational reality runtime truth runtime truth feels too literal. Use operational reality or rewrite the chapter title more naturally.
工程权衡 engineering trade-offs engineering balance Standard engineering wording is trade-offs.
架构选型 architecture selection architectural choice / architecture decisions Usually rewrite at sentence level as architectural choice, architecture decision, or design choice.
边界 limits / boundaries border Use boundaries for design scope and limits for hard capability ceilings.

2. Model and interaction fundamentals

Chinese Recommended English Avoid / Use with Caution Notes
大模型 large model / large language model / LLM big model Use large language model or LLM when the context is specifically language-model behavior. Use large model only if the text intentionally stays broader.
Token 化 tokenization token splitting tokenization is the standard term.
注意力机制 attention mechanism attention system Standard ML wording.
自回归生成 autoregressive generation self-regressive generation Standard wording.
温度 temperature randomness level Keep temperature; explain if needed for readability.
采样 sampling token picking Standard wording.
指令遵循 instruction following command following instruction following is the accepted model term.
上下文窗口 context window context frame Standard term.
多轮对话 multi-turn conversation multiple-round dialogue multi-turn conversation is more natural.
System Prompt system prompt system instruction set Keep the established English term lowercase in running prose unless style requires otherwise.
Few-shot few-shot prompting few-shot examples Use few-shot prompting when naming the technique.
Chain-of-Thought chain-of-thought reasoning step-by-step prompting Use the canonical term when discussing the technique specifically.
Lost in the Middle Lost in the Middle middle-attention decay Keep the recognized English phrase.
概率预测 probabilistic prediction probability guessing probabilistic prediction preserves the technical tone.

3. Agent systems and execution

Chinese Recommended English Avoid / Use with Caution Notes
Agent Agent intelligent agent (overused) Keep Agent capitalized only when treated as a named concept; otherwise lowercase in prose is fine.
Agentic agentic autonomous-style Preserve agentic where the manuscript is discussing the concept as such.
从问答到执行 from question answering to execution from Q&A to doing things Rewrite naturally in chapter titles and section headings.
执行循环 execution loop action cycle execution loop is the better default.
工具调用 tool calling function invocation Use tool calling as the umbrella concept. Reserve function calling for the specific mechanism.
Function Calling function calling tool calling Keep separate from tool calling: function calling is a mechanism, tool calling is the broader action.
ReAct 范式 the ReAct pattern ReAct paradigm pattern usually reads better than paradigm in this book's tone.
推理 reasoning thinking Use reasoning in technical contexts; avoid anthropomorphic thinking unless the prose deliberately uses quotes or rhetorical contrast.
行动 action / acting behavior In Reasoning + Acting, keep acting. In broader prose, action often reads better.
任务规划 task planning work planning Standard term.
任务分解 task decomposition task breakdown Both are acceptable; default to task decomposition when discussing system design.
自我反思 self-reflection self-thinking self-reflection is more standard in agent literature.
纠错 error correction / self-correction fix mistakes Choose based on context. self-correction often works well.
幻觉 hallucination fabrication Use hallucination for model behavior.
无限循环 infinite loop endless retry infinite loop is the correct engineering phrase.
过度自信 overconfidence false confidence overconfidence is more concise.
上下文污染 context pollution context contamination Prefer context pollution unless a more formal tone requires contamination.
规划失败 planning failure planning mistake Better for system-level discussion.
容错机制 fault-tolerance mechanisms error-proof design Use fault tolerance or fault-tolerance mechanisms.
回滚策略 rollback strategy revert strategy Standard engineering term is rollback strategy.
能力边界 capability boundaries / capability limits capability border Use capability boundaries when discussing scope; use capability limits when emphasizing hard ceilings.
失败模式 failure modes failure patterns failure modes is the standard engineering phrase.

4. Protocols, tools, and capability packaging

Chinese Recommended English Avoid / Use with Caution Notes
工具的标准化 tool standardization tool normalization Use tool standardization.
工具发现 tool discovery tool lookup Standard systems term.
能力声明 capability declaration feature declaration capability declaration is closer to MCP-style wording.
提供方 / 使用方 provider / consumer supplier / user side Use the standard systems pair provider and consumer where applicable.
传输层 transport layer transmission layer Use transport layer in MCP context.
Model Context Protocol Model Context Protocol (MCP) context model protocol Keep the canonical full name on first mention, then MCP.
预定义的能力包 packaged capability predefined capability pack For chapter titles and conceptual framing, packaged capability reads more naturally.
Skill Skill capability package only Keep Skill as the core term; use packaged capability to explain it, not replace it.
Prompt Template prompt template prompt pattern Keep the standard product term.
动态加载 dynamic loading live injection dynamic loading is the standard phrase.
工具编排 tool orchestration tool scheduling Prefer tool orchestration.
工作流定义 workflow definition process script workflow definition is stable and neutral.
生命周期 lifecycle life cycle Both are acceptable; use lifecycle consistently.
主 Agent orchestrator agent / main agent master agent Avoid master agent; prefer orchestrator agent.
SubAgent sub-agent / SubAgent child agent sub-agent is more natural in prose; keep SubAgent when mirroring a product or chapter term.
上下文隔离 context isolation isolated context Use context isolation for the design principle.
结构化报告 structured report formatted summary structured report fits system communication better.
协议化 protocolization turning into a protocol protocolization is acceptable in technical prose, but often a sentence-level rewrite reads better.
Agent Card Agent Card agent profile Keep the canonical protocol term if discussing A2A.
Agent 间互联网 the internet between agents / agent internet Agent Internet This is rhetorical language. Usually rewrite by meaning rather than force a fixed term.

5. Memory, context, and token economics

Chinese Recommended English Avoid / Use with Caution Notes
记忆体系 memory system memory architecture Use memory system as the default book-level term. memory architecture is acceptable when discussing design structure.
工作记忆 working memory active memory Keep the standard term.
短期记忆 short-term memory session memory short-term memory is the conceptual layer; session memory can be used if the implementation is session-scoped.
长期记忆 long-term memory persistent memory long-term memory is the conceptual term; persistent memory can describe implementation properties.
记忆写入 memory write / memory writing memory save In prose, writing to memory often reads best.
记忆检索 memory retrieval recall lookup Standard phrase.
记忆一致性 memory consistency memory coherence Use memory consistency unless the discussion specifically parallels distributed systems coherence.
跨会话记忆 cross-session memory multi-session memory More natural phrasing.
上下文工程 context engineering prompt engineering Keep separate from prompt engineering. The book treats them as different levels of abstraction.
Token 经济学 token economics token cost science token economics is natural and already used in the manuscript.
上下文压缩 context compression compressed context Prefer context compression.
摘要压缩 summarization-based compression summary compression Use summarization-based compression in explanatory passages; summary compression can work in tighter prose.
结构化压缩 structured compression schema compression Use structured compression as the broader term.
选择性注入 selective injection selective stuffing Use selective injection.
Prompt Caching prompt caching context caching Keep the product/industry term prompt caching.
输入 Token input tokens prompt tokens Use input tokens unless the pricing model specifically says prompt tokens.
输出 Token output tokens completion tokens output tokens is the safest neutral term.
缓存 Token cached tokens cache tokens Use cached tokens.

6. Knowledge injection and model adaptation

Chinese Recommended English Avoid / Use with Caution Notes
知识注入 knowledge injection knowledge feeding knowledge injection matches the current book framing well.
检索增强生成 retrieval-augmented generation (RAG) retrieval generation Always use the canonical full term on first mention, then RAG.
向量嵌入 vector embeddings vectorized representations Use vector embeddings when referring to the concept broadly.
向量空间 vector space embedding space only Use vector space when explaining the geometry, embedding space when the context is model embeddings specifically.
分块策略 chunking strategy slicing strategy chunking strategy is the accepted RAG term.
微调 fine-tuning model tuning Standard term.
长上下文 long context long-context window Use long context generally; use long-context window only if the sentence specifically focuses on the window.
私有数据 private data internal data only private data is the better umbrella term. internal data can be used for company-internal corpora.
检索不到 retrieval miss search failure In failure-mode lists, retrieval miss works well.
检索质量 retrieval quality search quality Prefer retrieval quality in RAG contexts.
混合检索 hybrid retrieval mixed retrieval Standard term.
语义检索 semantic retrieval semantic search Both work, but semantic retrieval aligns better with the chapter's architecture framing.
关键词检索 keyword retrieval keyword search Same principle: keyword retrieval is more parallel with other glossary items.

7. Specification-driven engineering

Chinese Recommended English Avoid / Use with Caution Notes
规范驱动编程 specification-driven development spec-driven programming For chapter-level framing, prefer specification-driven development. In running prose, spec-driven development is acceptable after the concept is established.
提示词 prompt prompt words Always prompt.
结构化提示词 structured prompt structured prompting Use structured prompting for the method, structured prompt for an artifact.
规范 specification / spec rule specification is safer for formal discussion. spec works well in running prose after the term is introduced.
声明式 declarative statement-based Use the standard engineering term.
分层的 layered multi-layered Use layered unless the distinction matters.
可组合的 composable combinable composable is the better engineering term.
可验证的 verifiable checkable Use verifiable.
持久化规范 persistent specification saved specification persistent specification preserves the architectural meaning.
全局规范 global spec global rule set Use global spec in prose after first mention.
项目规范 project spec project rules project spec is the cleaner form.
任务规范 task spec task rules Standard wording.
OpenSpec OpenSpec open specification format Keep the project/product name unchanged.
.cursorrules .cursorrules Cursor rules file Keep the filename literal when referenced.
约束空间 constraint space space of constraints constraint space reads naturally in technical prose.

8. Security, trust, and evaluation

Chinese Recommended English Avoid / Use with Caution Notes
安全与对齐 security and alignment safety and alignment The chapter can use security and alignment because the manuscript foregrounds trust boundaries and system design, not only model safety.
信任边界 trust boundary confidence boundary Standard security term.
威胁模型 threat model risk model Keep threat model when the security frame is explicit.
Prompt Injection prompt injection prompt attack Use the standard attack name.
数据泄露 data leakage data leak data leakage is more formal and better for principle-level writing.
输出安全 output safety output security Use output safety when discussing harmful model outputs; use output security only if the context is stricter access/control semantics.
权限最小化原则 principle of least privilege minimal permission principle Standard security phrase.
人工确认节点 human approval checkpoint manual confirmation point human approval checkpoint is more natural in workflow writing.
评估 evaluation assessment Use evaluation as the book's default term because it aligns with LLM engineering practice.
评估集 evaluation set benchmark set evaluation set is the safer default. benchmark set is narrower and more formalized.
离线评估 offline evaluation offline testing Keep offline evaluation.
在线评估 online evaluation live testing Use online evaluation.
回归门禁 regression gate regression barrier regression gate reads naturally in engineering contexts.
灰度 gradual rollout / canary rollout grayscale rollout Never translate literally. Choose by deployment context.

9. Non-deterministic systems and organization

Chinese Recommended English Avoid / Use with Caution Notes
非确定性 non-determinism / non-deterministic uncertainty Use the formal term when discussing system behavior.
工程化挑战 engineering challenges implementation difficulties engineering challenges fits the book's tone.
可观测性 observability monitorability Standard term.
延迟预算 latency budget time budget Use latency budget.
成本控制 cost control budget management Prefer cost control in systems discussions.
模型更新 model updates model upgrades Use model updates as the default neutral term.
治理 governance management governance is more precise for the book's organization-level discussion.
组织迁移 organizational migration team transition Usually rewrite based on local context. organizational migration is the closest default.
新型资产 new operational assets / new AI-native assets new resources Often rewrite by meaning. AI-native assets works well when referring to Prompt, Spec, Skill, memory, and evaluation sets as a governed class of assets.
静默退化 silent degradation hidden decay silent degradation fits engineering prose well.
变更管理 change management change handling Standard organizational term.
审计留痕 audit trail audit logging only Use audit trail unless the context specifically means logs.
平台角色 platform role infrastructure role platform role is closer to the manuscript's organizational framing.
团队级系统 team-level system organizational system team-level system is the safer direct rendering.
AI 原生架构 AI-native architecture architecture for AI AI-native architecture is the natural English form.
Agent 网络 agent network multi-agent network Use agent network for the broader future-facing concept; multi-agent system is fine when the text is discussing current architecture.

10. Terms that should usually stay in English

These terms should usually remain in English even inside otherwise translated prose:

  • Agent
  • SubAgent / sub-agent
  • MCP
  • Model Context Protocol
  • Skill
  • prompt
  • system prompt
  • function calling
  • tool calling
  • ReAct
  • RAG
  • OpenSpec
  • .cursorrules
  • Prompt Injection
  • Lost in the Middle
  • prompt caching

11. Translation notes for chapter titles

These are not final titles. They are directionally preferred renderings to keep terminology aligned.

Chinese Chapter Concept Directional English Rendering
大模型是怎么“思考”的 How Large Language Models Actually Work
与大模型交互的本质 What Interaction with an LLM Really Is
从问答到执行：Agentic 的运行原理 From Answers to Action: How Agentic Systems Work
工具的标准化：MCP 交互原理 Standardizing Tool Use: How MCP Works
预定义的能力包：Skill 交互原理 Skill as a Packaged Capability
一个不够用：多 Agent 协作原理 When One Agent Is Not Enough
Agent 的能力边界与失败模式 Agent Limits and Failure Modes
让 AI 记住你：记忆体系的原理与设计 Building Memory for AI Systems
Token 经济学：上下文工程的艺术 Token Economics and the Art of Context Engineering
知识注入的三条路：RAG、微调与长上下文 Three Paths to Knowledge Injection
规范驱动编程：从提示词到 OpenSpec From Prompts to OpenSpec
安全与对齐：AI 系统的信任边界 Security, Alignment, and Trust Boundaries
非确定性系统的工程化挑战 Engineering for Non-Deterministic Systems
AI 工程组织学：治理、评估与团队迁移 The Organizational Side of AI Engineering
AI 编程的边界与未来 The Limits and Future of AI Coding

12. Open questions to revisit later

These terms are good enough for the first translation pass, but should be revisited after 2-3 chapters are translated:

  • Should the book title stabilize on AI Coding from First Principles, The First Principles of AI Coding, or another variant?
  • Should AI 编程 remain consistently AI coding, or should some formal chapters switch to AI software engineering in limited contexts?
  • Should Skill always stay capitalized in chapter prose, or only when referring to the product-like concept explicitly?
  • Should 上下文工程 remain context engineering everywhere, or should some explanatory passages expand it as the engineering of context for readability?
  • Should 知识注入 remain knowledge injection, or be softened to bringing external knowledge into the model in more reader-facing passages?
  • Should 组织迁移 be rendered more idiomatically chapter by chapter rather than fixed globally?

13. Practical rule of thumb

When in doubt:

  1. Preserve the concept, not the Chinese surface form.
  2. Prefer the wording that an experienced English-speaking engineer would naturally use.
  3. If a term sounds translated rather than written, rewrite it.
  4. If a concept is already established in the book, do not rename it casually in later chapters.