Podcast Atlas
concept Updated 2026-07-09 Tags: Context, Knowledge-Management, Agents

Context Engineering

Context engineering is the practice of accumulating, organizing, and refining the information that helps AI produce useful work. 高手怎么用 AI？普通人怎么学 AI？投资人如何投 AI？｜对谈课代表立正 argues that as models and tools converge, differentiated performance will come from context: preferences, examples, standards, documents, workflows, and implicit judgment made explicit. OpenAI 和 Anthropic 共同看好的 FDE：AI 时代的新岗位出现，旧分工松动｜对谈 Rolling AI adds the enterprise version: Digital Employees need company processes, frontline knowledge, system access, and expert examples before they can work, so Forward Deployed Engineer practice becomes applied context engineering. 阿里千问离职余震，在几万人的铁球里如何体面生存 adds that context must include operational details such as data migration constraints and the user’s own internalized preparation for live situations. Community-Led SaaS Growth: How Ninety Hit $44M ARR adds an organizational-data example through Mas, which depends on Ninety’s accumulated context about vision, values, roles, accountability, rocks, metrics, and feedback. Agent 元年第 500 天：什么在消失，什么在诞生——为什么我们不该再投资 GUI 思维的软件？ treats context as the unchanged main thread of the first 500 agent-era days. 人类和 AI Agent 的最佳配合方式，还没被发明｜对谈 Paperboy adds the personal-computing version through OS-Level Context and Persistent Agent Memory, where the agent learns from the user’s actual computer activity instead of relying only on chat history or manually maintained prompt files. 探秘 Claude Code，搞懂 Agent Harness｜对谈来新璐 adds the harness version: context engineering includes working directory state, dependency and git state, system prompts, skills, memory, context compression, and handoff documents for the next agent.

EP108 Vibe Coding大地震：Cursor定价争议、Windsurf收购风波，模型厂商亲儿子们又将如何进场？ adds the Vibe Coding version. The source argues that good module boundaries and interface design help keep agent context tractable, and it contrasts Gemini CLI’s large-context behavior with Cursor’s likely chunking, indexing, and retrieval approach.

Episode 17: 向量模型工程师：AI 的隐藏瓶颈与新时代的信息迷宫 adds the retrieval and RAG version. N 同学 / N Student argues that useful context is not produced by dumping documents into a model: source quality, Document Chunking, Vector Model Engineering, Semantic Search Relevance, Reranking Models, and evaluation decide whether a system can find the right evidence before generation begins.

为什么硅谷开始重新定义「AI 记忆」｜ S10E20 adds the local personal-memory version. 康宏文 Henry argues that context engineering for personal assistants starts before retrieval: local videos, recordings, images, and files must be understood and converted into reusable memory through Data-to-Memory Transformation, otherwise even large context windows or RAG only operate over poorly shaped material.

AI 会写代码了，为什么你还是做不出产品？ adds a logging and observability version. The hosts argue that AI debugging improves when the user has asked the system to emit detailed logs, preserve test results, and document old-code behavior; without those traces, AI has too little context to diagnose logic errors that still compile and run.

OpenClaw 之后，谁将定义主动式 AI 的新战场？｜对谈 AirJelly 黄柏特 adds the AirJelly version: context engineering is not only stuffing more material into a model, but deciding when context has enough signal to save. Intent Context, event/entity extraction, memory merging, time decay, retrieval, and local privacy boundaries become product decisions for turning everyday computer activity into agent-usable context.

“AGI 来了？我用了一周，头皮发麻“｜对谈张昊然：Moxt 联合创始人 adds the Moxt version: context engineering becomes workspace design. Zhang Haoran argues that documents, meeting recordings, data definitions, project status, code changes, and comments need to live as Organizational Context in AI-readable formats before AI Coworkers can work with less repeated briefing.

130. 张月光创业两年首次访谈：妙鸭不是AI Native产品、流程到上下文设计、One Way Door和乙女游戏 adds the product-design version through 张月光. He argues that AI Native Product Design moves from designing deterministic user flows toward defining what context, examples, constraints, and quality boundaries the model needs before the user interface can be settled.

EP127 从 Skills 到自动化工作流，论 Agent 如何接管真实生产力 ⚙️ adds the skill-curation version. Requirement-grilling, architecture maps, project-local skills, podcast transcripts, 微信读书 notes, and repeated operations tasks all become context assets when they are written down clearly enough for agents to reuse.

Vol. 164 从苹果聊到软件未来：Agentic Software 真的要来了？ adds a limit case: more context and project memory help, but stale or compressed context can rot, and agents should sometimes challenge the user’s request instead of faithfully executing a weak premise.

E163.要完了？不！是要玩了！论养AI的心态与习惯 adds the personal external-brain version. 品哥 frames the context window as a scarce working-memory resource, so durable material should live in markdown files, prior writing, transcripts, notes, and indexed folders that can be pulled in only when relevant. The episode also stresses that context is not only documents; it includes why the task exists, who the user is, what output they can accept, and which parts of their taste or “soul” should guide the agent.

读书，就是在读一个人的 F adds a reading and knowledge-ecology version. Bookshelves, whiteboards, linked notes, conversation history, and the user’s own AI all become context for seeing and reshaping ideas. Personal Knowledge Ecology makes the context bidirectional: it helps AI understand the user, but it also helps the user notice their own X/F/FX Framework and decide what should be read or delegated.

140. 对姚顺宇的4小时访谈：请允许我小疯一下！在Anthropic和Gemini训模型、技术预测、英雄主义已过去 adds the training-side version through Long-Horizon AI. Yao Shunyu / 姚顺宇 argues that useful agents should be trained in finite context but used as if context were effectively much longer through interaction, selective forgetting, retrieval, and compression. This moves context engineering from user workflow into model behavior and post-training design.

135. 和自然选择创始人Tristan聊，Elys、赛博分身、灵魂、Context的获取与流动和AI社交网络 adds the AI-social-network version through Elys. Tristan argues that context becomes valuable when it can flow between Cyber Avatars inside AI Social Networks, letting agents do pre-interaction and filtering before handing valuable connections to real people. This shifts context engineering from private prompt quality toward networked Subjectivity As AI Asset and Context Flywheel design.

一个 AI 创始人的虚荣心、装，和愚昧之巅｜对谈 invoko.ai 创始人梦琪 adds the internal-company workflow version through invoko.ai / Invoqo. 梦琪 / Mengqi says an internal growth skill worked better than an external 2B product partly because the team had full context: target users, outreach details, pricing, negotiation state, content checks, communication history, and acceptance criteria lived inside the team’s own process.

Role In The Sources

  • Kedaibiao Lizheng treats context as a personal, team, and company advantage.
  • AI Skills are one way to package context into reusable procedures.
  • Agentic Workflow tools can use context more effectively than isolated chat sessions.
  • AI Assisted Software Development Risk shows that missing context around production state can cause real user harm.
  • Human Judgment Under AI shows that useful context must eventually become human understanding, not only model input.
  • Mas shows context as a SaaS product asset when organizational operating data becomes queryable by AI.
  • Rolling AI shows context as a deployment asset when AI must learn from strong frontline workers and existing systems.
  • Paperboy shows context as a personal-agent asset when OS activity, meetings, messages, code, and current app state become durable memory.
  • Claude Code shows context as a harness asset when the system decides which tool outputs to discard, what to summarize, and what the next agent must receive.
  • Gemini CLI and Cursor show context as a coding-product choice between direct long-context loading and engineered retrieval/indexing.
  • The Fuyou Tiandi vector-model episode shows context as retrieval engineering: chunks, vectors, relevance labels, hard negatives, and reranking determine which source text reaches the model.
  • S10E20 shows context as memory engineering: multimodal personal archives need structure, timestamps, entity labels, and source-grounded recall before they become useful agent context.
  • Shengpai Notice shows context as product/workflow knowledge made explicit enough for AI to implement a usable internal tool.
  • Detailed logs, tests, screenshots, and documentation become context assets for future AI debugging and refactoring.
  • AirJelly shows that the trigger for collecting context can be as important as the context itself; Enter-triggered capture tries to reduce random browsing noise.
  • Moxt shows context as workspace infrastructure: Markdown, CSV, JSON, HTML, file structure, meetings, data, and project traces are shaped for agent work.
  • Zhang Yueguang shows context as product-design infrastructure: the team starts from model input needs and smallest useful generation units before locking the flow.
  • EP127 shows context as skill material: repeated prompts, acceptance criteria, design preferences, reading notes, transcripts, and operational procedures can be compressed into reusable agent instructions.
  • Vol. 164 shows context as both asset and risk: memory helps only if it preserves the right signal, stays fresh enough, and supports agent pushback.
  • E163 shows context as an external-brain habit: personal archives, style files, memory notes, and acceptance criteria become the material from which agents learn the user.
  • The “读书，就是在读一个人的 F” source shows context as a reading and note ecology: shelves, whiteboards, friends, AI cards, and linked notes become material for training the user’s frame.
  • Episode 140 shows context as a long-horizon training target: the model must learn which state to keep, which to discard, and when to recover information during extended work.
  • Episode 135 shows context as a social-network asset: the product must collect, confirm, and circulate enough subjectivity context for avatars to create better human connections.
  • The InvokoAI source shows context as a reason internal tools can beat external products: full workflow history and direct acceptance criteria make the agent easier to improve.

Connections