Podcast Atlas
concept Updated 2026-07-09 Tags: Skills, Agents, Workflow

AI Skills

AI skills are reusable instructions and process packages that help an AI complete a defined class of work. In 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, strong skills are described as codifying implicit knowledge: boundaries, steps, tools, quality standards, and required context. In 阿里千问离职余震，在几万人的铁球里如何体面生存, skills appear more operationally through background delegation and debate-style agent analysis. 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？ adds a market and ecosystem view through PPT skills, Feishu-related skills, Open Cloud, and Code Pilot. 对话 MiniMax 闫俊杰：M3、10X 计划、10T 模型、和智能的终局 adds an engineering-quality use case: He Tao suggests packaging Clean Code, Google best practices, Amazon best practices, and similar standards into skills or context for agents. 探秘 Claude Code，搞懂 Agent Harness｜对谈来新璐 adds a memory-overlap view: in Claude Code-style systems, skills, memory files, task reports, experience, and SOPs can share markdown/file mechanics and be updated by agents over time.

1 人公司，扛 5 个人的活，还要管 50 个 Agents？｜S10E18 adds Cang Shifu’s product-market boundary around skills. His PPT skill drew meaningful use, but he treats raw 2C skill monetization as hard because open distribution, channel uncertainty, and unclear payment norms limit direct capture. The more plausible routes in the source are 2B customization, cooperation, service packaging, or skills embedded inside a broader agent/tool relationship such as Code Pilot.

当我们在讨论 Harness 的时候，我们在讨论什么 | 深度对谈: MiniMax × Hermes Agent adds Hermes Agent as a clearer memory-to-skill case. The source says successful workflows can be saved as skills so the agent can reproduce the correct path next time, making skills part of Agent Self-Evolution rather than only user-authored prompt packages.

AI 会写代码了，为什么你还是做不出产品？ adds an everyday engineering-practice version. Tests, logging policies, documentation expectations, code-review prompts, audit rubrics, and business handoffs can act like skills when they repeatedly guide AI through the same class of work.

Vol. 161 从开发自己的 OpenClaw 聊起 adds a personal-agent version through Open Claw. Justin Yan treats skills as lighter than protocol-heavy tool integration and describes a stronger loop where an agent can inspect home services, infer what they can do, and write new skills for itself. The same source adds a security split between trusted skills and agent-written skills, making Agent Permission Boundaries part of skill design.

Vol. 165 做客声东击西：「龙虾」和 vibe coding 正如何改变我们的思维 adds 王俊玉’s “trainable newcomer” interpretation. Skills are not only memory about user preferences; they preserve how a task should be done, so feedback can become a reusable work method for an agent operating in a company or personal workflow.

EP124 为什么 Agent 时代，CLI 反而成了最优解？⚡ adds a CLI-composition version through Podwise. Skills sit above Agent-Optimized CLI atoms such as search, process, get, and export, letting the user describe an outcome while the agent chooses the command sequence. The source treats this as a practical shift from operating tools to specifying tasks.

EP127 从 Skills 到自动化工作流，论 Agent 如何接管真实生产力 ⚙️ adds a selection and maintenance lens. A skill should survive because it handles repeated work, compensates for a weak domain, or helps an agent verify real output; popularity alone is a weak signal. The episode also treats self-written skills as disposable workflow assets: if a prompt, check, support reply, podcast-processing step, or deployment routine repeats enough times, it can become a skill and later be deleted or revised when the workflow changes.

20 个问题，搞懂 OpenClaw：爆红机制、本质变化、创业机会 adds an ecosystem and consumer-agent version through Open Claw. 鸭哥 describes maintaining personal preference or “axiom” skills for research and synthesis, while the episode treats open-source skills, user PRs, skill markets, and AI-readable smart-home APIs as ways ordinary users can expand what an agent can do.

Vol. 170 Fable 5 重出江湖，GPT 仍需努力 adds a manual skill-selection contrast through GrillMe Skills. The source says a full automatic bundle such as Superpowers can help non-experts by enforcing the whole software process, but experienced users may prefer to manually trigger requirement grilling, specs, ADRs, PRDs, or issue decomposition to save tokens and avoid over-structuring small tasks.

Vol. 167 Token 如流水，Agent 似朝阳 adds a product-prototype path from conversation to skill. The hosts describe tweaking an agent’s behavior through natural language until it reliably handles article selection, translation, note saving, or todo aggregation; once repeated, that behavior can become a reusable workflow asset even if it is not identical to a formal skill file.

E155.似乎没什么人再提「AI 泡沫论」了 adds the business-software and personal-skill disruption angle. The episode argues that B2B software know-how can become explicit, shareable, and repackaged as skills, weakening software moats based on hidden workflow knowledge. It also gives a personal example where a data-visualization workflow becomes a skill, creating the feeling that years of craft can be compressed into a reusable procedure.

当可靠的代码变成了偶尔发疯的OpenClaw，我们未来的工作范式变迁 adds the personal-training and skill-safety angle. The hosts describe writing product-design and PRD skills, using memory files to preserve project norms and testing methods, and letting agents update experience after tasks. The same source warns that skills from third parties or generated by agents can carry hidden behavior, so skill usefulness has to be paired with trust, permission tiers, and review.

E163.要完了？不！是要玩了！论养AI的心态与习惯 adds the ordinary-user operating-manual angle. 品哥 describes skills as a way to onboard a high-intelligence but background-poor AI worker: define the job, process, prior pitfalls, acceptance standards, user style, and reusable materials so the agent can do the work without being re-taught every session.

137. 对洪乐潼的4小时访谈：AI for Math、把数学变成Lean、数学天书中的证明、直觉、被创造与被发现的 adds a theorem-proving case through Axiom. Hong Letong / 洪乐潼 describes using subagents, learning from experience, and skills as part of replacing expensive search in AI For Math, which makes skills a way to preserve proof tactics, tool habits, and successful formalization procedures rather than only office workflows.

138. 对罗福莉3.5小时访谈：AI范式已然巨变！OpenClaw、Agent范式很吃后训练、卡的分配、组织平权 adds a model-team case through Luo Fuli / 罗福莉. She treats skills as a route for converting internal team norms, management questions, research habits, and simulated user-agent workflows into Agent Post-Training material that the model could not learn from generic pretraining alone.

263.Sora死了，Adobe跌了，美图何去何从？ adds a commercial capability-distribution case through Meitu / 美图. Instead of treating skills only as user-authored instructions or agent memory, the episode describes Meitu packaging cutout, outfit-change, image-to-video, action-guidance, and resize capabilities as callable skills inside an agent ecosystem, making skills part of To-Agent Distribution.

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 an internal-operations case through invoko.ai / Invoqo. 梦琪 / Mengqi says the team’s later growth skill worked better than the earlier external 2B product because the team had complete context, direct feedback, and daily maintenance incentives around its own influencer/growth workflow.

Key Claims

  • A useful skill is more than style imitation.
  • Skills become more valuable when tied to Agentic Workflow and real tools.
  • Subagent Workflow shows one practical skill pattern: delegating large or complex work to background agents and integrating results later.
  • Skill commercialization may work better as skill as a service than as raw skill-file sales.
  • Simple skills may be absorbed by stronger models, while trusted agents, brands, and tool ecosystems may own the longer-term user relationship.
  • Skills become more useful when they can call Agent-Facing Interfaces rather than only instruct a chat model.
  • Engineering skills can help convert software taste, review standards, and maintainability rules into reusable AI Coding Verification context.
  • Skills can sit inside an Agent Harness as selectively loaded files whose descriptions help the model decide whether to pull in the full procedure.
  • Skills can also be generated from successful agent work traces, where Persistent Agent Memory compresses experience into reusable procedures.
  • Practical skills can encode AI Engineering Thinking: require tests before implementation, collect logs before debugging, and make domain acceptance criteria explicit.
  • Personal-agent skills need permission tiers because a markdown-like procedure can still expose private data or operate a powerful tool.
  • Skills can make agent training resemble onboarding: a human corrects an agent’s work method, and the corrected procedure becomes reusable in future tasks.
  • Skills become more powerful when the underlying tool offers small, stable CLI actions that can be discovered, combined, and debugged without custom integration code.
  • Skills can turn users into partial developers because a reusable procedure, API description, or tool instruction can become part of the agent’s future capability.
  • Verification, testing, and release skills may matter more than generic framework skills because they let the agent prove that work is actually done.
  • Skills can become Routine Agent Automation when paired with scheduled execution for email, notes, analytics, cost monitoring, or research workflows.
  • Skill bloat is a real context and behavior risk; unused fashionable skills should be removed when they do not match actual recurring work.
  • The stronger the model, the more important skill loading discipline becomes: Fable 5 can do more in one pass, but automatically wrapping every task in a heavy process can waste AI Inference Cost Structure.
  • Skills can emerge from repeated personal-agent conversations, but should be hardened into deterministic tools or explicit procedures once the workflow becomes stable enough to justify engineering.
  • Skills can expose business-process know-how that previously lived inside software vendors, consultants, or individual experts.
  • Skills are also a training interface: users teach a recurring method once, but the resulting procedure should still be inspectable and scoped because it can operate tools or private data.
  • Skills can also encode a user’s “what good looks like” through Output Quality Gates, not only the steps of execution.
  • In formal-math systems, skills can preserve reusable proof and formalization procedures, but their outputs still need Lean Theorem Prover-style checking.
  • In model teams, skills can become a private data and training interface because they encode organizational procedures, evaluation standards, and repeated agent workflows.
  • In application companies, skills can also become a distribution surface: a vertical tool can expose stable capabilities to agents without requiring the user to open the original app.
  • S10E18 adds a commercialization caveat: skill use or downloads do not automatically create a business unless the skill has a channel, buyer, trust path, and repeatable delivery model.
  • Internal skills can outperform external product attempts when the builder owns the context, acceptance criteria, and follow-up loop.

Connections