Podcast Atlas
entity Updated 2026-07-07 Tags: Host, Builder, Agents

Justin Yan

Justin Yan is a 枫言枫语 host who uses Vol. 160 一年多以后，再聊AI写代码Vibe Coding to revisit a year of AI-assisted coding through NewSpot, YOLO-style agents, AI search, testing, and the boundary between generated code and product judgment. Vol. 161 从开发自己的 OpenClaw 聊起 adds his simplified Open Claw-like personal agent: he approached the project largely through Vibe Coding, focusing less on reading every line of code and more on learning the architecture, product assumptions, and safety tradeoffs exposed by the build. Vol. 162 科技快乐星球44: 新模型“SOTA们”齐贺新春 adds Justin as a hands-on model/tool evaluator who compares Codex, Claude Code, Gemini, Xcode, and domestic models through Model Workflow Fit rather than treating SOTA as a single winner. Vol. 164 从苹果聊到软件未来：Agentic Software 真的要来了？ adds Justin’s broader software-design view: Agentic Software should not be reduced to adding an AI button, and old SaaS may need Atomic Capability Services plus human review, communication, and taste. Vol. 165 做客声东击西：「龙虾」和 vibe coding 正如何改变我们的思维 adds Justin in a crossover role with 声东击西, where he frames OpenClaw as an early historical turn toward agentive software and warns that vibe coding still needs foundations, engineering judgment, and production responsibility. Vol. 166 闲聊: 从 Gemini 到 AI 的加速与混沌 adds Justin as a heavy hands-on user of coding agents and agent orchestration during a period of fast AI change. Vol. 169 高考只是个开始，Don’t Waste Your Life adds his student-and-parent-facing view of College Major Choice, AI As Tutor, and the need to balance income responsibility with long-term interest. Vol. 170 Fable 5 重出江湖，GPT 仍需努力 adds his more mature workflow split: use Fable 5 for planning, requirements, and review judgment, then use Codex and skills for execution and validation. Vol. 167 Token 如流水，Agent 似朝阳 adds Justin’s view of Codex as a remote personal technical assistant and of Open Claw/Hermes Agent-style IM agents as fast product-idea test beds.

Source Position

  • Justin treats making an agent as more revealing than merely using one because the build exposes tools, skills, channels, permissions, and trigger design.
  • In Vol. 160, he treats NewSpot as proof that AI can write most implementation code while the human still owns architecture, review, tests, product taste, and final acceptance.
  • He warns that natural language should not make users treat AI as a wish-granting person; clear expression and review remain part of the toolchain.
  • His implementation focuses on Telegram rather than the multi-channel approach attributed to Open Claw.
  • He connects personal-agent experimentation back to NewSpot, imagining deeper multi-source research and better news insight workflows.
  • He argues that expert judgment, personal taste, and distinctive curation remain important even as AI can build more generic apps and features.
  • In Vol. 164, he separates the excitement of one-week demos from the slower work of shipping, feedback, and product readiness.
  • In Vol. 165, he distinguishes AI’s strength on quantifiable, repeatable work from harder high-end judgment, taste, and creative expression.
  • In Vol. 166, his examples include Superpowers planning flows, Codex-built self-use tools, Cloudflare operations, and physical/attention costs from supervising agents for long stretches.
  • In Vol. 169, he treats programming as still worth learning for students who enjoy building, while warning that AI-generated prototypes still differ from maintainable software.
  • In Vol. 170, he treats GrillMe Skills as a way to keep the useful parts of structured planning while avoiding automatic Superpowers overhead on small tasks.
  • In Vol. 167, he uses multi-session Telegram-like agent workflows, article collection, calendar/reminder/Obsidian summaries, and daily todo generation to show how IM Agent Interfaces can become small, configurable products.
  • In Vol. 162, he treats model choice as an operating decision: Codex can be stronger for review and planning, Claude Code can be faster for bounded execution, and Xcode integration is useful where IDE context matters.

Connections