Offline Dev Talks | 离线思考

A working software engineer spots a paradigm shift in his own side projects: we used to ship software, but now agents hand you the doc, the podcast, the image directly. What does that mean for the job of writing code?

00:07  Two cases, one contrast: building software vs shipping the result
01:45  Software is just the middle product: users want the outcome
03:07  Two breeds of agents: the software-builder vs the result-deliverer
04:35  The Coding Agent's blind spot: prompts wired for building
06:05  The chat is the asset: repeat interactions become Agent Skills
07:34  Skill is the new app, Runtime the new platform
08:53  The middle step collapses: no more third-party developer
10:06  The role thins out: demand drops, two kinds remain
11:24  Real AI Native: build Skills that deliver results directly

一个软件工程师在自己的两个小项目里，撞见了一场范式转移：以前我们交付软件，现在 Agent 直接把你要的文档、播客、图片递到手上。这场变化，对「写代码的人」到底意味着什么？

00:08  两个 case 的范式反差：造软件 vs 直接给结果
01:29  软件只是中间产物：用户要的是结果不是软件
02:36  两类 Agent 分野：造软件 vs 直接给结果
03:52  Coding Agent 的软肋：系统提示词为造软件而生
05:05  对话即资产：高频交互沉淀成 Agent Skill
06:15  Skill 即应用，Runtime 即平台：Agent 是 OS 之上的新抽象层
07:18  中间过程被压缩：不再依赖第三方开发者
08:22  岗位会被稀释：需求大减，但两类软件工作仍在
09:35  真正的 AI Native：在 Agent 平台建 Skill 直接产出结果

A small story: I gave my note folders recognizable icons so I can tell them apart at a glance. But the real lesson isn't the feature — it's why it was me, not the AI, who first noticed the friction.

00:09  My note workflow: one folder per note
00:52  Folder names use the note ID: unique and lean
01:28  The pain: a wall of number IDs, hunt folder by folder
02:05  The spark: give the folder an icon you recognize at a glance
02:37  Icon design: a visual plus the title, instant ID
03:05  Building it with AI: speak the idea, settle the plan, it executes
03:36  Collaboration snags: no Finder access, font too small
04:09  The real insight: the change is trivial, spotting the problem isn't
04:42  The loop: from spotting to solving, the human starts it
05:07  A reminder for everyone: spotting problems is yours to train

这期从一个真实的小功能讲起：我给一堆数字 ID 命名的笔记目录，换上了能一眼认出的图标。但真正值得聊的不是这个功能本身，而是——为什么是我，而不是 AI，先看到了那个「找起来好麻烦」的问题。

00:08  我的笔记工作流：每篇一个目录
00:45  目录名用笔记 ID：唯一又精简
01:17  痛点：一串数字 ID，找产物得逐个翻
01:51  灵光一现：给目录换个一眼认出的图标
02:20  图标设计：视觉图加标题，一眼认出
02:44  和 AI 一起落地：口述想法、定方案、它执行
03:10  协作小坑：够不到 Finder，字体太小我让它调大
03:41  关键认知：功能改动不重要，发现问题才重要
04:09  闭环：发现问题到共同解决，人才是起点
04:35  给所有人的提醒：发现问题得自己练

I deleted every script, JSON config and validator from my AI image pipeline — and the results got dramatically better. This episode is about the deeper philosophy of working with AI: restraint.

00:07  Wrapping AI in scripts and validators made its images worse

01:23  Automation strangled the model's artistic ability

02:27  Deleting the rules unlocked the AI

03:37  Claude Code cut 80% of its system prompts

04:31  The engineer's habit of constraining everything

05:37  Work with AI as a partner, not a commander

06:46  No system is 100% controllable

08:01  Too many constraints create complex systems and surprises

09:08  Two rules: don't constrain what you don't master, keep systems simple

我把给 AI 出图的那套脚本、JSON、校验器全删了，AI 反而画得更好。这期聊聊跟 AI 合作的底层哲学——克制：真正的高手，懂得在不懂的地方放手。

00:08  给 AI 套脚本和校验器，出图反而越来越呆板
01:20  自动化把大模型的艺术创作能力抹杀掉了
02:18  删掉规则后 AI 自由发挥，效果反而惊艳
03:16  Claude Code 适配时删掉了 80% 的系统提示词
04:07  软件工程师的「代码约束一切」通病
05:07  跟 AI 是合作不是指挥
06:02  系统不可能 100% 可控
06:56  约束太多会变成复杂系统，产生涌现行为
07:47  两条正解：别在不擅长处加约束，系统做简单