Alchemy Adopts OpenAI Codex Code Review to Prevent Production Outages

Alchemy采用OpenAI Codex代码审查以预防生产故障 AI编程智能体在部署前捕获生产级软件缺陷，并将初创公司的产品开发周期从18个月缩短至一周以内。 Matias Castello分享了Alchemy如何在整个组织中集成Codex，通过基于Slack的文档更新以及预防生产事故的自动化追溯性代码审查，重塑内部开发者的工作流程。他详细介绍了自己的异步构建策略，利用定制的智能体配置文件与Linear集成来派发任务，使其在夜间自主运行。此外，Castello还展示了多个副业项目——包括通过Codex App Server在Apple Watch上实现的语音触发工作流，以及一个名为Snapcat的长期个人模型基准测试——以展示当前的AI能力如何消除交付软件的传统障碍。 Alchemy 在一次事后复盘中发现 Codex 能瞬间识别出遗留系统迁移中隐藏的竞态条件 Bug，彻底改变了团队认知，证实 AI 已具备生产级代码审查与事故预防能力。 (1:51) AI 辅助开发极大降低了创业壁垒，过去需要 15 人工程团队耗时一年半才能交付的 MVP，如今独立构建者借助 AI 不到一周即可完成。 (6:29) 通过标准化配置文件（如 agent.md）消除隐性假设，开发者可将 Linear 转为 AI 执行看板，由 Codex 自主规划架构、拆解任务并闭环交付 150 多个 Issue。 (10:44) 基于开源的 Codex App Server，开发者可构建移动端工作流，仅凭 Apple Watch 录制的一句语音便在后台自动执行编码并提交 GitHub Pull Request。 (20:29)

Alchemy Product Lead Matias Castello and OpenAI's Romain Huet argue that advancing AI capabilities and agentic workflows have dismantled traditional software engineering barriers, allowing solo builders to research, architect, and deploy production software asynchronously. They contend that maximizing these tools requires treating autonomous agents as infrastructure consumers while adopting disciplined prompt communication rather than viewing model errors as technical limitations. Automated Code Review and Production Infrastructure (1:51) At Alchemy, internal AI adoption shifted from basic documentation updates in Slack to critical engineering safeguards following a postmortem on a legacy migration race condition. Testing confirmed that Codex flagged the bug retroactively, prompting engineers to integrate the model directly into GitHub pull request workflows. Romain Huet cited a Datadog retrospective from January indicating that Codex could have prevented over 20% of production incidents, while both Huet and Castello projected that advanced models like GPT-5.5 could eventually catch between 50% and 90% of system incidents before deployment. This capability shift has altered platform architecture. Alchemy now operates under the assumption that 100% of human developers rely on AI assistance, while autonomous agents are increasingly direct consumers of developer infrastructure—provisioning API keys, selecting implementations, and executing on-chain transactions without human intervention. Compressing Startup Timelines and Democratizing Roles (6:29) Castello and Huet contrasted traditional startup execution from roughly 2018 with current AI-native development. Whereas launching a minimum viable product previously required raising capital and managing 3 to 15 engineers over 6 to 18 months, Castello estimated that a solo builder using AI can now assemble an MVP in under a week with minimal subscription costs. The primary development friction has consequently shifted from manual code syntax and hiring to idea clarity and product direction. To scale product management internally, Alchemy deployed a centralized repository of reusable "skills"—structured prompts and workflows for market research, user feedback synthesis, and PRD authoring. This shared library enables non-product managers across technical and non-technical teams to execute product tasks with standardized, PM-level fidelity. Asynchronous Delegation via Standardized Configurations (8:04) To counteract "builder's anxiety," Castello transitioned from continuous manual coding to an asynchronous model where Codex executes multi-hour tasks overnight. To eliminate errors caused by unstated assumptions, he standardizes coding style, architecture preferences, and conventions in upfront agent.md configuration files. In a custom macOS and iOS writing application, Codex referenced these specifications to autonomously generate an architectural roadmap, populate over 159 Linear issues, and implement pull requests without manual issue writing. This workflow also automates early-stage feature exploration. By directing Codex to conduct competitor research and structure experimental features behind modular feature flags, Castello can test and toggle working prototypes directly within the application each morning, removing human bottlenecks from the ideation cycle. Ubiquitous Entry Points and Multimodal Execution (18:51) Leveraging the Codex CLI and Codex App Server, Castello extends agent workflows beyond desktop IDEs into mobile, wearable, and multimodal environments. Through "OpenClaw," a home server setup running an agent named Lou, dedicated Discord channels connect directly to specific GitHub repositories, enabling remote code generation from mobile chat messages. Similarly, a custom Apple Watch complication transcribes 10-second voice memos, identifies repository context, and generates pull requests autonomously. Multimodal execution also addresses UI and device management tasks. In one weekend experiment, Codex used computer use over SSH to parse URLs and automate data entry on a Raspberry Pi web admin console. To benchmark model progress against "Snapcat"—a cat-selfie app originally built by a five-person team over a 24-hour hackathon a decade ago—Castello used GPT-5.5 to one-shot the core logic and iteratively convert prompted UI design images directly into functional front-end code. Castello advises builders to assume technical feasibility and treat unexpected LLM outputs as prompt communication failures rather than tool limitations. However, while solo execution speeds have accelerated dramatically, the workflows presented depend heavily on strict configuration files, explicit constraints, and continuous curation through feature toggles to prevent autonomous drift.

Alchemy 产品负责人 Matias Castello 在访谈中阐述了软件工程与初创企业构建范式的根本转变：通过将 AI 模型融入深度代码审查、标准化跨职能技能库及全天候异步代理工作流，个人开发者如今能在数天内完成以往多名工程师耗时数月的交付目标，而现代软件基础设施的设计重心也已转向同时服务人类开发者与自主 AI 代理。 代码审查与技能库：AI跃迁为核心支柱 (1:51) Alchemy 最初仅在内部 Slack 频道通过 Codex 进行自然语言交互式文档更新，以规避在本地拉取代码仓库的繁琐流程。内部对大模型能力的信任转折点源于一次生产事故的事后复盘：在一次大规模遗留代码迁移引发竞争条件（race condition）故障后，工程团队测试发现 Codex 在审查事故发生前的代码时，能够立即精准标记出该缺陷。这一结果促使工程师开始在 GitHub Pull Request 中将模型作为人类对等同行进行代码协同审查。OpenAI 团队的 Romain Huet 补充表示，Datadog 在 1 月份的复盘中也发现 Codex 能预防超过 20% 的线上故障；两人预测，随着 GPT-5.5 等模型的发展，自动化代码审查有望在上线前拦截 50% 甚至高达 90% 的系统故障。 在研发流程之外，Alchemy 还将大模型应用扩展至产品管理全流程。团队在跨部门共享代码库中沉淀了包括编写产品需求文档（PRD）、提炼客户反馈及竞品调研在内的“技能”（skills）资产。这套标准化的提示词与流程库，让非技术人员与工程人员均能以专业产品经理的标准执行复杂的分析任务，实现了产品管理能力的去中心化与内部民主化。 基建转型与初创效率：面向AI代理的开发 (4:08) 随着开发环境的演进，Castello 提出开发者平台在架构设计上面临着底层逻辑重构。Alchemy 的平台建设建立在两项现实判断之上：其一，100% 的人类开发者均已在 AI 辅助下编写代码；其二，完全自主运行的 AI 代理正在成为平台基础设施的直接调用者，它们能够自主注册账户、配置凭证密钥、决定技术实现方案并独立发起区块链交易。因此，平台必须专门为无人工介入的自主代理打造完全顺畅、零摩擦的 API 接口体系。 在企业创办效率方面，Huet 与 Castello 比较了过去与当前的开发成本差异。Huet 回忆称，约 6 至 7 年前，他曾需要 15 名工程师组成的团队耗时约一年半才能交付支持广泛客户的 V1 版本产品。Castello 则评估认为，过去需要 3 至 4 名全职工程师数月才能搭建的最小可行产品（MVP），如今个人开发者配合 AI 在一周内即可独立交付。创办企业的核心摩擦力已彻底从代码语法实现、人员招聘和早期资本筹集，转移到对产品意图的清晰构思与方向决策上。 异步自主开发：规约约束与功能开关实验 (8:04) 为克服总觉得未能充分利用前沿 AI 技术的“构建者焦虑”，Castello 设计了一套异步后台开发体系。他将日常碎片时间（如午餐小憩、外出饮酒或睡前）作为分发任务的“Codex时刻”（Codex moments），由模型在后台自主执行长达数小时的需求调研、代码编写和测试流程。为了确保模型产出的工程质量，他在每个项目根目录中维护标准化的 agent.md 规范文件，详细约束架构偏好、编码风格与设计标准。Castello 强调，模型的不良输出主要来自未经明说的隐藏假设，清晰的前置约束是让代理对齐工程标准的核心。 在任务管理层面，Castello 将 Linear 面板作为代理的执行控制台。在某个包含 159 项任务的项目中，所有工单均由 Codex 依据架构里程碑自动生成、认领编码并提交 PR 关闭，无需人工手动编写一条任务。此外，该流程还支持模块化特性探索：模型能在夜间自主分析竞品功能点并编写对应模块，将其统一封装在应用内的特性开关（feature flags）中，使开发者在早晨仅凭勾选开关就能直观评估实验特性，不再受限于个人的创意发散瓶颈。 随行触发与系统操控：从手表到Snapcat (18:51) 在交互入口上，Castello 借助开源的 Codex App Server 构建了多终端控制链路。他在家用独立主机上运行名为 Lou 的 OpenClaw 代理服务，将特定 Discord 频道与对应的 GitHub 代码仓库绑定，实现在移动过程中通过聊天即时下发工程指令。针对修复落地页拼写等微型单次任务，他开发了 Apple Watch 表盘组件与 iPhone 配套应用，通过录制 10 秒语音便能完成意图识别，结合已绑定的账户上下文远程生成拉取请求（PR）。 在系统级交互与模型能力极限测试方面，Castello 验证了计算机使用（Computer Use）功能，命令 Codex 通过 SSH 远程登录连接电视的树莓派设备，以模拟光标和窗口操作的方式自动在 Web 控制台中录入域名。此外，他将十年前 5 人黑客松团队耗时 24 小时编写的猫咪自拍应用“Snapcat”作为基准测试：利用 GPT-5.5 结合先生成界面设计图、再由 Codex 转换为前端代码的图生布局流程，模型仅用单次提示词（one-shot）在一夜之间交付的核心逻辑与界面精致度，已显著超越当年 5 人团队的全天成果。 Castello 总结了 AI 构建者的三项核心思维：坚信目标可达成、克服对自身技能边界的心理障碍，以及在模型输出未达预期时放下自负，优先反思指令沟通与上下文约束的不足而非工具本身的局限。Huet 对此表示赞同，认为开源运行框架与模型自主能力的结合，正在使任何人独立构建任何复杂软件成为现实。

## Background 本期视频由 OpenAI 于 2026 年 6 月 3 日发布在官方 YouTube 频道的《Builders Unscripted》系列第三期。节目由 OpenAI 开发者体验负责人 Romain Huet 主持，特邀 Alchemy 产品负责人 Matias Castello 展开对谈。节目简介指出，本集核心围绕“你完全可以直接动手构建一切”的心态展开，探讨了 Matias 团队如何利用 Codex 推进代码审查与产品管理、如何借助 Codex App Server 开发个人副业，以及为何他选择持续重建 Snapcat 应用作为评估新模型的个人基准。 ## TL;DR Matias 分享了非科班出身的自己如何借助 Codex 在 Alchemy 推动全员编码、精准复盘线上事故并在 PR 审查中将 AI 视作资深队友，进而推动基础设施向支持自主智能体演进。为了攻克项目构建的心理焦虑，他探索出了一套高度自治的异步工作流，通过 agents.md 配合 Linear 与 Feature Flag，让 Codex 在夜间自主完成竞品调研、里程碑拆解及功能代码开发。他现场演示了基于 Codex App Server 构建的全局语音写作助手、Apple Watch 腕上任务派发链路以及用 Computer Use 接管树莓派的自动化操作。最后，他以跨越十年的个人评测项目 Snapcat 为例展现了 GPT-5.5 多模态代码生成的飞跃，并鼓励创作者放下自尊、坚信想法皆可实现。 ## Atmosphere & style 整场对谈洋溢着浓厚的技术极客氛围与强烈的时代乐观主义，节奏明快而充满启发。Romain 与 Matias 的互动自然松弛，两人在分享“戴耳麦被当成话务员”或“看 Computer Use 挪光标入迷”等日常开发趣事时笑声不断，展现出同行之间强烈的共鸣。在密集的实操演示与真诚的心路剖白交织下，观众能切身感受到个体从被繁琐技术绑架到彻底释放创造力时的兴奋与震撼，具有极强的动手感染力。 ## Bulletin - [12:55](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=775s) Alchemy 产品负责人 Matias Castello 揭示非工程师如何借助 Codex 与 GPT-5.5 实现个人高产，通过将个人规范沉淀为 agents.md，实现了在 Linear 看板中自动拆解并交付逾百个 Issue 的全自动化研发闭环。 - [01:51](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=111s) 企业引入大模型的关键转折点在于代码审查与事故复盘：Codex 曾成功提前捕获复杂架构迁移中隐蔽的竞态条件，促使工程师形成提交 PR 后反复调用 @Codex review 的人机结对协作流。 - [08:04](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=484s) 借助异步开发与 Feature Flag 机制，开发者可通过 Codex 在夜间自主调研竞品并打包前 10 大核心功能的实验性开关，实现“睡前指派任务、次日晨间直接验收试用”的无感交付体验。 - [20:15](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1215s) 基于 Codex App Server 与开源架构，开发者可将编码任务延伸至 Apple Watch 与 Discord 等非传统终端，实现仅凭一句语音指令即可在后台自动向 GitHub 提交 Pull Request。 ## Key takeaways - [00:13](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=13s) 非科班工程师借助底层大模型能力的跃升，在项目交付数量与速度上已能大幅超越传统全职工程团队。 - [00:58](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=58s) 在 Slack 等办公协同工具中接入 Codex 修改文档，是让团队全员以极低门槛拥抱 AI 的最佳落地切入点。 - [01:51](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=111s) 线上事故复盘（Postmortem）验证了大模型应对海量复杂生产代码的能力，Codex 能精准捕捉人类易忽略的竞态条件等致命 Bug。 - [04:04](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=244s) 基础设施与开发者平台的范式转移：既要适配使用 AI 辅助编码的人类工程师，更需构建面向自主决策智能体（Autonomous Agents）的全自动交互接口。 - [06:28](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=388s) AI 辅助彻底改变了创业研发的经济学模型，过去需要十数人团队耗时数月至一年半的 MVP 原型，如今单人一周即可交付。 - [08:04](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=484s) 为缓解“构建焦虑”，可构建睡前派单、醒来验收的离线自治工作流，Codex 会自主完成技术规划、编写测试与部署打包。 - [10:08](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=608s) 利用碎片化时间触发任务，配合后台并发执行，能够让个人在 Linear 看板上同时推进十余个高复杂度的进行中项目。 - [10:59](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=659s) 通过全局快捷键唤起基于 Codex App Server 的写作助手，能够将口语化原声即时重写为高质量的专业沟通文本。 - [12:55](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=775s) LLM 产出不达预期通常源于隐性假设偏差，将工作规范与架构偏好前置写死在 agents.md 中是建立全自动项目管理闭环的前提。 - [15:06](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=906s) 全自主竞争调研与夜间功能孵化：利用 Codex 耐力拉满的特性自动抓取竞品亮点，并以 Feature Flags（特性开关）的形式打包供人工筛选裁决。 - [18:51](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1131s) 借助常驻后台的 AI 助手与 Discord 机器人，开发者在移动端即可随时触发远端主机的代码拉取与执行链路。 - [20:15](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1215s) 通过 Apple Watch 语音录制指令，配合 iPhone 端意图解析及后端 Codex App Server 会话，几分钟内即可在 GitHub 自动创建对应 PR。 - [22:52](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1372s) OpenAI 开源 Codex CLI、Harness 与 App Server，旨在鼓励社区在官方功能覆盖前自由扩展端到端智能体在各种形态设备上的可能性。 - [24:01](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1441s) Computer Use（电脑使用）在自动化运维场景表现亮眼，能自主通过浏览器 SSH 登录树莓派并完成繁重的机械式数据迁移任务。 - [25:22](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1522s) 多模态 UI 构建新范式：先通过图像生成模型产出视觉参考图，再交由 Codex 依据设计稿像素级还原前端代码。 - [28:24](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1704s) 给创造者的三大心智建议：坚信功能在技术上可行、坚信自己有能力完成、遇到偏差时放下自尊反思是否未能准确传达意图。 ## Detailed walkthrough ### [00:13](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=13s) 跨越领域的构建者：非工程师背景的高产实践 OpenAI 开发者体验负责人 Romain Huet 热情欢迎了 Alchemy 的产品负责人 Matias Castello。Matias 的职业生涯横跨多个截然不同的领域：他曾涉足面向终端用户的消费级产品，在早期的 Facebook 参与过开发者平台的建设，如今则在 Web3 与加密基础设施龙头 Alchemy 执掌产品。然而，Romain 指出最令人惊叹的一点是，Matias 并非科班出身的软件工程师，但在实际产出的项目数量上，却远远超过了许多全职工程师。 面对这种反差，Matias 坦言这完全归结于对诸多事物保持广泛的好奇心，以及多年来不间断地亲自尝试与摸索。随着底层大模型能力的持续跃升，他驾驭这些工具的能力也水涨船高。他不仅在全职工作中高度依赖 AI，在业余生活中也保持着极其高频的项目构建节奏。 --- ### [00:58](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=58s) 引入 Codex：从 Slack 中微调文档开始 谈及在 Alchemy 引入 AI 的“灵光一现”时刻（Aha Moment），Matias 回忆大约一年前，团队最初是在企业 Slack 中接入了 Codex 机器人。此前，如果团队想要修改开发者文档中的哪怕一小段内容，都必须在本地拉取代码、启动本地环境运行整个文档站点，流程既冗长又繁琐。而通过将 Codex 引入内部 Slack 的文档频道，任何人只需直接以自然语言指示 Codex 进行更改，它就能自动完成修改并推进流程。 Romain 笑着回应称，OpenAI 内部也一直在以完全相同的方式使用这项能力，虽然现在构建了更加成熟复杂的自动化方案，但“在 Slack 里直接使唤模型改文档”始终是推动全员拥抱 AI 最低门槛、最立竿见影的切入点。这一小步让整个 Alchemy 团队第一次真切体会到了 AI 落地日常工作流的便利。 --- ### [01:51](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=111s) 转折点：代码审查与追溯性 Bug 复盘 真正让整个技术团队建立信心的分水岭，发生在代码审查（Code Review）流程中。Alchemy 曾经遭遇过一次线上小事故，起因源于几个月前完成的一次涉及海量代码的大规模架构迁移。在迁移后的复杂代码库中潜藏了一个竞态条件（Race Condition）。在团队针对该事故进行内部复盘分析（Postmortem）并修复问题后，有人突发奇想：何不让 Codex 逆向审查当初引发问题的 Pull Request，看看它是否能提前识别出这个 Bug？ 测试结果令人震惊——Codex 准确指出了那个微小而致命的竞态条件。团队随后又用多个历史问题进行了验证，反复确认了其强大的分析能力。几天后，Matias 便在 Slack 中观察到一位资深工程师已经把 Codex 当作全职队友协同工作：他在提交 PR 后，直接在评论区调用 `@Codex review`，根据模型提出的修改建议逐一优化代码，随后再次触发审查，在多轮高强度的推拉讨论中反复打磨代码。这彻底打破了此前大家认为“大模型无法胜任海量、复杂且高可靠性生产环境代码库”的偏见。 ```mermaid flowchart TD accTitle: Codex代码审查与反馈循环 accDescr: 展示工程师如何利用Codex在PR中进行多轮互动式代码审查与修复 subgraph Postmortem["事故复盘与验证"] A["发生竞态条件事故"] --> B["定位并修复Bug"] B --> C["回溯运行 Codex 审查"] C --> D{"Codex 成功命中 Bug"} end subgraph Workflow["日常研发协同流"] E["提交 Pull Request"] --> F["触发 @Codex review"] F --> G["Codex 提出修改意见"] G --> H["工程师针对性修改"] H --> F end D -.->|"打破偏见，建立信任"| Workflow ``` - **行业验证数据**：Romain 提到，Datadog 在同年 1 月也公开过类似统计，发现团队经历的线上事故中有超过五分之一完全可以通过 Codex 自动审查来预防；随着 GPT-5.5 级别模型的演进，这一拦截比例如今完全可以达到一半乃至十之八九。 --- ### [04:04](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=244s) 演进的开发者生态：赋能人类团队与自治智能体 在 Alchemy 的日常产品管理中，Matias 和团队已经将 Codex 嵌入了全流程。他们建立了一套共享的企业级“技能库”（Skills Repo），涵盖了自动撰写产品需求文档（PRD）、提炼分析海量客户反馈等核心场景。这种模式不仅让现有的产品经理工作成倍提速，更打破了职能壁垒，让没有受过专业 PM 训练的工程师和其他岗位成员也能无缝输出高质量的产品文档。 随着研发模式的剧化，Alchemy 对自身开发者平台的定位也在经历双重范式转移： 1. **全面假设 100% 的人类开发者都在使用 AI**：作为底层基础设施服务商，必须重塑 SDK、API 和文档设计，确保类似 Codex 这类辅助编码工具能以最丝滑的方式调用 Alchemy。 2. **面向全新的自主智能体（Autonomous Agents）构建工具**：未来的“开发者”不再只是人类，还会包含全自动运行的 AI 智能体。这些智能体有自主决策权，需要能够全自动注册账号、完成集成，并在区块链上执行复杂的链上交易或任务。两者的需求目前仍存在明显差异，产品团队的核心使命就是同时兼顾并逐步融合这两类开发者。 --- ### [06:28](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=388s) 创作者的时代红利：从数月煎熬到一周单挑 对比当下的研发环境，Romain 与 Matias 一起回顾了过去没有 AI 辅助时的创业历程。六七年前，非技术背景的创业者如果想将构想付诸现实，道路极其痛苦且漫长：只能四处东拼西凑、复制粘贴代码硬憋出一个简陋的原型，随后四处奔走融资，拿到资金后再招募 3 到 4 名工程师团队耗费数月时间开发最小可行产品（MVP）。 Matias 感慨道，当年耗费重金、耗时数月、动用多名工程师才构建完成的初版 App，如果放在今天利用最新的 AI 编码工具，他自己一个人用不到一周就能完全交付。Romain 也分享了自己的切身体会：当年他创业时，需要动用多达 15 名工程师的高规格团队，耗时整整一年半，才终于打磨出能够正式交付给客户的 V1 版本。两人一致感慨，当下无疑是有史以来创办公司和构建产品最好的时代，任何有想法的人都已被赋予了即刻动手的全套军火。 --- ### [08:04](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=484s) 战胜“构建焦虑”：夜间自主研发与 Feature Flag 实验流 当话题转向业余项目的构建时，Romain 惊叹 Matias 如今几乎保持着“每天产出一个新项目”的恐怖节奏。然而，Matias 却吐露了背后的心路历程：当构建门槛被彻底抹平后，他一度陷入了严重的心理焦虑，哪怕只是周末外出放松都会产生强烈的负罪感，总觉得自己是在虚度光阴，没有充分利用眼前这项革命性技术不断去创造。 为了摆脱这种被代码绑架的生活状态，Matias 探索出了一套高度自治的“异步开发工作流”，核心目标是最大限度减少肉身守在电脑前的时间： - **新项目立项**：只需用简练语言向 Codex 描述创意构想，Codex 就会自主完成技术方案规划、代码编写与全面测试，全过程可能自主运行数小时，全部搞定后再通知他验收。 - **现有产品迭代与竞品调研**：给 Codex 设定专属 Skill，让其自主搜索分析竞品动态与市场新点子，并结合产品既有目标和个人偏好，推演应当新增哪些功能。 - **“睡前下发，醒来验收”**：在入睡前给 Codex 下发指令，要求它提炼出前 10 个最有价值的功能点，并以模块化架构和功能开关（Feature Flags）的形式全部直接实现并打包进主程序分支。 - **掌控权过滤**：第二天清晨醒来，应用中已经生成了 10 个可随时开闭的独立实验性功能开关。Matias 无需成为创意生成的瓶颈，同时又能保持绝对的人工裁决权，只需在界面上一一试用开关，决定哪些功能值得合并，哪些应该舍弃。 ### [10:08](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=608s) 抓住碎片时间的“Codex时刻”与并发项目管理 Matias 分享了自己如何在日常繁重的工作与生活之余推进个人副业项目。他善于捕捉一整天中的各种碎片化间隙——比如在公司午休、周末外出聚餐、喝酒前，或是晚上临睡前。在这些被称为“Codex时刻”的短暂窗口期，他会打开个人电脑，利用预先写好的各种 skills 和自动化流程，将一堆工作像任务单一样派发给 Codex，然后直接合上电脑去干别的事情。让 Codex 在自己毫不费心的后台状态下自主执行，这套投资策略带来的产出回报极其惊人。 当切换到屏幕演示时，Matias 展示了自己的 Linear 工作台，界面上赫然并存着 12 个同时处于“进行中”状态的项目。虽然他不可能在同一时刻兼顾所有项目，但这种极高密度的并发管理正是借助 Codex 的自主能力才得以维持的。 ```mermaid flowchart TD accTitle: Codex离线异步工作流 accDescr: 展示Matias利用碎片时间派发任务给Codex并离线执行的流程 subgraph Trigger["碎片时间触发（Codex时刻）"] A["午休 / 晚餐前 / 睡前"] --> B["语音口述想法 / 派发任务"] end subgraph Execution["后台自主执行"] B --> C["Codex 依据 Skill 与 agents.md 规划"] C --> D["在 Linear 中自动拆解 Issue 与里程碑"] D --> E["Codex 离线编写代码与执行构建"] end subgraph Review["空闲验收"] E --> F["用户上线：Feature Flag 灰度切换 / 测试验收"] end ``` ### [10:59](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=659s) 基于 Codex App Server 开发的 AI 写作助手 Matias 重点挑选了一个正在进行的项目作为案例：这是一款他最初为提升自己办公写作效率而开发的 Mac 应用，后来还移植为了 iOS 版本及键盘扩展（Keyboard Extension）。该应用在底层直接连接了 Codex App Server，使得推理请求能够直接走他个人的 ChatGPT 订阅额度，无需额外的 API 配置。 在日常交互中，Matias 彻底摒弃了键盘打字，而是完全依赖语音输入。他调侃道，一年前在办公室戴着那种通话耳麦对着电脑自言自语时，同事们经常笑话他看起来就像个“呼叫中心话务员”（call center operator）；不过时过境迁，如今办公室里到处都是戴着同款耳机对着 AI 说话的同事，大家已经习以为常。 在演示现场，Matias 按下全局快捷键 `Command + Shift + Space` 唤出极简弹窗，随口录入了一句极其口语化且并不通顺的话：“Hey, I'm not totally sure I understand what you're saying. Maybe we should just get together and figure it out”（嘿，我不太确定是否完全理解了你的意思，要不我们碰一下一起理清楚）。随后他按下快捷键 `Command + Return`，系统瞬间借助刚刚发布的 GPT-5.5 模型，将其重写为一段专业、得体、消除了错别字的职场沟通文本。 - **全局唤起**：通过 `Command + Shift + Space` 调出轻量化浮窗，`Command + Return` 一键触发改写。 - **底层架构**：基于 Codex App Server，直接复用现有的 ChatGPT / Codex 订阅做模型推理。 - **模型支持**：内置模型选择器（Model Picker），在 GPT-5.5 发布后第一时间无缝接入。 - **跨平台延伸**：除 macOS 桌面客户端外，还延伸出了带键盘扩展的 iOS 手机应用。 ### [12:55](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=775s) 架构与规划委派：agents.md 与 Linear 的自动化闭环 在谈到如何彻底解放双手时，Matias 展示了一个令人惊叹的事实：在当前 Linear 看板中，包含已完成的 159 个 Issue、正在代码审查的 Issue 以及若干进行中的任务，这多达上百个任务条目中，“没有一条是我亲手创建或编写的，全部由 Codex 自动生成”。 Matias 总结了与大语言模型协作的核心认知：每当 LLM 交付的结果出乎意料或者令人失望时，根本原因通常是模型在面对模糊信息时不得不自行做出各种假设，而这些假设恰好与你的预期相违背。为了彻底消除这种隐性偏差，他的工作流高度依赖 upfront（前置明确化）流程与核心配置文件： ```mermaid flowchart LR accTitle: 基于 agents.md 的全自动项目规划与执行闭环 accDescr: 从初期构想到Linear看板全自动任务创建与执行的流转示意 subgraph Input["标准化上下文输入"] A["初始产品创意"] B["agents.md<br/>（编码风格/工作习惯/架构偏好）"] end subgraph Planning["规划阶段（Codex Skill）"] A & B --> C["Codex 规划 Skill"] C --> D["拆解多阶段里程碑 (Milestones)"] end subgraph Delivery["执行阶段（Linear 闭环）"] D --> E["自动批量生成 150+ Linear Issues"] E --> F["Codex 自动接单开发并关单 (In Review/Done)"] end ``` - **`agents.md` 单文件沉淀**：把个人的工作哲学、代码规范、架构风格与决策偏好集中整理在一个 `agents.md` 文件中。启动任何全新项目时，只需注入该文件即可完成初始化。 - **规划 Skill 自动化**：通过定制的规划 Skill 与 Codex 展开对话，Codex 会自主完成架构设计，将产品愿景拆解为底层基础架构（Foundation）、偏好设置（Preferences）及各功能模块。 - **项目管理全权委托**：Codex 在 Linear 中自动建卡、细化里程碑、按顺序领取任务编码并推进状态变更，实现“项目经理 + 核心研发”两项角色的双重委派。 ### [15:06](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=906s) 竞品调研与夜间 Feature Flag 自动化实验生成 在完成了产品的 V1 版本后，Matias 面临着下一个瓶颈：如何持续为产品迭代新功能，同时避免让自己成为创意的堵点（bottleneck）。他既不希望所有点子都得靠自己大脑憋出来，又希望在产品方向上保留最终的把关权。为此，他设计了一套“全自主竞争调研与实验功能孵化”的 Codex Skill。 在临睡前，Matias 只需要向 Codex 下达一条宏观指令：“我正在给自己做一个 AI 写作助手。你去网上调查一下市面上还有哪些同类竞品、它们都在做哪些功能；给最有趣、最具影响力的功能打分；把它们全部构建出来，但必须采用模块化架构包裹在 Feature Flag（特性开关）后，确保我明天醒来能自由开关测试。” Codex 在调研方面展现出了远超人类的耐力，它永远不会对枯燥的网页检索感到厌烦。凭借预设流程，Codex 一口气跑完了竞品抓取、价值排序、功能代码编写与特性开关配置的全套链路。 - **睡梦中构建（Built overnight）**：整套长列表的实验功能全部在夜间离线一次性编译部署完毕。 - **即插即用测试**：醒来打开 Mac 客户端的设置面板，所有新功能以开关形式并列排布；勾选“输出语言翻译”或“Slack 风格输出排版”等功能后，主界面瞬间实时生效。 - **极佳的筛选体验**：通过 Feature Flag 机制，Matias 可以在数秒内亲自体验功能效果，并迅速决定哪些特性保留至主干产品、哪些直接淘汰，既省时又保证了产品质量。 ### [18:51](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1131s) 探索另类入口：OpenClaw、Discord 与随身编码实验 随着对 Codex App Server 潜力的挖掘，Matias 的探索进一步扩展到了更多非传统的编程交互界面。为了验证能否从任何碎片化设备随时启动一次编程任务，他不仅折腾了移动端应用，甚至还写了一款 Apple Watch 手表应用。 其中一个颇具趣味的核心玩具是基于 OpenClaw 构建的常驻 AI 助理，名叫“Lou”。屏幕上展示出了 Lou 的幽默头像，引来两人一阵欢笑。Matias 介绍说，Lou 运行在家中一台专属的独立主机上，平时主要承担两项职责： 1. **群聊整活**：在朋友的社交群组里专门接梗、插科打诨。虽然没有什么实际生产力，但充满了乐趣。 2. **移动端代码代理**：深度打通了 OpenClaw、Codex 与 Discord。Matias 为不同的代码仓库各自绑定了专属的 Discord 频道。当他出门在外、手头只有一部手机时，只需在对应的 Discord 频道里给 Lou 留下一条指令，Lou 就能通过 Codex App Server 在远端工作机上自动拉取仓库、执行开发任务，真正实现了不受设备和场景限制的编码自由。 ### [20:15](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1215s) 腕上编程：通过 Apple Watch 与 Codex App Server 触发代码任务 Matias Castello 展示了他最近开发的一个极具未来感的侧边项目：在 Apple Watch 表盘上开发一个复杂功能（Complication）小组件，结合配套的 iPhone 应用，实现通过随时随地录制语音备忘录直接触发 Codex 任务。想象一下在路上散步或浏览网页时突然发现某个网站有拼写错误，无需掏出电脑，只需对着手表面表说一句：“嘿，帮我修复这个网站落地页上的错字。”系统便会自动完成后续流程。 Matias 现场演示了这套工作流：他对着 Apple Watch 发送了一条简短语音——“在我的 skills 代码仓库中创建一个 test.md 文件”。该录音立刻推送到配套的 iPhone 应用，应用在本地转写语音并解析出开发者的真实意图。由于应用拥有 Matias GitHub 账户的完整上下文，它准确地将“skills repo”路由定位到他个人的 GitHub 仓库，并通过后端启动任务执行。 在架构设计上，应用设置中提供了两种运行机制：一种是常规的直接配置 OpenAI API Key 运行；另一种被称为“后端会话”（Backend Session），即直接连接运行在其他服务器上的 Codex App Server。这种方式直接绑定 Matias 的个人账户并加载全部上下文，让他可以真正实现移动办公。尽管这种方式更适合单次搞定的小型离散任务（one-shot tasks），但它让开发者离“将所有实现细节彻底抽象化”的目标又近了一步。 ```mermaid flowchart TD accTitle: Apple Watch 驱动 Codex 自动化架构 accDescr: 展示从 Apple Watch 语音输入到 Codex App Server 执行并在 GitHub 创建 Pull Request 的端到端流程 subgraph Client["客户端输入层"] A["Apple Watch 录制语音备忘录"] --> B["iPhone 配套应用转写文本"] B --> C["意图解析与上下文路由"] end subgraph Server["后端与执行层"] C --> D["Codex App Server 会话"] D --> E["GitHub 个人仓库 (skills repo)"] E --> F["自动提交变更 / 创建 PR"] end ``` ### [22:52](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1372s) 开放生态：Codex CLI、Harness 与 App Server 的开源价值 几分钟后，Matias 的手机立刻收到了通知，GitHub 仓库中已经准确无误地提上了新建 `test.md` 的 Pull Request。虽然这只是一个极简测试，但它直观证明了整套链路的稳定性。Matias 开玩笑地敦促 OpenAI 团队：“我觉得你们应该亲自把这个功能做进官方产品里，把 Apple Watch 表盘组件和 iPhone 桌面小组件都做出来，这样我就不用自己写了。” Romain Huet 对此深表赞同，并指出这正是 OpenAI 决定全面开源 Codex CLI、Codex Harness 以及 Codex App Server 的初衷。OpenAI 希望构建一个开放的生态系统，使像 Matias 这样的创造者即便在官方尚未正式发布某些端到端功能前，也能基于开放接口按自己的设想自由扩展编码智能体（Coding Agent）。无论是把智能体集成进手表、手机还是 Discord 机器人中，开发者都能以自己喜欢的方式调度 AI。 - **多端触达**：利用开放协议将 Codex 接入即时通讯（如 Discord 机器人 Lou）或可穿戴设备（Apple Watch）。 - **极速交付**：从对着手表说话到 GitHub 收到提交代码的 Pull Request，全程仅需数分钟。 - **生态定位**：Codex CLI 和 Harness 提供核心运行载体，App Server 则充当远程代理桥梁。 ### [24:01](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1441s) 惊艳瞬间：Computer Use 与 GPT-5.5 带来的能力跃迁 聊到近期 OpenAI 发布的一系列新工具与模型时，Romain 询问 Matias 是否在 Codex 与最新模型 GPT-5.5 的配合下遇到过令人震撼的时刻。Matias 坦言最近“电脑使用”（Computer Use）功能的表现远超他的预期。上周末他在家里折腾一块连接在电视机上的树莓派（Raspberry Pi），面临一项极其枯燥的机械任务：需要将大量域名 URL 逐一复制粘贴到某个管理后台中。 Matias 没有手动操作，而是直接指示 Codex 通过浏览器 SSH 登录到那台树莓派并接管操作。Codex 自动找到了包含 URL 清单的文件，打开后台页面，行云流水地完成了一系列复制粘贴操作。最搞笑的细节在于，Matias 本想利用 Computer Use 节省时间，结果却目不转睛地盯着屏幕看完了全过程——“看到光标在屏幕上平滑移动、准确打开窗口并填入表单，那套动画效果实在太迷人了，你既想插手接管，又忍不住想静静看它表演。” - **自动化运维**：直接通过浏览器建立 SSH 连接接管嵌入式设备（树莓派）。 - **人机共存体验**：任务既能在前台以拟人化光标动画执行，也能静默转入后台无感运行。 ### [25:22](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1522s) 个人评测基准：十年老项目 Snapcat 与多模态 UI 生成 谈及 GPT-5.5，Matias 提到了自己维持了十多年的个人评测基准（Eval）项目——Snapcat。这是一个十年前他在黑客马拉松上开发的趣味应用，专门给猫咪拍自拍：屏幕呈现全黑背景，上面有一颗红色的激光红点在随机移动；当猫咪扑抓红点触碰到屏幕视口时，前置摄像头就会立刻抓拍一张照片；而猫咪按不到的机身音量键则是人类退出“猫咪模式”的隐藏开关，退出后便能查看猫咪视角的自拍相册。 在十多年前的黑客松上，5 个人组成的团队花费了整整 24 小时以上才勉强完成这个项目。而如今借助 GPT-5.5，只需一段精心设计的 Prompt 和几项预置技能，AI 在一轮交互（One-shot）中就能独立交付完整的应用骨架。甚至连现场 Romain 假装成猫咪点击屏幕测试时，应用也顺畅完成了拍照与图库联动。 更具革命性的是 Matias 构建应用界面的新方式：他先用自然语言描述出“轻快、色彩丰富、充满趣味”的设计风格，通过图像生成模型直接生成 UI 视觉稿；随后将设计图丢给 Codex，命令其依照图片还原前端代码；在完成相册视图后，他又要求 AI“参考当前风格重新设计并实现设置页和主屏幕”。这种“文本描述生成视觉参考图，再由编码智能体将其转写为代码”的多模态工作流，颠覆了传统 UI 开发的范式。 | 维度 | 十年前黑客松阶段 | 现阶段（GPT-5.5 + Codex） | |:--- |:--- |:--- | | **团队与耗时** | 5 人团队，耗时超过 24 小时 | 单人完成，前一天晚上一次性生成（One-shot） | | **UI 开发方式** | 手动切图、手写界面布局代码 | 图像生成工具产出设计稿，Codex 直接还原实现 | | **代码与体验** | 勉强运行的原型 | 细节精致、体验丝滑、视觉统一的完整应用 | ### [28:24](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1704s) 给创造者的三大建议：打破心智枷锁 对谈最后，Romain 请 Matias 为那些渴望像他一样精通 Codex 的创业者与开发者分享建议。Matias 总结了从自身经历中沉淀出的三个核心心智假设（Mental Assumptions）： 首先，“假设它是可能的”（Assume it's possible）。很多时候开发者在脑海中闪过一个创意时容易打退堂鼓，但现实往往是：只要你想得到，现在的技术能力基本都能实现。 其次，“假设你能够做到”（Assume you can do it）。许多人迟迟不愿动手构建项目，本质卡点是心理上认定自己缺乏能力。Matias 建议即便带着一种“盲目自我催眠”（be in denial）的自信，也要坚信自己能行，而事实证明这种假设通常是对的。 最后，“当大模型未能交付理想结果时，放下自尊，假设是自己的沟通问题”（Put your ego aside, assume it's your fault）。这是最难做到的一点：不要轻易断定工具能力不足，而应反思自己是否尚未找到清晰传达意图的方式。不断调整 Prompt、更换表达并反复尝试，正是把想法最终变成现实的关键。正如 Romain 最后总结的，尽管听起来有些老生常谈，但事实确实如此：“任何人都可以亲手造出任何东西（Anyone can just build anything）。” ## Notable quotes > [00:58](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=58s) 在 Slack 里直接使唤模型改文档始终是推动全员拥抱 AI 最低门槛、最立竿见影的切入点。 — Romain Huet > [10:59](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=659s) Hey, I'm not totally sure I understand what you're saying. Maybe we should just get together and figure it out — Matias Castello > [12:55](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=775s) 没有一条是我亲手创建或编写的，全部由 Codex 自动生成。 — Matias Castello > [15:06](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=906s) 我正在给自己做一个 AI 写作助手。你去网上调查一下市面上还有哪些同类竞品、它们都在做哪些功能；给最有趣、最具影响力的功能打分；把它们全部构建出来，但必须采用模块化架构包裹在 Feature Flag（特性开关）后，确保我明天醒来能自由开关测试。 — Matias Castello > [20:15](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1215s) 在我的 skills 代码仓库中创建一个 test.md 文件 — Matias Castello > [22:52](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1372s) 我觉得你们应该亲自把这个功能做进官方产品里，把 Apple Watch 表盘组件和 iPhone 桌面小组件都做出来，这样我就不用自己写了。 — Matias Castello > [24:01](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1441s) 看到光标在屏幕上平滑移动、准确打开窗口并填入表单，那套动画效果实在太迷人了，你既想插手接管，又忍不住想静静看它表演。 — Matias Castello > [28:24](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1704s) 假设它是可能的（Assume it's possible）。 — Matias Castello > [28:24](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1704s) 假设你能够做到（Assume you can do it）。 — Matias Castello > [28:24](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1704s) 当大模型未能交付理想结果时，放下自尊，假设是自己的沟通问题（Put your ego aside, assume it's your fault）。 — Matias Castello > [28:24](https://www.youtube.com/watch?v=8QKqENa_eQQ&t=1704s) 任何人都可以亲手造出任何东西（Anyone can just build anything）。 — Romain Huet ## Mentioned Matias Castello, Romain Huet, Alchemy, OpenAI, Codex, Codex App Server, Codex CLI, Codex Harness, GPT-5.5, Snapcat, Datadog, Slack, Linear, GitHub, Discord, OpenClaw, Apple Watch, iPhone, Mac, Raspberry Pi, Facebook ## Video description Builders Unscripted spotlights the stories behind real projects and the mindset that makes them possible: you can just build things. In this episode, Romain Huet, Head of Developer Experience, chats with Matias Castello, Product Lead at Alchemy. Matias shares how his team uses Codex for code review and product work, how he builds side projects with Codex App Server, and why he keeps rebuilding Snapcat as a personal eval for each new model. Chapters: 00:00 Intro 00:58 Bringing Codex to work at Alchemy 01:51 Using Codex for code review 08:04 Building side projects with Codex 18:51 Codex App Server projects 24:01 Computer use, GPT-5.5, and Snapcat 28:24 Advice for builders [Source](https://www.youtube.com/watch?v=8QKqENa_eQQ)

[0:58] Romain Huet: 在 Slack 里直接使唤模型改文档始终是推动全员拥抱 AI 最低门槛、最立竿见影的切入点。 [10:59] Matias Castello: Hey, I'm not totally sure I understand what you're saying. Maybe we should just get together and figure it out [12:55] Matias Castello: 没有一条是我亲手创建或编写的，全部由 Codex 自动生成。 [15:06] Matias Castello: 我正在给自己做一个 AI 写作助手。你去网上调查一下市面上还有哪些同类竞品、它们都在做哪些功能；给最有趣、最具影响力的功能打分；把它们全部构建出来，但必须采用模块化架构包裹在 Feature Flag（特性开关）后，确保我明天醒来能自由开关测试。 [20:15] Matias Castello: 在我的 skills 代码仓库中创建一个 test.md 文件 [22:52] Matias Castello: 我觉得你们应该亲自把这个功能做进官方产品里，把 Apple Watch 表盘组件和 iPhone 桌面小组件都做出来，这样我就不用自己写了。 [24:01] Matias Castello: 看到光标在屏幕上平滑移动、准确打开窗口并填入表单，那套动画效果实在太迷人了，你既想插手接管，又忍不住想静静看它表演。 [28:24] Matias Castello: 假设它是可能的（Assume it's possible）。 [28:24] Matias Castello: 假设你能够做到（Assume you can do it）。 [28:24] Matias Castello: 当大模型未能交付理想结果时，放下自尊，假设是自己的沟通问题（Put your ego aside, assume it's your fault）。 [28:24] Romain Huet: 任何人都可以亲手造出任何东西（Anyone can just build anything）。

[6:29] Romain Huet: It took me a team of like 15 engineers for maybe a year and a half to get to like the first milestone of a V1 that we shipped to many customers. [28:24] Matias Castello: When you try to get something out of an LLM and it doesn't work... you have to put your ego aside, assume it's your fault. Don't assume that the tool isn't capable. Assume you haven't found a way to communicate it or convey what you want just yet.

[0:13] (speaking french) [0:15] (speaking french) [0:17] Matias, thank you for joining. [0:19] Thank you for having me. Really happy to be here. [0:21] I'm really excited to chat because you have a very unique background. [0:25] You've crossed a bunch of worlds. [0:27] You work in consumer products. [0:29] You worked on developer platforms like at Facebook in the early days [0:32] and now leading product for Alchemy in the crypto space. [0:35] But what I find even more interesting is that you're not an engineer by training, [0:40] and yet you're building more than many engineers that I know. [0:44] How do you go about it? [0:46] I think it just comes down to being interested in a lot of things [0:50] and trying a lot of things over time [0:53] and getting better and better as models, I guess, get better as well. [0:57] That's awesome. [0:58] So today you're using AI a ton at work and outside of work. [1:03] But let's start with, like, Alchemy. [1:06] Like what was, aha moment you had maybe kind of bringing AI to work? [1:11] The first thing that I remember the first time we used AI, [1:14] I guess it was Codex in Slack. [1:16] Right. [1:17] It's probably a year ago or so at this point. [1:19] Yeah. [1:20] The first thing we used it for was to make small edits [1:23] to our documentation, our developer docs from Slack. [1:26] So we went from having to run the site locally and doing that whole process [1:30] that is fairly involved. [1:32] To just being able to add Codex on Slack in the docs channel in our company Slack [1:36] and to be able to just tell it to make changes. [1:38] We do that all the time at OpenAI too. [1:41] We still do it. [1:41] Now we have a bit more of a sophisticated setup [1:43] but that was really the first thing that we used [1:47] Codex for and AI in general at the company. [1:50] From there, the other big moment I remember [1:53] that was probably the bigger turning point, was using code review. [1:56] And so there was this instance, we had a small incident [2:00] that was related to a big migration that had happened months prior. [2:03] And in big migrations, some bugs will happen. [2:06] It was like a lot of code. [2:08] We had an internal postmortem about the incident [2:12] and we basically identify what the race condition was. [2:15] We fixed it and someone on the team had the idea to retroactively run code review, Codex, [2:21] to see if Codex would have caught the bug. [2:24] And it did. [2:25] Wow. [2:25] And so we did that a couple of times just to see. [2:28] People were starting to get curious and [2:30] you could tell that was a bit of a turning point in terms of model capability [2:34] and like a lot of the ingredients were right for for teams to start adopting those tools more. [2:38] From there, I remember a couple of days after seeing one of the engineers on the team [2:43] essentially using Codex review in pull request comments as his teammate. [2:49] He was just he submitted a PR that I saw go through Slack. [2:52] I click through it and I saw him do at Codex review, [2:55] address the comments that Codex would come up with, then review again. [3:00] Do that again and he was just doing this back and forth with Codex. [3:03] You could tell that people were getting over the hurdle of, [3:08] of like thinking that LLMs weren't good enough to help in a professional setting [3:13] with a really large complex code base, [3:15] with code that needs to be good enough to be served to a lot of people. [3:19] That’s really cool. [3:19] That was really a turning point. [3:20] Yeah, that's a very interesting point on code review. [3:22] I remember talking to many companies were just getting into like, you know, AI coding. [3:28] Code review was the first moment when they realized [3:31] that many incidents they had could be caught actually by Codex automatically. [3:35] And they replayed some of those similar to Alchemy. [3:38] I remember Datadog, for instance, back in January, said that [3:41] like more than one incident out of five would have been saved by Codex. [3:47] But I'm sure the model capabilities now are so good [3:49] with GPT 5.5, it could be more than half, easily. [3:52] Nine out of ten, maybe. [3:54] Maybe most of it. [3:55] Yeah. So code review was really a turning point for many companies [3:59] to get into Codex and really realize the power of AI. [4:04] Can you tell me a bit more about how you're using Codex at work today? [4:08] I use it a lot for a lot of different things. [4:11] A lot of those typical product management sort of tasks, writing documents. [4:16] We've created all these skills internally to help us [4:19] do the PM job more easily and better and faster. [4:23] That includes writing PRDs, analyzing customer feedback, like doing all those things. [4:29] I use Codex for that [4:30] and several members of the team reuse those skills as well. [4:33] We have this shared, repo of skills across the company, [4:37] across all the different functions, so that not only we can do the same [4:41] job and the same tasks faster and better [4:44] but also more people who are not necessarily PMs are able to do those things. [4:48] I'm sure, yeah. [4:49] And I think it's also telling a lot about, like, how we think about developers now, [4:54] not just as humans but also as agents consuming our own platforms and infrastructure. [4:59] We have to think about these agents and making sure that, like, tools like Codex [5:03] can actually integrate, like OpenAI’s API or Alchemy’s infrastructure very quickly. [5:07] Yeah, it's really several things that change at the same time. [5:10] First, the way we work has changed and Alchemy has a lot of very technical products. [5:16] It's an infrastructure company first. [5:19] And so all the developers internally have changed the way they work. [5:22] And so it's that's very clear to everyone now. [5:25] But that also means as builders of a developer platform, we have to change [5:30] the way we build tools so that other developers that we know [5:34] are also changing their workflows with AI are able to use Alchemy better. [5:38] So it's, the developer we're building for has evolved. [5:42] Now, we are very much assuming that 100% of developers are building [5:46] software with the help of AI. [5:48] Right. [5:48] And so it's evolving your platform to fit that new reality as one thing. [5:53] And then the other thing that's a bit more nascent [5:55] but that's also very interesting is, I guess, [5:58] this new type of developer that's an agent sometimes, autonomous agent. [6:01] And they don't always need the same tooling. [6:03] They need different things. [6:04] And if an autonomous agent shows up on Alchemy [6:07] and as far as implementation decisions go, they're fully autonomous. [6:11] They have to be able to sign up and integrate something [6:15] and use the blockchain to perform some kind of tasks. [6:18] We need to build the tools for that. [6:20] And it's building for those two. [6:21] And over time those two probably converge. [6:23] But for now they're still fairly different needs. [6:25] And my job is basically to try to figure out what we do for both. [6:28] Awesome. [6:29] Zooming out of Alchemy a little bit, I was reflecting recently [6:33] about my own startup. [6:34] And you were a founder before, as well. [6:37] How do you think now, about this time where you had to build without AI? [6:42] It was very different. [6:43] And it was hard, even if you were not an engineer. [6:45] Even more difficult, like the typical way you'd go about it 6 or 7 years ago [6:50] was to build some kind of prototype, to try to raise enough money [6:54] to hire a small team, to then build the real thing. [6:56] And that that's pretty much what I did with with my company at the time, [6:59] built the first prototype myself. [7:01] It was very difficult because it was essentially copy pasting code left and right until something worked. [7:07] Managed to raise a bit of money, hired a team of 3 or 4 engineers [7:10] to help build the first MVP and not only that took, [7:15] a decent amount of money and several people, but also took months. [7:18] I mean I haven't thought about this much [7:19] but I wonder how long it would take me to build the first version of the app [7:22] we’re a building at the time with AI today, just by myself. [7:26] My guess is that it would take less than a week. [7:29] I don't know what your experience was like. [7:30] You are building a much more complicated product than mine, I guess. [7:34] How long did I take it back then and how many people? [7:37] It took me a team of like 15 engineers for maybe a year and a half [7:42] to get to like the first like milestone of a V1 that we shipped [7:46] to like many customers. [7:48] It's pretty incredible how far we've come. [7:50] Like it's just never been easier to do something. [7:53] Yeah. [7:54] And to start something from scratch. [7:55] Yeah. [7:56] It's probably the best time to be a founder. [7:58] Like any idea you have, you can just build it. [8:00] Yeah. [8:00] Anyone who has an idea now has the tools to basically try it out. [8:04] Speaking of things that you do on the side, it's a great segue [8:07] because I wanted to ask you about how you're using AI outside of work. [8:11] It sounds like you're building so many projects these days [8:14] it's almost like one a day. [8:16] You know, it's amazing because it's never been easier to build things. [8:19] And yet I just kept feeling this sort of anxiety [8:22] that I, I wasn't doing enough pretty much all the time. [8:25] I think I'm a bit past that. [8:27] But the reason I am is because I try to work [8:30] on a setup that allows me to spend little time on my computer [8:33] and have essentially Codex work for me for hours while I'm doing something else. [8:37] It's like [8:39] I had gotten to a point a few months ago where I felt bad about going out [8:43] because I felt like I should be taking advantage of this [8:46] amazing technology we have and I should just be building things all the time. [8:49] I didn't want to do anything else. [8:50] I didn't want to feel like that either. [8:51] I didn't like that trade off. [8:53] And so I invested in a set of skills and a process and several things [8:58] that I set up for myself to basically be able to very quickly [9:01] either describe a new idea and have Codex come up [9:04] with a plan, implement it, test it, and let me know when it's done. [9:07] And it's very easy to to get Codex to work for hours at that point. [9:11] Or if I have an existing product and I want to explore new things, [9:14] I came up with this way of essentially having Codex [9:17] do research for me on what new features I should build. [9:20] It will look at other competitors. [9:22] It will look at other ideas. [9:24] It will try to figure out what my goals are [9:25] and what my own preferences are. [9:28] And I created this set of skills that basically gets Codex [9:31] to build all those things. [9:33] So it comes up with ideas. [9:34] It builds them in a modular way. [9:37] So it's going to feature flag them and build them [9:38] as experiments in the main app so that I can prompt Codex before going to bed [9:43] to go do some research about an app that I have. [9:45] Tell it to build the top ten features it can come up with, [9:48] build them as experiments in the product. [9:50] Wake up to ten feature flags that I can toggle on and off [9:54] automatically to basically [9:57] not be the bottleneck when it comes to generating new ideas [10:00] but still having the control of deciding what is good enough to stay in the project or not. [10:04] I love that. That's really cool. [10:05] It's like without spending a ton of time on a computer [10:08] have these moments, like during my lunch break at work, [10:11] I have my personal laptop and I try to send things to Codex [10:14] so that it can work while I go back to work [10:16] before going out on weekends to dinner, or to have drinks before going to bed. [10:20] I have these Codex moments, I guess, where I just try to dispatch a bunch of work, [10:25] use all these processes and these workflows I have with skills [10:28] and other actions so that Codex can work while I'm not paying attention. [10:33] And it's amazing how much you can do if you kind of invest in it that way. [10:36] That's so cool. [10:37] I mean, honestly, I would love to see your setup and how you're using the Codex app [10:41] and kind of bringing all of these tools together. [10:43] Yeah. Let's check it out. [10:44] Let's check out linear to start. [10:46] So what you see here is basically a collection of projects. [10:50] You see there's this 12 of them in progress. At the same time. [10:53] I'm not working on all of them [10:55] in one go, but maybe we can pick one and I can show you how this works. [10:59] This one, for example is a mac app that since then [11:02] I've also built it as an iOS app that uses Codex App server [11:06] to essentially help me write better while I'm at my computer. [11:09] It's something that I built for myself [11:10] for work first, and that, I kept building out since. [11:14] The way it works is that you have a global shortcut. [11:16] In my case, it's like command shift spacebar, [11:19] and it brings up this little window that you see there. [11:21] That's cool. [11:22] And this is using Codex App Server. [11:24] So in the settings it's using my GPT subscription behind the scenes. [11:28] And if you were to write something in my case, I'm not even going to write. [11:33] I usually talk to my computer just because it's so much faster for everything. [11:36] Like people at work a year ago would make fun of my little headset. [11:40] I basically looked like a call center. [11:42] Call center operator. [11:44] And now a lot of people have those similar headsets, so it's okay. [11:47] But if you say you're on slack and you want to write something, [11:52] Yeah. [11:52] but you need to clean it up [11:53] and you want AI to make it sound more professional or fix typos or things like that [11:57] like I'm going to record something that would be like, [11:59] “Hey, I'm not totally sure I understand what you're saying. [12:01] Maybe we should just get together and figure it out” or something like that. [12:06] Obviously, this is not a great sentence. [12:08] I just dictated that to my computer. [12:10] And then if you hit command return in this case, [12:13] it's going to rewrite it for me in a way that sounds more professional [12:17] because the professional mode was selected. [12:20] This is basically using Codex Server behind the scenes and GPT 5.5. [12:23] You have a model picker. [12:25] I just added 5.5 since it just came out. [12:28] And this is just an example of a little app that I made for myself for Mac. [12:32] I put the same setup in an iOS app after to have [12:37] like a keyboard extension and help you rewrite things quickly and so on. [12:40] I don't know if a lot of people use their, [12:42] their Codex subscription for things that are not just coding. [12:44] I know that people use the Codex app server to build other [12:47] coding interfaces, but in this case I basically used it [12:50] because I wanted the inference to be done, through my ChatGPT account. [12:55] And if we go back to to the process and how to find time, [12:59] what you see here is this linear project. [13:02] I didn't create this project. [13:04] There are like 159 issues that are done, a bunch that are in review [13:08] and some that are in progress. [13:10] I did not create or write a single one of these. [13:12] Wow. [13:13] Codex did. [13:13] What I did is to talk to Codex about what I wanted [13:18] through these skills that I wrote for myself, to basically help [13:23] me create a plan that turns into milestones [13:27] that you see here, like the foundation and the preferences and so on, a bunch of the features. [13:31] Codex will create that plan, [13:33] break it down into milestones, create the tasks, [13:36] work on them in a way that makes sense based on this agent’s .md file that I created, [13:42] that essentially summarizes the way I like to work. [13:44] And I put in a bunch of time and, putting that in a single file. [13:48] So when I create a new project or I have a new idea, I can basically just [13:53] initialize that project with that agent’s .md file. [13:56] From there, I use this skill I have to create a plan [14:00] and after that it's just telling Codex, build the plan. [14:04] That's really cool because you're using linear as the interface [14:08] as like your backlog and your release milestones and things like this. [14:11] But like Codex is not just writing [14:15] all the issues, it's also picking them up to close them and work on them. [14:18] Yeah. [14:19] It's of both delegating the, I guess, the project management [14:22] and the setting up of the project and the milestones and the tasks as well as the [14:27] the execution of the tasks too. [14:29] I’m initiating the project, sharing the initial idea, [14:34] giving Codex enough information about my preferences [14:36] and what I want the product to be so that it builds [14:39] the version of the product that I want. [14:41] I think the biggest gap when you code with LLMs, when [14:45] if the LLM produces something that's surprising to you, in a negative way, [14:49] it didn't do what you wanted, usually it's because it had to make one or [14:53] multiple assumptions and it just didn't make them in the same way you would have. [14:57] And so in order to avoid that, I have this process [15:00] that requires me to clarify all those things upfront. [15:03] And once Codex has all of that, it can just go and build. [15:06] That's really cool. [15:07] And so that's the first part and if you go back to [15:10] to this writing app as an example, the next step after [15:14] setting up a new project, building a V1 of a product [15:18] and having Codex essentially do all the work for me, usually while I'm not even on my computer. [15:23] I wanted to find a way to generate [15:26] new features and keep working on a [15:29] on an existing product without being the bottleneck. [15:32] I didn't want all the ideas to have to come from me [15:35] but I still wanted some level of control [15:38] when it comes to deciding whether or not I like that feature. [15:41] Right. [15:41] So I came up with this process that is basically a way [15:44] to build experiments with Codex on an existing project. [15:47] And it's a skill, again, that [15:51] essentially has me tell Codex I'm interested [15:54] in building features in this direction, or this is the high level goal that I have. [15:58] Go do some research about what types of features you could build. [16:02] Go figure out what the competitors to my products are building, things like that. [16:06] And it will come up with a lot of features based on what I'm saying. [16:10] And it's going to set them up and build them with feature flags [16:15] as experiments in the code base. [16:17] And what that means is that you're able to do things like what you see here. [16:22] You have [16:23] You can flip any switch of any feature. [16:25] That's right. [16:26] So you have a long list of of experimental feature flags. [16:29] These were all built in one go. [16:31] One evening while I was sleeping. [16:33] So before going to bed I said, [16:36] “Hey Codex, I'm building this AI writing assistant for myself. [16:39] Go figure out who else is building that. [16:41] What kind of features do they have? [16:43] Score the ones that are most interesting, most impactful, something like that. [16:48] Build all of them but build them in a modular way so that when I wake up, [16:52] I can toggle each one on or off independently to see if I like them. [16:57] And so you basically see a list of all of those there. [17:01] We can pick a few if you want to see. [17:02] Like we're going to output language is basically going to translate [17:05] Sure. [17:06] just going to [17:06] It’s fascinating that you did not use Codex just to make the plan and implement [17:10] but also use Codex [17:12] even before to make the research online about what the feature should even be. [17:16] Yeah. [17:16] And so Codex is great. [17:18] It's it's way better than me at doing research. [17:21] It doesn't it doesn't get bored of searching for things. [17:24] And so as long as you tell it what to research [17:27] and how you expect the findings to be structured [17:30] and in my case, you also tell me what to do with those findings. [17:33] It's basically just a sequence of fairly small steps. [17:36] but it will do all of them in one go, because it's all encoded in a skill or a set of skills. [17:41] And so it's able to come up with all of this build all of it, [17:44] build in a way that makes it easy for me to test [17:47] and makes it easy for me to decide whether or not I want it to stay in the product. [17:51] And so we selected a few here. [17:53] If we go back to the rewriting. [17:56] They just showed up instantly. [17:57] They just show up there. [17:58] And so in this case you see that you have a translation feature. [18:02] In this case you can change the [18:03] the output format of the message as an example. [18:06] It's like Slack style. [18:07] Yeah. [18:07] And so those are all small features that I got Codex to ideate [18:13] and build in a way that, [18:15] makes it really easy for me to essentially decide [18:18] if I want that to stay in the product or not. [18:19] So, those are all the things that I ended up coming up with to try not to spend [18:24] too much time building in a way but still end up building quite a few things. [18:28] It's really cool. [18:29] Instead of starting from scratch, it's basically [18:33] trying to think about the things you do repeatedly, or even trying to think [18:37] about the ways you like to do things, putting that into a skills, [18:41] putting some of that in an agent’s .md file. [18:43] Sometimes it's additional tooling and other things that make you go even beyond that. [18:47] And then having AI do most of the work. [18:51] That's awesome. [18:52] Have you built any other like projects on top of the Codex Observer? [18:55] Maybe more like on the coding side? [18:57] I built a few things like, this is an AI writing assistant. [19:01] It's a mac app. It's a consumer. [19:04] I use OpenClaw and I've built a couple of other things on my phone, even an Apple Watch app. [19:10] That are [19:11] It was just me tinkering, trying to see if I could have [19:16] different or new entry points to start a coding job, essentially. [19:20] So, in OpenClaw, for example, I set up my OpenClaw that my assistant's [19:24] name is Lou has got a pretty funny profile picture if you want to see it. [19:28] Yeah. [19:29] It’s [19:31] look at this guy. [19:33] [Laughs] [19:34] So what do you do with your OpenClaw here? [19:36] What do you talk about? [19:38] This little guy, Lou, it basically, it mostly has two jobs. [19:42] The first one is to make jokes in group chats with friends. [19:46] It's not very useful but it's quite funny. [19:48] The second one is to just help me code. [19:50] And so it's connected at home. [19:52] It's running on a dedicated machine. [19:54] It's got access to Codex and it's got integrations [19:58] that make it easier for me to do certain things on the go. [20:00] For example, I have an OpenClaw discord and Codex integration, [20:04] so that I have a bunch of discord channels. [20:07] Each one is bound to a specific repo. [20:09] And so if I want to do something from my phone on the go and I wanted to [20:15] to be able to run at home and do things that maybe aren't so easy [20:18] to do yet on the cloud or things like that. [20:19] I can just do it while I'm walking around somewhere. [20:23] I talk through discord by talking to Lou and it's going to work on Codex. [20:28] That's an example. [20:29] Another thing that I did recently was try to see if I could [20:33] build an Apple Watch complication on the home screen [20:36] so that you could record a short voice memo and have that trigger a Codex job. [20:42] So, like, imagine you found a typo on a website or something like that. [20:46] Obviously it mostly works for one shot things but you would just [20:51] very briefly like you would record a voice memo to say, “Hey, [20:55] fix this typo on the landing page of this website.” [20:58] And you have the watch and you have the companion iPhone app to do that. [21:01] Yeah. So, let me show you this, this thing that I built. [21:04] This is my Apple Watch. [21:06] What I essentially built here is a way to just send a voice memo [21:12] like create a test.md file in my skills repo. [21:16] And that basically sent that memo to this iPhone app. [21:22] That is now, it transcribed the message, it understood the intent. [21:27] Based on that message is going to route it to the correct place. [21:30] In this case, it's going to understand that the skills repo means my own personal skills repo [21:35] because it has context on my GitHub account [21:38] and then it's going to run the execution of that task. [21:43] So in a couple of minutes I'm basically going to get a notification that says it created a [21:47] That second one is now successful? [21:48] Yeah. [21:48] So now it's running the execution here. [21:51] And this is all running with like Codex app server behind the scenes? [21:54] Yeah. [21:54] So if you if you go to the settings here, [21:56] you basically see that, there's two different ways it can run. [21:59] There's the typical you put an openAI API key and you can do it that way. [22:04] And the other one, what is here called back in session [22:07] is basically Codex app server running somewhere else. [22:10] That's so cool. [22:11] Roaming off of my account, connected to my GitHub with all my context. [22:15] I love this. [22:16] I can basically just start Codex jobs on the go. [22:19] It's like [22:22] it's not useful for everything but it's one of the ways [22:25] that I kind of like to play around with what becomes possible. [22:28] It's like one step closer to making [22:31] I guess, the entire implementation and implementation detail. [22:34] It's like I just talked to my watch for 10 seconds and it did absolutely everything the right way. [22:38] Incredible. [22:39] I don't think we're always there yet, depending on, you know, how complex [22:42] the actual ask is, but for small things it's kind of convenient. [22:46] It was kind of fun to play around with just to see how to work with Codex App Server. [22:50] And yeah, I just kind of try to do that. [22:52] And by the way, this is exactly what I love to see, right? [22:55] Like we open sourced the Codex CLI, the Codex harness, [22:58] and we also open sourced the Codex app server just for like fostering this like open ecosystem. [23:04] Like a builder, like you can extend the way [23:07] you use Codex in this case from your watch or your phone. [23:10] Even though, you know, we have not shipped that yet. [23:13] So this is really, really cool how you can use your coding agent the way you want. [23:17] I think you guys should build that [23:18] I think we should. [23:20] That complication on Apple Watch and on my iPhone home screen so I don't have to do it myself. [23:24] That'd be great. [23:25] Sounds like a great idea. [23:26] Let's see. [23:26] Look at that. [23:27] The whole request just arrived. [23:29] The request is right there. [23:30] Alright, out of this test.md file obviously not the most useful task but [23:34] but it just shows that it works. It's kind of fun to build as well. [23:37] Pretty cool. [23:39] That's really cool that you build this. [23:40] And by the way this is exactly why we open source not just to the Codex CLI [23:45] and the Codex harness but also the Codex App Server. [23:48] We want builders like you [23:49] to be able to kind of poke at it and also integrate that in their own ways. [23:53] It's really cool that you can have a coding agent [23:55] from your watch, which we don't have it just yet, you know. [23:58] You better build it. [23:59] We we probably should build it. [24:01] Yeah. [24:01] It's a great idea. [24:02] I'm curious, [24:03] like, we launched a ton of things in the past, like, few weeks, [24:06] both on the Codex side and GPT 5.5 our latest model. [24:11] Have you had some, like, surprising moments with these two? [24:14] Yeah, there's a couple of things. [24:16] The first one I use computer use recently that was, [24:19] it was a lot better than what I remembered or I guess I was surprised by how good it was. [24:23] I was at home last weekend doing something on a Raspberry Pi [24:26] and I needed to to do something pretty tedious. [24:29] Like copy paste a bunch of different domain URLs and go put them in some admin panel somewhere. [24:34] And I didn't really want to do it by hand. [24:36] So I told Codex to basically SSH into the Raspberry Pi. [24:40] It was like dangling from a TV to go do it for me. [24:43] And it just did. [24:45] Wow. [24:46] It found a way to just connect to the Raspberry Pi through my browser. [24:51] It basically found all these URLs that I had in a file that I gave it. [24:55] It just did the whole thing. [24:56] What was funny is that I was trying to use computer use, I guess, to save time. [24:59] And I spent the entire time just watching it. [25:02] That’s so funny. [25:02] Doing the thing for me because I was impressed by how I was going to the right place, [25:06] opening the right thing, copy pasting stuff for me. [25:09] It was just the funny thing. [25:10] I love how the cursor moves also in that you can kind of follow along the animations. [25:15] It's pretty but you don't want to touch it but you kind of do at the same time. [25:18] Because you can actually let it disappear in the background too. [25:20] Yeah, that was pretty cool and recent. [25:22] The other thing I guess is 5.5 just launched [25:26] and I have this old hackathon project that I've had for over ten years [25:30] that I use almost like my personal eval in a way, that is, [25:34] I know that models nowadays can one shot this project [25:37] but every time a new model comes out I try to do it again [25:40] just to see how much better it gets. [25:41] How many more details it gets right. How the design improves. [25:45] It's a bit of a silly project but I can tell you about it. [25:47] It was, it's a thing that was called SnapCat, which is a bit absurd. [25:51] It was basically this app for cats to take selfies. [25:54] And the way that it works is that, I mean, I guess I could just show you. [26:00] But it's essentially [26:02] a black screen with a red laser dot that moves around. [26:07] And [26:07] For the cat to catch? [26:09] Right. And so if you are a cat, we can pretend you're a cat for a second. [26:12] And you try to catch that. Every time you did [26:14] and if you don't mind trying to tap to catch the dot. [26:17] I guess I'm not as fast as a cat. [26:19] You're really not a very good [26:20] I just did. [26:21] I think I got it. [26:21] There you go. [26:22] So as you do that, every time you hit the viewport [26:25] you’re basically taking a photo with the front facing camera. [26:27] Woah [26:28] And so you can exit this cat mode with the volume buttons, [26:31] which cats don't really have easy access to. [26:34] At that point you go to the gallery. [26:36] You see the upside down pictures you just took of yourself as a cat. [26:40] You can basically select one, share it and that kind of stuff. [26:43] And what's cool here is two things. [26:45] The first one, it easily one shot this project. [26:49] It's something that over ten years ago we did at a hackathon. [26:51] It took us over a full day to build. [26:53] We had a team of five people. [26:55] It was quite difficult to do really quickly. [26:57] Now this is one really well crafted prompt and a couple of skills [27:01] and it can build the whole thing for you in one go. [27:03] How did you build the UI? [27:05] Was it like 5.5 building the UI? Was it image gen? [27:08] That's the second thing, so the first thing, it can one shot the skeleton of this app really easily. [27:14] What it did in this case after that was [27:17] I described the style of an app I wanted. [27:21] I said I wanted something light and colorful and playful and that kind of stuff given the theme. [27:26] I generated an image of that UI, and then I told Codex to build that. [27:31] And once it did that with the camera roll view that you're looking at right now, [27:35] I like that style. [27:36] And I told it “Go redesign the settings page based on the same style.” [27:39] And it generated another image and then implemented it. [27:42] Then it did the same for the home screen. [27:45] So it's a pretty different and new way to go about building UI. [27:50] So that's your eval, every time there's a new model you're just rebuilding SnapCat? [27:53] I basically just look at how easy it is for cats to take selfies. [27:56] Is it the best you've seen so far? [27:58] It's the best by far. [28:00] It is, it’s a lot better than the last time I tried this a few months ago. [28:05] It is infinitely better than what we managed to build in a full 24 hours. [28:09] Wow. [28:09] And I basically made this in one shot last night in preparation for today. [28:15] Incredible, I love this. [28:17] Well, I guess now you have selfies of me among your cats. [28:20] I'll text you one of them. [28:23] That's so cool. [28:24] Thanks so much for showing me your setup. [28:26] I love watching users building with the Codex app. [28:29] Maybe to to close out, [28:31] what would be your advice for other founders or builders, who would like to master the codex the way you do? [28:37] You know, I think it boils down to a few assumptions [28:39] that I've made early on that I think have been really, really helpful for me. [28:43] The first one is assume it's possible. [28:45] Like if you have an idea, more often than not, it can probably be done. [28:51] The second one is assume you can do it. [28:53] That one's a bit more personal [28:54] but a lot of people could be building these things. [28:57] And I think the blockers that they think they can't. [28:59] I think assume you can. [29:01] In this case, maybe be in denial even that you can't do it. [29:05] It's usually true. [29:08] And the third one is when you try to get something out of an LLM and it doesn't work, [29:12] this one's a bit difficult, [29:13] you have to put your ego aside, assume it's your fault. [29:16] Don't assume that the tool isn't capable. [29:19] Assume you haven't found a way to communicate it or convey what you want just yet. [29:24] And I guess try again. Like I've done a lot of trying again. [29:26] And I think that's what's gotten me quite a long way into building things. [29:30] So that's probably the advice I give to people. [29:33] I love that and it feels like the perfect note to end. [29:37] Anyone can just build anything. [29:38] It's true. [29:39] It sounds cheesy but it's true. [29:41] Thank you so much Matias. This was a great conversation. [29:44] Merci. [29:44] Thank you. [29:45] (speaking french)