Lucien

Speed up your Gemma 4 workflows by up to 3x with Multi-Token Prediction (MTP) drafters. Standard LLM inference is fundamentally memory-bandwidth bound, creating a latency bottleneck as billions of parameters travel from VRAM just to generate a single token. We're working to ease this bottleneck with MTP drafters for @googlegemma 4.

developers.openai.com/api/docs/guide… **NEW: GPT-5.5 Prompting Guide** "GPT-5.5 works best when prompts define the outcome and leave room for the model to choose an efficient solution path. Compared with earlier models, you can often use shorter, more outcome-oriented prompts: describe what good looks like, what constraints matter, what evidence is available, and what the final answer should contain. Avoid carrying over every instruction from an older prompt stack. Legacy prompts often over-specify the process because earlier models needed more help staying on track. With GPT-5.5, that can add noise, narrow the model’s search space, or lead to overly mechanical answers. For more detail on GPT-5.5 behavior changes, start with the Using GPT-5.5 guide. This guide focuses on prompt changes that follow from those behavior changes. The patterns here are starting points. Adapt them to your product surface, tools, evals, and user experience goals."

NEW POST Thoughtworks internal IT use a workflow for agentic programming called Structured-Prompt-Driven Development (SPDD). @WeiZhang595190 and Jessie Jie Xia describe how this works with a simple example plus details in a github project. martinfowler.com/articles/struc…

@garybernhardt AI will not replace programmers. Software engineers, who know how to wrangle AIs, will replace “vibe“ programmers. And when those software engineers meet Dykstra in heaven, he will smile and give them all hugs.

Starting June 1st, GitHub Copilot will move to a usage-based billing model as GitHub Copilot supports more agentic and advanced workflows. In early May, you'll see a preview bill experience, giving visibility into projected costs before the transition. 👉 Read more about the upcoming change: github.blog/news-insights/…

作者画的这图效果挺好，不过好像没提供提示词，我仿照着写了一个 GPT Image 2 的提示词模板： > 请创作一张关于「【数学概念/知识点】」的数学可视化信息图，目标是帮助读者直观理解：它是什么、为什么成立、几何/结构直觉是什么、在不同场景下如何表现。画面要像高质量数学讲义 + 手绘教育海报，优雅、清晰、信息丰富，但不要杂乱。 > > 视觉风格：竖版或横版均可，干净的浅色纸张背景，深蓝标题，黑色/深灰正文线条，搭配少量优雅的蓝色、青绿色、金色、红色强调色。使用圆角卡片、细线边框、编号标签、手绘箭头、局部放大框和总结栏。整体要美观、平衡、有学术感，让人一眼看懂这个概念的“结构”和“为什么”。

Anthropic 的哲学家@AmandaAskell 最近参加了一个访谈，在访谈中她分享了自己探索好奇领域的一个方法。 提示词大概是： 我希望你从「xx」领域里选一个大概研究生水平的概念。然后我希望你通过写一个寓言的方式，间接地把这个概念完整讲出来。最好一直到快结尾时，人才会慢慢意识到这个概念究竟是什么。然后在故事之后，再补一段解释，把你刚才真正要讲的概念说清楚。

这个通过寓言故事学习概念的提示词真的很棒，我做了一点点小调整，方便你使用 --- 提示词 --- 我在学习：【Agent Harness】 这个概念。 我希望你通过写一个寓言的方式，间接地把这个概念完整讲出来。最好一直到快结尾时，人才会慢慢意识到这个概念究竟是什么。然后在故事之后，再补一段解释，把你刚才真正要讲的概念说清楚。 ---- 实际结果 --- 玉室里的先生 很久以前，青云山里有一间密封的玉室。 玉室没有门，没有窗，墙壁厚达三尺。里面住着一位先生。没有人知道他从哪里来，只知道他极其聪明——天文地理、治国用兵、农桑水利，什么都懂。 但他有一个问题：他看不见外面，听不到外面，也走不出去。 第一个找到玉室的是齐国。齐国的做法很简单：在墙上凿了一个小洞，派人对着洞大喊问题，然后等先生把答案写在竹简上，从洞里递出来。 一开始效果还行。"今年麦子什么时候种？"先生答了，齐国照做，收成不错。 但问题很快就来了。齐国问："北边的敌军怎么打？"先生答了一套方略，头头是道。齐国照做，大败。 齐国大怒，觉得先生是骗子。但事后复盘才发现：先生根本不知道齐国的兵力有多少，不知道地形是山地还是平原，不知道敌军用的是骑兵还是步兵。他在玉室里凭空想出来的方略，只是"一般来说应该这么打"。 而齐国从头到尾，只问了一句"怎么打"。 楚国也找到了一间一模一样的玉室，里面住着一位同样聪明的先生。 楚国有个年轻的官员叫沈括。沈括第一次去玉室的时候，也只是对着洞喊话。先生的回答聪明但空泛，和齐国的经历差不多。 沈括没有骂先生。他想了三天，然后开始做一件奇怪的事。 他先派了两个识字的文吏，常驻在玉室旁边。一个负责把问题写清楚，连同所有背景材料一起整理好，从洞里递进去。另一个负责把先生递出来的答案仔细誊抄，遇到看不懂的地方，再递回去追问。 仅这一步，效果就明显不同。先生开始给出具体的、能用的建议。 但沈括没有停下来。他发现先生经常说"需要看看地图""要知道土壤的情况"，可先生看不见任何东西。于是沈括又安排了一批人，专门负责把地图、账册、田间观察的记录做成文字描述，送进玉室。 先生有了"眼睛"。 然后沈括注意到另一个问题。先生会说"把河堤往东移二十丈"，但先生自己动不了。说完就说完了。于是沈括组建了一支执行队，他们的职责是：接收先生的指令，去现场执行，然后把执行结果——成功的、失败的、意外的——全部记录下来，再送回玉室。 先生有了"手脚"，而且能知道手脚做了什么。 到这一步，楚国已经靠先生解决了好几个棘手问题。但沈括又发现了一个麻烦：先生偶尔会说错。他会用一种极其自信的口吻说出明显离谱的话。比如他声称某种草药能治疫病，但楚国的医官翻遍药典，找不到这种草药——先生编了一个不存在的东西，还说得言之凿凿。 沈括的做法很务实。他不要求先生不犯错，他在外面加了一层校验：每次先生给出关键建议，都有专人去交叉核实。能核实的才采纳，核实不了的标记存疑。先生的错误从流程上被拦截了，不需要先生自己改正。 先生有了"免疫系统"。 还有一个问题更微妙。先生每次回答问题的时候，并不记得上次说过什么。同一个问题问两遍，可能得到两个矛盾的答案。沈括发现先生的"记忆"是断裂的。 于是他安排了一个专职的记录官。这个人的工作只有一件事：把历次和先生的对话全部整理成册，每次递新问题进去的时候，连同之前相关的记录一起送进去。先生拿到的不再只是一个孤立的问题，而是一个问题连同它的全部历史。 先生有了"记忆"。 几年后，周边的国家都听说了玉室的事情。很多国家纷纷去找自己的玉室。他们都找到了——每间玉室里的先生都一样聪明，一样什么都看不见，一样走不出去。 但结果天差地别。 有的国家只凿了一个洞，全靠喊。先生说什么他们做什么，做砸了就骂先生蠢。 有的国家给先生递了一些材料，但执行全靠将军们自己理解，先生永远不知道自己的方案最后被执行成了什么样。 有的国家什么都做了，唯独没有安排校验。先生一本正经地胡说八道的时候，全国上下照单全收。 只有楚国，围绕先生搭建了一整套体系：感知、行动、校验、记忆。先生还是那个先生，但在楚国，他像是拥有了一副完整的身体。 有一天，沈括的学生问他："老师，先生到底有多聪明？" 沈括想了想，说："我不知道他有多聪明。我只知道一件事——他的聪明能发挥出几成，从来不取决于他自己。取决于我们在外面给他造了一副什么样的身体。" 学生又问："那我们做的这些事情，叫什么？" 沈括说："我管它叫'造身体的学问'。先生是脑子，我们造的是眼耳手脚、筋骨血脉。脑子谁都一样，谁的身体造得好，谁的先生就更有用。" 这个寓言在说什么 "玉室里的先生"就是大语言模型（LLM）。 它极其聪明，但它是一个封闭系统——没有感知，没有行动能力，没有持久记忆，也没有自我纠错的机制。你给它什么输入，它就基于这些输入来"想"，然后输出一段文字。仅此而已。 沈括围绕先生搭建的那套体系，在今天有一个专门的名字：Harness。 故事里的每个角色都对应着 Harness 的一个组件： 文吏（感知层）：负责把外部信息整理好，送进模型的上下文窗口。这包括 system prompt 的设计、上下文的组装、用户输入的预处理。模型的"视野"取决于你往上下文里塞了什么。 执行队（行动层）：模型说"执行这个命令"或者"调用这个 API"，但它自己做不了。Harness 负责接收模型的工具调用指令，真正去执行，再把结果反馈回来，形成一个"思考→行动→观察→再思考"的循环。 校验官（容错层）：模型会产生幻觉，会犯错。Harness 不指望模型自己不犯错，而是在外部设置护栏——linter、测试、交叉验证、结构化输出校验——从流程上拦截错误。 记录官（记忆层）：模型的上下文窗口是有限的，而且每次对话之间默认没有记忆。Harness 负责管理短期记忆（当前对话中哪些信息该保留、哪些该丢弃）、长期记忆（跨对话的持久化信息）和项目级知识（主动为模型准备的文档、规范、历史记录）。 而沈括做的事——设计、搭建、持续改进这整套体系——就是 Harness Engineering。 故事里最关键的一句话是沈括说的那句：先生的聪明能发挥几成，从来不取决于他自己，取决于外面给他造了什么样的身体。 翻译成今天的话就是：模型能力是地板，Harness 质量才是天花板。 同一个模型，套上不同的 Harness，表现可以天差地别。Claude Code 和 OpenClaw 用的是同一个 Claude 模型，但因为 Harness 不同，一个是程序员的身体，一个是项目经理的身体，能做的事情完全不一样。 这就是为什么越来越多的人意识到：在 Agent 时代，真正决定产品好坏的，往往不是你选了哪个模型，而是你围绕模型搭建了什么样的体系。

Introducing Expect Let agents test your code in a real browser 1. Run Claude Code / Codex to QA your app 2. Watch a video of every bug found 3. Fix and repeat until passing Run as a CLI or agent skill. Fully open source

Introducing the new dev-browser cli. The fastest way for an agent to use a browser is to let it write code. Just `npm i -g dev-browser` and tell your agent to "use dev-browser"

We just released Claude Code channels, which allows you to control your Claude Code session through select MCPs, starting with Telegram and Discord. Use this to message Claude Code directly from your phone.

Released today: /loop /loop is a powerful new way to schedule recurring tasks, for up to 3 days at a time eg. “/loop babysit all my PRs. Auto-fix build issues and when comments come in, use a worktree agent to fix them” eg. “/loop every morning use the Slack MCP to give me a summary of top posts I was tagged in” Let us know what you think!

Markdown-first repo automation lands in technical preview. • Define issue triage, PR review, CI debug, maintenance, in Markdown. • Run gh aw to compile to standard Actions that use Copilot CLI or other agents. Check out Agentic Workflows → github.blog/changelog/2026…

The whole SDD (Spec Driven Design) thing is very valuable for AI development; but it carries the risk of BUFD. We'd like to avoid that. So the iterative nature of software development should be maintained. Just enough specs for the stories in this sprint -- just like we used to say that QA should write just enough acceptance tests for this sprint. Also, we will be modifying those specs as the sprints go by. Specs will be co-authored by the humans and the AI, but with final approval, ferociously defended, by the humans. And don't fret about the word "sprint", even if your sprint size is a day, an hour, or a microsecond, the same rules apply. Small Steps.

In case it’s not clear in the docs: - Ancestor CLAUDE.md’s are loaded into context automatically on startup - Descendent CLAUDE.md’s are loaded *lazily* only when Claude reads/writes files in a folder the CLAUDE.md is in. Think of it as a special kind of skill. We designed it this way for monorepos and other big repos, tends to work pretty well in practice.

📣 Open call to agent builders: Let's read agent skills from `.agents/skills`, so people don't have to manage separate folders per agent. Today we pulled the trigger for Codex to read `.agents/skills`. Goal is to deprecate `.codex/skills`. Pls like/tag/RT for momentum.

I guess iz a smart boy. I even made a slash command for this directly inside of claude code. What's cool about this is it takes advantage of the AskUserTool so Claude will first ask you questions to improve the quality of the prompt It will then call Codex via the command line and write a very detailed plan, which will even tell you which tasks can be run in parallel so you can launch swarms to complete tasks with no remaining dependencies. its just a slash command /codex-plan.md it will save the plan to codex-plan.md in your CWD That way you never even have to leave Claude Code, and you take advantage of the AskUserTool which is more robust in Claude. --- description: Create a detailed implementation plan using Codex 5.2 with xhigh reasoning argument-hint: "" allowed-tools: Read, Write, Bash, AskUser --- # Codex Plan Command You are being asked to create a detailed implementation plan using a Codex subagent. Your job is to: 1. Understand the user's planning request 2. Ask clarifying questions using AskUser to improve plan quality 3. Craft an excellent, detailed prompt for Codex 4. Execute Codex to generate and save the plan **Always uses:** `gpt-5.2-codex` with `xhigh` reasoning ## User Request ``` $ARGUMENTS ``` ## Step 1: Analyze the Request Look at what the user wants to plan. Identify: - What is the core goal? - What technology/domain is involved? - What aspects are ambiguous or underspecified? - What decisions would significantly impact the plan? ## Step 2: Ask Clarifying Questions **Use AskUser to ask 3-6 targeted clarifying questions** before generating the plan. Good clarifying questions: - Narrow down scope and requirements - Clarify technology choices - Understand constraints (time, budget, team size) - Identify must-haves vs nice-to-haves - Uncover integration requirements - Determine security/compliance needs ### Example Question Patterns **For "implement auth":** - What authentication methods do you need? (email/password, OAuth providers like Google/GitHub, SSO, magic links) - Do you need role-based access control (RBAC) or just authenticated/unauthenticated? - What's your backend stack? (Node/Express, Python/Django, etc.) - Where will you store user credentials/sessions? (Database, Redis, JWT stateless) - Do you need features like: password reset, email verification, 2FA? - Any compliance requirements? (SOC2, GDPR, HIPAA) **For "build an API":** - What resources/entities does this API need to manage? - REST or GraphQL? - What authentication will the API use? - Expected scale/traffic? - Do you need rate limiting, caching, versioning? **For "migrate to microservices":** - Which parts of the monolith are you migrating first? - What's your deployment target? (K8s, ECS, etc.) - How will services communicate? (REST, gRPC, message queues) - What's your timeline and team capacity? **For "add testing":** - What testing levels do you need? (unit, integration, e2e) - What's your current test coverage? - What frameworks do you prefer or already use? - What's the most critical functionality to test first? ## Step 3: Gather Context After getting answers, also gather relevant context: - Read key files in the codebase if applicable - Check existing architecture/patterns - Note any existing plans or documentation ## Step 4: Craft the Codex Prompt Create a detailed prompt that includes: 1. **Clear objective** - What plan needs to be created 2. **All requirements** - Everything learned from clarifying questions 3. **Constraints** - Technology choices, timeline, team size 4. **Context** - Relevant codebase info, existing patterns 5. **Plan structure** - What sections the plan should include 6. **Output instructions** - Write to `codex-plan.md` in current directory **Critical instruction to include:** Tell Codex to NOT ask any further clarifying questions - it has all the information it needs and should just write the plan and save the file. ## Step 5: Execute Codex ```bash codex exec --full-auto --skip-git-repo-check \ -c model=gpt-5.2-codex \ -c model_reasoning_effort=xhigh \ --output-last-message /tmp/codex-plan-result.txt \ "YOUR_CRAFTED_PROMPT_HERE" ``` Then show the results: ```bash cat /tmp/codex-plan-result.txt ``` ## Example Full Flow **User:** `/codex-plan help me implement auth into my webapp` **You ask (using AskUser):** ``` I'd like to create a detailed auth implementation plan for you. A few questions first: 1. What authentication methods do you need? - Email/password - OAuth (Google, GitHub, etc.) - Magic links - SSO/SAML 2. Do you need role-based access control (admin, user, etc.) or just authenticated vs not? 3. What's your backend stack? (e.g., Node/Express, Python/FastAPI, Next.js API routes) 4. Where will you store sessions? (Database, Redis, JWT stateless) 5. Do you need any of these features? - Password reset - Email verification - Two-factor authentication (2FA) - Remember me / persistent sessions 6. Any compliance requirements? (SOC2, GDPR, HIPAA) ``` **User answers:** "Email/password and Google OAuth. Using Next.js with Prisma and Postgres. Need basic roles (admin/user). Yes to password reset and email verification. No 2FA for now. No special compliance." **You craft prompt for Codex:** ``` Create a detailed implementation plan for adding authentication to a Next.js web application. ## Requirements - Authentication methods: Email/password + Google OAuth - Framework: Next.js (App Router) - Database: PostgreSQL with Prisma ORM - Role-based access: Admin and User roles - Features needed: - User registration and login - Password reset flow - Email verification - Google OAuth integration - Session management - NOT needed: 2FA, SSO, special compliance ## Plan Structure Use this template structure: ```markdown # Plan: [Task Name] **Generated**: [Date] **Estimated Complexity**: [Low/Medium/High] ## Overview [Brief summary of what needs to be done and the general approach, including recommended libraries/tools] ## Prerequisites - [Dependencies or requirements that must be met first] - [Tools, libraries, or access needed] ## Phase 1: [Phase Name] **Goal**: [What this phase accomplishes] ### Task 1.1: [Task Name] - **Location**: [File paths or components involved] - **Description**: [What needs to be done] - **Dependencies**: [Task IDs this depends on, e.g., "None" or "1.2, 2.1"] - **Complexity**: [1-10] - **Test-First Approach**: - [Test to write before implementation] - [What the test should verify] - **Acceptance Criteria**: - [Specific, testable criteria] ### Task 1.2: [Task Name] [Same structure...] ## Phase 2: [Phase Name] [...] ## Testing Strategy - **Unit Tests**: [What to unit test, frameworks to use] - **Integration Tests**: [API/service integration tests] - **E2E Tests**: [Critical user flows to test end-to-end] - **Test Coverage Goals**: [Target coverage percentage] ## Dependency Graph [Show which tasks can run in parallel vs which must be sequential] - Tasks with no dependencies: [list - these can start immediately] - Task dependency chains: [show critical path] ## Potential Risks - [Things that could go wrong] - [Mitigation strategies] ## Rollback Plan - [How to undo changes if needed] ``` ### Task Guidelines Each task must: - Be specific and actionable (not vague) - Have clear inputs and outputs - Be independently testable - Include file paths and specific code locations - Include dependencies so parallel execution is possible - Include complexity score (1-10) Break large tasks into smaller ones: - ✗ Bad: "Implement Google OAuth" - ✓ Good: - "Add Google OAuth config to environment variables" - "Install and configure passport-google-oauth20 package" - "Create OAuth callback route handler in src/routes/auth.ts" - "Add Google sign-in button to login UI" - "Write integration tests for OAuth flow" ## Instructions - Write the complete plan to a file called `codex-plan.md` in the current directory - Do NOT ask any clarifying questions - you have all the information needed - Be specific and actionable - include code snippets where helpful - Follow test-driven development: specify what tests to write BEFORE implementation for each task - Identify task dependencies so parallel work is possible - Just write the plan and save the file Begin immediately. ``` **Execute and return results.** ## Important Notes - **Always ask clarifying questions first** - Don't skip this step - **Use AskUser tool** - This is interactive planning - **Always use gpt-5.2-codex with xhigh reasoning** - No exceptions - **Tell Codex not to ask questions** - It should just execute - **Output file:** `codex-plan.md` in current working directory - **Use --full-auto** Now analyze the user's planning request above, ask your clarifying questions, and then generate and execute the Codex plan.