Martin

🎬 把流程变成技能，一次构建反复用｜2/5 以前每次做类似项目，我都要重新写一遍提示，烦得要命。 结果视频里那哥们直接把整个流程打包成了「技能」。Codex 把分析 YouTube 评论的全套流程固化下来，下次直接调用就行，完全不用从头再来。 别重复造轮子了，让 AI 记住你的套路，这才是真正的降智商操作 😂

🎬 就一句提示，我让 Codex 直接吐出一个完整 Excel 分析报告｜1/5 以前我做数据分析，最后一步总要手动搞表格、插图表、刷条件格式，烦得要死 😩 结果昨天我随手甩给 Codex 一句话，它直接给我整了三个工作表： 📊 评论统计 + 情感分析 + 关键词排名，还自带图表和条件格式！ 我当时看傻了……🤯 这他妈就是数据分析的最后一公里啊！ 程序员的快乐，原来可以这么简单 ⚡

@An_yhl 是的，很多东西没跑之前都会被想得很复杂。真正动起来之后，问题反而会变具体，也更容易一步步解决。

@php_martin 先跑起来这一步真的最容易被想复杂

🎬 5 分钟跑通第一个任务 我以为装个 AI 工具至少得折腾半小时。结果从下载到跑通第一个任务，只花了 5 分钟。 打开终端，敲了几行命令，配置好 API Key。然后它开始自己干活了——读取项目、分析代码、生成修改。我甚至没来得及喝完一杯水。 原来最大的门槛不是技术，是敢不敢先跑一次。

@billy119941 我要笑死了😂

@php_martin 哦 搞错了 是我两倍速看的

🚀 Codex APP 吊打 Claude Code：额度更高、功能更全，免费账号随便用还不限速封号！ 但……90% 的人只用到了它的皮毛，核心战斗力完全没激活。 视频用一个【宠物护理网站】真实项目，手把手拆解 12 个硬核实战章节： - 沙箱权限系统如何又安全又极致高效 - Steer 引导功能如何中途精准纠偏 - Git Worktree 实现并行开发零干扰 - Skills 封装可复用专业工作流 - MCP 协议直连外部数据库 - Computer Use 让 AI 直接操控电脑完成全自动化 看完这套视频，你会把 Codex 从“有趣玩具”彻底变成真正的开发利器，安装→开发→部署全流程一次性打通！ 👉 想要把 Codex 用出 10 倍生产力？点进来看完整拆解

@An_yhl 哈哈标题确实猛了点，先把人拉进来。真正想让大家看的还是后面的项目拆解，跟着做一遍比只看功能介绍有用多了。

@php_martin 标题有点猛，项目拆解倒是挺适合跟一遍

🎬 44 秒，从代码到线上 以前部署一个项目至少 10 分钟。结果它一键部署到 Netlify，44 秒后，线上已经能访问了。 它自动构建、上传、配置域名。我什么都没做，就看着终端滚动。44 秒后，URL 出现在屏幕上。 部署不应该是一个步骤，应该是一个按钮。

@Xudong07452910 @xiangyuli 了解了，不过还好的基本上是看代码状态，输出有限，可以忽略不计

@php_martin @xiangyuli 肯定消耗呀，不过消耗很少，但是我用的模型是拉满的，所以有点浪费

我就是懒得再开一个terminal 用来git clone项目了 我就在codex里直接输了代码， 这么一个简单的任务 居然已经跑了24分钟过去了 我直接一个大后悔🥲 token也花了，时间也花了 这么久了， 居然就干了一个git clone的活儿 妈个der

@CycleDecoded 这个方向挺关键：不是简单把代码塞给模型，而是先把 codebase 变成可检索、可理解的结构。省成本只是结果之一。

给大模型接入 codebase，最心疼的就是那如流水般烧掉的 Token 费用。😭 今天安利一个帮你省下 99% Token 账单的开源 MCP 服务器 —— ！ 为什么说它是代码可视化的终极平替？ 不仅是搜索，更是脑图：它把代码资产重构成高持久性的知识图谱，AI 找 Bug、重构代码犹如神助。 快到飞起：平均仓库索引以“毫秒”计，查询速度达到“子毫秒”级。 不挑食：通吃 158 种编程语言，无论你是玩 Rust、Go 还是前端，通通秒杀。 最爽的是，它是一个单一静态二进制文件，没有任何依赖。两分钟部署完，直接白嫖极致的 AI 代码理解力。 👇 体验地址/源码：github.com/DeusData/codeb…

@UFOCaptainTing 恭喜上岸！59 天就拿到结果很猛，新的阶段继续冲。

喜大普奔！🥳🥳🥳 加入 X 的第 59 天，成功上岸！🥹😆😆 今天开始新的征程！💪💪💪

@WukongNumber1 这个确实适合入门，Codex 很多价值不在“会聊天”，而在能把实际开发流程跑起来。

mark

@Tipwotip Yeah, hitting the limit mid-flow is the worst timing. The real fix is usually being more deliberate about context, not just opening a fresh chat.

How to Stop Hitting Claude Limits Imagine you are making good progress in Claude and then suddenly it says You've reached your usage limit Annoying right? Usage limit hits is one of the biggest disappointment for all Claude users across all plans Why Does Your Limit Run Out? Claude won't count the messages It counts words (called tokens) Every time you send a message Claude re-reads the whole chat from the start So a long chat eats more limit even for a short question How Many Messages Do You Get Roughly On Estimation? These are estimations, nothing official, your actual number changes based on what you do -Free: ~30–50 short messages -Pro: ~200–400 messages -Max 5x: ~1000–2000 messages -Max 20x: ~4000+ messages How To Check Your Usage? Go to Settings -> Usage (paid plans only) You will see how much you used this session and this week What You Should Do Now To Keep Claude Running Longer 1. Edit Your Message When Claude gets it wrong do not type "no I meant..." this makes the chat longer Instead: >Hover over your message >Click the pencil icon >Fix it >Send again The wrong reply disappears and the chat stays short 2. Start a New Chat Every 15–20 Messages Long chats are always expensive After 15-20 messages try to do this: >Ask Claude: `Write a short summary of what we discussed` >Copy that summary >Open a new chat >Paste the summary as your first message >Keep going 3. One Topic = One Chat Talked about emails and then code all in one chat, stop doing this New topic = open a new chat Old messages from unrelated topics still cost you tokens 4. Ask Multiple Things in One Message Bad (costs more): > Message 1: "Summarize this" > Message 2: "Now bullet points" > Message 3: "Now a title" Good (costs less): > "Summarize this, give bullet points, and suggest a title" One message but same result and less limit used 5. Tell Claude What Part Is Wrong Not To Redo Everything If only section 2 is bad, say: "Only fix section 2, keep everything else" Don't say "redo it all" That wastes tokens on parts that were already fine. Also add this when you know what you want: "No explanation, just the output" Claude talks less more, you save more 6. Use Voice to Type Your Prompts When you type, you may write vague things like "make it better" Claude guesses wrong, you correct, then more messages and so more cost When you speak, you naturally say more: "Make section 2 casual, keep the data, don't touch anything else" Claude gets it right at the very first try 7. Write Short Prompts Let Claude Ask You Questions A long prompt repeats in every future message which is a bad deal Try: "I want to [task]. Ask me questions before you start." Cheap and easy 8. Convert Files Before Uploading A PDF page = ~1500–3000 tokens A screenshot = ~1300 tokens Make it cheaper: >Go to (https://doc .new) in your browser >Paste only the text you need >Download as Markdown (.md) >Upload the .md file not the PDF Cropping screenshots tight also helps A tight crop can go from 1300 tokens to under 100 9. Use Projects for Files You Use Again and Again When you are uploading the same file to different chats You pay for it every time Fix it: >Go to Projects in the left sidebar >Click New Project >Upload your file in Project Knowledge >All chats inside this project use the file for free after that Claude only loads what is relevant not the whole file everytime 10. Don't Upload Files the Task Doesn't Need Only attach files Claude actually needs right now No files needed? Start with zero files 11. Use the Right Model >Quick simple question = Haiku >Everyday writing, emails = Sonnet >Hard thinking, complex work = Opus How to switch: Click the model name at the top of the chat and pick a different one >Haiku = cheapest >Opus = most expensive Don't use Opus to write a short email 12. Turn Off Features You Are Not Using Web search, Research mode and connectors add tokens even when you don't need them How to turn off: >Click the + icon next to where you type >Turn off Research, web search and any connected tools you don't need >Click the model name -> turn off Extended Thinking if you don't need it Default: keep everything off Turn things on only when needed 13. Lower the Effort Level for Easy Tasks Claude has an effort setting High effort = more tokens used For simple tasks: click the model name -> find the effort or thinking level -> set to Normal 14. Set Up User Preferences Without this, you waste messages every new chat re-explaining yourself How to do it: >Click your name / avatar -> Settings >Go to General -> Personal Preferences >Write your name, job, writing style, what you like >Save Claude knows who you are in every new chat so zero setup messages wasted 15. Set a Default Writing Style Go to the + icon or model selector -> find Styles -> set to Concise Claude will give shorter answers by default Shorter answers = fewer tokens used 16. Turn On Memory Memory makes Claude remember you across chats without you re-explaining How to enable: >Settings -> Capabilities* >Toggle on Generate memory from chat history Or: make a short .md file (under 2000 words) with your context, upload it to a Project This works better for detailed preferences 17. Keep Project Instructions Short Project instructions load before every chat in that project Long instructions = wasted tokens before you type anything Keep them under 500 words General stuff only Put task details in the chat 18. Save a Handoff Note for Long Projects At the end of a work session: > "Write a short handoff note: what we decided, what's next" Save it and on Next session, start with: > "Read this first: [paste note]. Now let's continue" No need to explain everything again 19. Use the Same Prompt Structure for Repeated Tasks Claude partially remembers repeated prompt patterns If you use the same format every time like always starting with "I want to [task] to achieve [goal]" it gets cached Build one template Swap only what changes inside it 20. Use Claude's Chat Search Claude can search your old chats No need to copy-paste old context Just ask: "What did we talk about last time regarding [topic]?" Make sure it is on: Settings -> Capabilities -> Search and reference chats ON 21. Plan in Chat First Create Files After File creation (making a spreadsheet, doc, presentation) costs more than chatting Do the thinking in chat first Once you know exactly what you want then say "build it" Don't rebuild files 3 times because the plan wasn't clear That is very expensive 22. For Claude Code Be Very Specific Claude Code explores files on its own if you give it free hand and that costs tokens fast Bad: "Help me fix the dashboard" Good: "In /src/Dashboard.jsx add a loading spinner when data is fetching do not touch any other file" 23. Use CLAUDE.md for Claude Code Settings Claude Code reads a file called CLAUDE.md before every task Put your standing rules there: folder, language, what not to touch How to make it: >In your project root folder, make a new file called CLAUDE.md >Write your rules (keep it under 300 words) >Claude Code reads it automatically every time Keep it short A long CLAUDE.md makes Claude ignore your actual instructions 24. Spread Work Across the Day The limit window is 5 hours rolling Don't use everything in one go Split into morning, afternoon, evening sessions By the time evening comes your morning usage has mostly cleared. 25. Heavy Tasks Only On Off-Peak Hours Claude's limits feel tighter during busy hours (~5 AM–1 PM PT on weekdays) If you have a big task, try doing it in the afternoon or evening You will get more out of it 26. Watch Your Usage Don't Wait to Hit Check it before a big task Settings -> Usage You will see: >How much of your 5-hour session is used >How much time is left >Weekly limit progress If you are close, pause Come back in an hour Better than burning it all and waiting 27. Don't Use Claude for Things It's Bad At Use This Instead: >Make an image = Midjourney, DALL-E, Gemini >Latest news = Grok, Perplexity, Google >Simple search = Google, DuckDuckGo Spending 5 messages on something Claude cannot do well wastes 5 messages 28. Ask Do I Even Need AI for This? Before opening Claude, ask yourself: can I Google this? Can I do it myself in 2 minutes? If yes then do that Save Claude for the stuff that actually needs it Note: All of the settings discussed over here might not be available in free plan, choose accordingly.

@Xudong07452910 @xiangyuli 什么？claude命令行执行 !ls 也消耗token？我也经常用 !gst 啊

@xiangyuli 哈哈，我有个坏毛病就是经常使用ls命令去查看本地文件目录，我有时候还在Claude code命令行里敲这个命令😂，这是不嫌token少呀

@xiangyuli 这种小任务最适合直接在本地终端解决，交给 agent 反而容易把简单事变复杂。成本感一下就出来了。

@keyan898786 @Amateur8888 播放量很容易刷，但稳定的商业转化才是真本事。这组数据挺说明问题的。

@Amateur8888 报价70万，不等于最终成交70万；一个月12单，也不等于全部都是定制视频。 但这组数据至少说明：真正值钱的不是播放量，而是稳定的内容能力和商业转化能力。

影视飓风，一个月变现154万，这是光抖音的，他收入大头应该在B站，按照1单定制，12单植入算的话，光抖音单月就154万商单收益，吸金能力极强

@zqiusu001223 @Lonely__MH @thsottiaux 哈哈哈，就是帮你存着，你想什么时候取出来用都行

@php_martin @Lonely__MH @thsottiaux 什么是存款式重置？

不好意思，昨天说话声音确实大了点😅 @thsottiaux enjoying！🥳🥳

@Forest_GoGoGo 是的，最大的好处就是稳定可控：不用担心 API 额度和网络波动，很多日常修 bug、改页面、跑脚本的任务，本地模型已经够用了。

@php_martin 本地模型方便

🎬 定时任务，让 AI 自己上班 你以为自动化就是写脚本。结果它用插件，设置好时间，到点自己跑——不需要我盯着。 我设了一个每天早上 8 点自动更新依赖的任务。第二天醒来，它已经跑完了。133 秒配置，以后每天省 10 分钟。 AI 最厉害的不是帮你干活，是它自己知道什么时候该干活。

@keyan898786 这个流程很适合编排一个可以串联的多个skill处理流👍

@php_martin 很多人把 Codex 当成更强的聊天工具，但它真正有价值的地方，是把专业经验封装成可以重复执行的工作流。 我现在也在尝试把标书审核、资源匹配和项目可行性分析做成 Skills，期待这套实战拆解。