49 块钱的桌面看板,我用 Codex 做的第一个项目A 49-Yuan Desktop Dashboard: My First Project Built with Codex


AI free token/AI免费积分

本文来源:https://mp.weixin.qq.com/s/OmBULWsUE2YjdPjT617hMQ
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

“

49 块 的墨水屏,第一次把 Codex 从聊天框,用成每天会抬头看 的东西。

—— 老冬瓜 Daniel

花了两周,踩过 BLE 连不上、额度对不上、bug 越修越多。

闲鱼成交价 49 块,一块 4.2 + 2.13 的双屏墨水屏,现在搁在我键盘前面。上屏是 Codex 额度和 GitHub 贡献,下屏是自带的时钟日历。

这是我第一次把 Codex 从聊天框,用成一个每天会抬头看的东西。

49 块钱的桌面看板,我用 Codex 做的第一个项目

本文看点

01

先调研再画界面

02

额度冻住不是渲染错

03

修好一个旁边炸了

01

WHY

一开始我并不知道要做什么

刚学会用 Codex 的时候,没有项目。

打开工具,第一反应不是「今天做哪个能力」,而是「我到底该拿它开发什么」。

后来刷到小红书上很多人在做墨水屏看板。想法不是慢慢想出来的,是一下子蹦出来的:把最常盯的两件事,从终端里挪到桌上。

49 块钱的桌面看板,我用 Codex 做的第一个项目

第一件必然是额度

Plus 加上 5 小时窗口之后,额度变得一言难尽。写得顺的时候最容易忘,等限制弹出来,节奏已经断了。我需要的不是再开一个网页去查,是抬头就能看见:这 5 小时还剩多少,这周还剩多少。

第二件是项目有没有在往前走

这块屏只回答两个难题:Codex 还剩多少额度;手头这几个仓库,这周有没有好好提交。

49 块钱的桌面看板,我用 Codex 做的第一个项目

「它不是 AI 很酷的摆件,是一张会说话的便利贴。」

02

PLAN

先调研,再拍板,最后才画界面

49 块钱的桌面看板,我用 Codex 做的第一个项目

想法有了,下一步不是让它直接开写。

规划交给 sol-high,让它先当机型经理,用提问把最小 MVP 和边界问清楚。顺序必须是:先调研思路,思路定了再细化。界面再关键,也不能倒过来先画图。

这类看板并不新。GitHub 上早有人做过,所以第一件事是让它去调研,再对着这块能走蓝牙同步的 EPD 墨水屏给思路。

它给出的第一版很朴素:电脑采集 Codex 额度,再跑终端脚本,把画面推到屏上。

我原本想把采集放到服务器上,24 小时跑。它觉得这条链路不安全。可公司电脑也不是一直开着,采集不能绑在工位上。

最后折中落在家里那台几乎不停机的飞牛 NAS 上:Docker 里跑采集;结果写成共享目录里的文件;电脑开机且屏在旁边时,终端只读这份数据,再经蓝牙写入墨水屏。

「电脑负责看见屏的时候再推,NAS 负责额度一直有人盯。」

采集和表明拆开,凭据也就不必跟着桌面脚本满天飞。思路定了,才把上屏 UI 画死,后面所有生成和返工都对照这张图。

49 块钱的桌面看板,我用 Codex 做的第一个项目

03

BLE

BLE 反而不是最大的坑

蓝牙这块,压力比我预想的小。这块板子的作者自己就有公开仓库。直接让 Codex 去读,大致就能弄明白怎么连、怎么把图画进槽位。

「真正反复沟通的,是数据。」

04

QUOTA

额度不对,多半不是渲染错了

sol 把任务规划完之后,执行交给了 luna-high。对 Plus 来说,这个模型量大管饱,缺点也很显著:没那么聪明。采集的数据来源、采集时间、更新时间,这些边界本该在第一天就写死。

!踩坑提示

NAS 端的 Codex 授权缓存过期。采集停了,额度不会自己长出来,看板上的数字也就冻在那里。

我当时没找到根因,只是一顿让它改。这就是第三个坑的开始。

05

FIX

小能力修好了,旁边那个能力又炸了

49 块钱的桌面看板,我用 Codex 做的第一个项目

bug 越改越多。额度终于更新了,底下的 Token 使用量又错。这个小能力活了,另一个小能力就炸。不是模型不肯干活,是任务边界没说清,它只能在局部打补丁。

于是我把规划权交回 sol-high,让它重新调研。多数人用 Mac 本机采,因为 Mac 几乎 24 小时开着。我立刻改成 Mac 采集,写入 NAS 共享目录,公司电脑再读这份文件,执行蓝牙同步。

49 块钱的桌面看板,我用 Codex 做的第一个项目

!踩坑提示

家里那台 Mac 一直插着电,但只要一休眠,采集照样停。

链路被改成一波三折,最后还是回到原点:仍然由 NAS 采集。只是加了一层保险,缓存失效时用 Bark 推到手机,重新授权一次就行。

看板能用了。长期跑的机器可以休眠、可以掉授权,平台必须把这件事当成正常情况,而不是意外。

∞

END

这 49 块钱教会我的

一开始就要把任务边界和架构想清楚。每个能力尽量解耦,不要一次做太复杂。我不盯它代码漂不漂亮,但必须能指出业务逻辑错在哪,才能把走歪的路拽回来。

这块 49 块的屏能亮着,不是因为第一版思路完美,是因为错了三次之后,我终于肯把「授权会过期、电脑会睡觉」写进外观里。

「有想法很关键,规划更关键。」

不知道 AI 能干嘛的时候,先做起来就行。只有在不断的实践中,才能一步步学到一些东西。

END

我是老冬瓜 Daniel,平时分享 AI 观察和折腾记录。觉得有收获,欢迎点赞、在看、转发。

"A 49-yuan e-ink screen is the first time I turned Codex from a chat box into something I look up at every day."

— Old Winter Melon Daniel

It took two weeks, with BLE connection failures, mismatched quotas and bugs multiplying as I fixed them.

The Xianyu closing price was 49 yuan for a 4.2 + 2.13 dual-screen e-ink panel, now sitting in front of my keyboard. The top screen shows Codex quota and GitHub contributions; the bottom screen shows a built-in clock and calendar.

This is the first time I turned Codex from a chat box into something I look up at every day.

A 49-Yuan Desktop Dashboard: My First Project Built with Codex

Highlights of this article

01

Research first, then design the UI

02

A frozen quota is not a rendering bug

03

Fix one and the neighbour blows up

WHY

At first I had no idea what to build

When I first learned to use Codex, I had no project.

Opening the tool, my first reaction was not "which capability should I build today", but "what on earth should I develop with it".

Later I saw many people on Xiaohongshu building e-ink dashboards. The idea did not come slowly - it hit me all at once: move the two things I stare at most out of the terminal and onto my desk.

A 49-Yuan Desktop Dashboard: My First Project Built with Codex

The first thing had to be quota

After Plus added the 5-hour window, the quota became hard to put into words. You most easily forget when writing flows; by the time the limit pops up, the rhythm is already broken. What I needed was not another webpage to check, but something I could see by looking up: how much of this 5-hour window is left, how much of this week is left.

The second is whether the project is moving forward

This screen answers only two hard questions: how much Codex quota is left; and whether these few repos I am handling got committed properly this week.

A 49-Yuan Desktop Dashboard: My First Project Built with Codex

"It is not a cool AI ornament; it is a talking sticky note."

PLAN

Research first, decide, then design the UI

A 49-Yuan Desktop Dashboard: My First Project Built with Codex

With the idea in hand, the next step is not to let it start writing.

I handed planning to sol-high, letting it act as model manager first, using questions to pin down the minimal MVP and the boundaries. The order must be: research the approach first, then refine once the approach is set. However key the UI is, you cannot reverse it and draw first.

This kind of dashboard is not new. People on GitHub built one long ago, so the first step was to have it research, then give an approach against this EPD e-ink screen that can sync over Bluetooth.

Its first version was very plain: the computer collects the Codex quota, then runs a terminal script to push the screen to the panel.

I originally wanted to put collection on a server, running 24 hours. It thought this chain was insecure. But the work computer is not always on either, so collection cannot be tied to the workstation.

The compromise finally landed on the Feiniu NAS at home that almost never shuts down: collection runs in Docker; results are written to a file in the shared directory; when the computer is on and the screen is nearby, the terminal only reads this data, then writes it to the e-ink screen over Bluetooth.

"The computer is responsible for pushing when it sees the screen; the NAS is responsible for always keeping an eye on the quota."

Splitting collection and display also means credentials do not have to fly around with the desktop script. Once the approach was set, the top-screen UI was locked down, and all later generation and rework were checked against that picture.

A 49-Yuan Desktop Dashboard: My First Project Built with Codex

BLE

BLE was not actually the biggest pitfall

On the Bluetooth side, the pressure was less than I expected. The board's author has a public repo. Just letting Codex read it, you can roughly figure out how to connect and how to draw into the slots.

"What really took repeated back-and-forth was the data."

QUOTA

Wrong quota is mostly not a rendering error

After sol finished planning the task, execution was handed to luna-high. For Plus, this model delivers volume and quantity, but its drawback is also obvious: not that smart. The data source, collection time and update time - these boundaries should have been nailed down on day one.

! Pitfall tip

The Codex auth cache on the NAS side expired. Collection stopped; the quota would not grow back on its own, so the numbers on the dashboard froze there.

I had not found the root cause then, just kept telling it to change. That is where the third pitfall began.

FIX

The small capability got fixed, and the one next to it blew up

A 49-Yuan Desktop Dashboard: My First Project Built with Codex

Bugs multiplied as I fixed them. The quota finally updated, but the Token usage below was wrong again. This small capability lived, another small capability blew up. It is not that the model refused to work; the task boundaries were not made clear, so it could only patch locally.

So I handed planning back to sol-high and let it re-research. Most people collect on a Mac locally because a Mac is almost always on 24 hours. I immediately switched to Mac collection, writing to the NAS shared directory, and the work computer reads this file and executes the Bluetooth sync.

A 49-Yuan Desktop Dashboard: My First Project Built with Codex

! Pitfall tip

That Mac at home is always plugged in, but as soon as it sleeps, collection stops just the same.

The chain was changed into a twisty three-act, and finally returned to the starting point: still collected by the NAS. Only a layer of insurance was added - when the cache fails, push to the phone via Bark, and just re-authorize once.

The dashboard works. The machine that runs long-term can sleep, can drop authorization; the platform must treat this as normal, not an accident.

END

What these 49 yuan taught me

Think through task boundaries and architecture from the start. Decouple each capability as much as possible; do not make it too complex at once. I do not watch whether its code is pretty, but I must be able to point out where the business logic is wrong, to pull the strayed path back.

This 49-yuan screen can stay lit not because the first version was perfect, but because after failing three times, I finally agreed to write "auth expires, computers sleep" into the UI.

"Having ideas matters; planning matters more."

When you do not know what AI can do, just start doing it. Only through continuous practice can you learn bit by bit.

END

I am Old Winter Melon Daniel, sharing AI observations and tinkering notes. If you found this useful, likes, "Wow" and shares are welcome.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry