jianshuo/claude-skills / wjs-x-improving-content
wjs-x-improving-content
VerifiedUse when 王建硕 wants to systematically improve his X (Twitter) content by iterating on the content-generation prompt (prompts/x/prompt.md, used by the every-6h tweet Action) and finding which prompt version produces the highest-reach tweets. Each prompt edit is a git-SHA-versioned, numbered experiment with a hypothesis; tweets are attributed to the version live at post time and judged on median impressions per tweet. Also mines per-tweet impression data for content-feature signals (angle / length / topic) that feed the next prompt edit. North-star = impressions per tweet. Triggers — "改 X 的 prompt", "X 内容改进", "哪版 prompt 最好", "什么内容 impression 高", "improve my tweets", "A/B test the X prompt", "/wjs-x-improving-content".
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 6.2 KB |
| inbox/.gitignore | other | 12 B |
| scripts/_common.pyexecutable | script | 1.5 KB |
| scripts/analyze-content.pyexecutable | script | 4.1 KB |
| scripts/evaluate.pyexecutable | script | 3.1 KB |
| scripts/ingest-tweets.pyexecutable | script | 4.3 KB |
| scripts/ledger.pyexecutable | script | 2.9 KB |
| scripts/scoreboard.pyexecutable | script | 3.7 KB |
| state/.gitignore | other | 6 B |
| state/SCOREBOARD.md | markdown | 1.1 KB |
| state/tweets.jsonl | other | 42.1 KB |
| state/versions.jsonl | other | 523 B |
state/tweets.jsonl
View on GitHub{"tweet_id": "2052061432964169934", "date": "2026-05-06", "impressions": 390, "engagements": 15, "likes": 2, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 239, "text": "AI feedback in serif type reads like a letter, not a chat bubble. That's intentional. Cathier uses Instrument Serif for all AI reflections — not aesthetics, but pacing. Same words, 40% longer dwell time in testing. Type is rhythm. (App Store: search Cathier) https://t.co/sytoz8c9fj", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 24, "mature": true}
{"tweet_id": "2053147005149712768", "date": "2026-05-09", "impressions": 483, "engagements": 25, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 231, "text": "Just shipped /jianshuo-wechat-mp-publish — Claude Code skill for 微信公众号 articles. → light polish, never overwrites your voice → auto cover (2.45:1) + illustration via gpt-image-2 → rich-text clipboard publish helper, ~2 min to ship https://t.co/UU6Xjhw5Dv #ClaudeCode", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 21, "mature": true}
{"tweet_id": "2053146611552051266", "date": "2026-05-09", "impressions": 496, "engagements": 28, "likes": 4, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 236, "text": "Just shipped /translate-video — end-to-end video localization. → Whisper transcribe → translated SRT (zh / en first-class) → time-aligned TTS dub: Volcano 豆包 + edge-tts neural → burn-in, original-audio bed, mouth-movement diarization https://t.co/hgfZ81TAto #ClaudeCode", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 21, "mature": true}
{"tweet_id": "2053142836489630156", "date": "2026-05-09", "impressions": 361, "engagements": 6, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 230, "text": "Just shipped /multicam — a Claude Code skill for multi-cam footage. → audio sync via envelope cross-correlation (handles low SNR + clock drift) → director-style auto-edit (Viterbi over per-second scores) → PiP, virtual crop-zoom, partial-coverage sources #ClaudeCode", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 21, "mature": true}
{"tweet_id": "2053491380497031580", "date": "2026-05-10", "impressions": 630, "engagements": 30, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 171, "text": "translate-video skill 更新 — 加了三道用户确认门, 防止误触烧 CPU/烧 API: 🛑 烧入字幕前: 先渲 30s 预览 + 抽帧 + 你确认字号才跑全片 🛑 全片配音前: 强制 sample 3-4 个 voice 让你挑 🛑 承诺 Volcano 声音前: 5 词 smoke test, 401 直接提示去开通 https://t.co/hgfZ81TAto #ClaudeCode", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 20, "mature": true}
{"tweet_id": "2053490579275964659", "date": "2026-05-10", "impressions": 1854, "engagements": 85, "likes": 5, "replies": 2, "reposts": 1, "new_follows": 3, "char_len": 164, "text": "微信公众号 skill 更新了 — 修了一个会让题图横向文字被裁掉的 bug。 之前 SKILL.md 和 cover-prompt.md 把比例写成 2.45:1, 但脚本实际裁到 900×383 (= 2.35:1)。AI 按 2.45 构图, 裁剪时左右各被切掉一截。 9 处统一为 2.35:1。 https://t.co/UU6Xjhw5Dv #ClaudeCode", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 20, "mature": true}
{"tweet_id": "2053374245045207396", "date": "2026-05-10", "impressions": 1833, "engagements": 47, "likes": 7, "replies": 1, "reposts": 2, "new_follows": 2, "char_len": 68, "text": "AI 时代最反直觉的一件事:要学的越多,越应该学得越少。 不是少学,是少而精——选一两个真正想解决的问题,沿着问题学。 其余的,让 AI 帮你查。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 20, "mature": true}
{"tweet_id": "2054570913618964511", "date": "2026-05-13", "impressions": 7236, "engagements": 204, "likes": 10, "replies": 2, "reposts": 0, "new_follows": 1, "char_len": 161, "text": "机器翻译进中文,处处都是「这个」「那个」「那份」「那种」。 不是翻译错了 —— 中文是 high-context 语言,源语言里多数指代都该删掉。 「那份能量」→ 「能量」 「正是在这合一里」→ 「合一中」 规则写进 prompt,AI 就照着改。 源码:https://t.co/ViA4q9XAG9 https://t.co/D1tnyNVceO", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 17, "mature": true}
{"tweet_id": "2054569776358490216", "date": "2026-05-13", "impressions": 28819, "engagements": 512, "likes": 14, "replies": 2, "reposts": 3, "new_follows": 9, "char_len": 200, "text": "Whisper API 直接吐的 SRT,几乎不能用。 两个失败模式: - 30 秒一大块字幕,没人读得完 - 安静段循环幻觉「你很难的」× 50 修法:response_format=verbose_json + timestamp_granularities[]=word,自己拼 cue。 别让不懂你需求的工具替你做边界决定。 源码:https://t.co/dKQ0jmi3ha https://t.co/Hk5S8iOhf5", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 17, "mature": true}
{"tweet_id": "2056396303110992319", "date": "2026-05-18", "impressions": 587, "engagements": 2, "likes": 0, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 23, "text": "https://t.co/Raznaxn5uv", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 12, "mature": true}
{"tweet_id": "2056395932019875897", "date": "2026-05-18", "impressions": 117, "engagements": 0, "likes": 0, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 94, "text": "@garrytan Book Mirror is a great idea. I just did the mirroring - suddenly changes any book into interesting one.", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 12, "mature": true}
{"tweet_id": "2056271940906787011", "date": "2026-05-18", "impressions": 436, "engagements": 7, "likes": 0, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 63, "text": "每天写一段 prompt,跟每天画一幅画是一回事。 不是非要写得多漂亮,就是让今天的句子跟昨天有点不一样。 不是变聪明,是手感来了。", "slug": "2026-05-18-mo-bi-tou", "angle": "B", "source": "bot", "prompt_sha": null, "age_days": 12, "mature": true}
{"tweet_id": "2057454735654109364", "date": "2026-05-21", "impressions": 2770, "engagements": 74, "likes": 7, "replies": 1, "reposts": 2, "new_follows": 1, "char_len": 97, "text": "我半年前是这个观点,现在我确认只会用现代编程语言(Python,C等)的程序员已经全部失业了,只是他们的老板们还没有发现。这个过程已经完结了,是瞬间的事情,只是需要一个渐进的过程才会被人意识到。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 9, "mature": true}
{"tweet_id": "2057839010811388286", "date": "2026-05-22", "impressions": 917, "engagements": 109, "likes": 5, "replies": 0, "reposts": 0, "new_follows": 1, "char_len": 72, "text": "装上试试,看看你天天在用的工具,背地里都说了些什么。 开源:https://t.co/99thvxcoqh,欢迎翻源码、提 issue、一起来贡献。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057838998056476701", "date": "2026-05-22", "impressions": 1005, "engagements": 53, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 82, "text": "response 标签:token 的账一清二楚。这次输入 2 个,缓存创建 53940 个,输出 116 个。 看得见,才谈得上理解。 https://t.co/SXEgxPzFtX", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057838984588673026", "date": "2026-05-22", "impressions": 467, "engagements": 31, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 84, "text": "tools 标签:这一次普通请求就挂了 48 个工具。翻源码,光工具目录下就有 43 个,主体全是 Anthropic 写好的内置工具。 https://t.co/ezUhsqKrjK", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057838959649264126", "date": "2026-05-22", "impressions": 521, "engagements": 45, "likes": 1, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 104, "text": "messages 标签:真正发过去的对话。你会看到一堆 <system-reminder>——这些是 Claude Code 自己塞进去的上下文,不是你打的字。 https://t.co/DZ6LcvbOcy", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057838946521154028", "date": "2026-05-22", "impressions": 707, "engagements": 100, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 134, "text": "system 标签:发给模型的系统提示词。第一句就是「You are Claude Code, Anthropic's official CLI for Claude.」,后面一长串行为约束,连 billing header 里的版本号都看得见。 https://t.co/0grvhjj8dX", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057838930318496101", "date": "2026-05-22", "impressions": 765, "engagements": 9, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 93, "text": "零依赖,两行就跑起来: npm install -g ccglass ccglass 它会问你看哪个 client——Claude Code / Codex / Kimi。也可以直接 ccglass claude。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057838920243769664", "date": "2026-05-22", "impressions": 8650, "engagements": 697, "likes": 45, "replies": 3, "reposts": 7, "new_follows": 12, "char_len": 130, "text": "Claude Code 从 node 升级成二进制以后,claude-trace 那些工具全失效了——再也看不到它到底给大模型发了什么。 我自己写了一个:ccglass。npm 装上、跑一行,它发给服务器的每个细节都摊在你眼前。 https://t.co/h4SPniy8Zi", "slug": "2026-05-22-ccglass", "angle": "thread", "source": "bot", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057837546537918880", "date": "2026-05-22", "impressions": 713, "engagements": 7, "likes": 3, "replies": 0, "reposts": 1, "new_follows": 0, "char_len": 79, "text": "全世界用的是同一个模型——版本号一样,就没有一丝区别。 人和人的差别,全在这一大坨随身带着的信息里。它越厚,越精确地刻画出你是谁。 模型是同一个。那本书,才是你。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057837533237801160", "date": "2026-05-22", "impressions": 740, "engagements": 4, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 106, "text": "你用 Claude Code 根本不是在聊天。 你是在借一大批工具、和你自己都没意识到的技能,替你垫背景。就像写 C 语言:你只敲一行,一编译,几十万行库代码跟着一起进去。 这,就是 ChatGPT 和 Claude Code 的区别。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057837520235442378", "date": "2026-05-22", "impressions": 435, "engagements": 1, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 116, "text": "而且这还只是没上下文时。多聊几轮,你的输入、它的输出、每次工具调用的结果、所有中间变量,都一层层往上叠。 Claude Code 的上下文是一百万词元,一部《三国演义》才七十五万。 用不了多久,你每打一个词,它收到的就是一整本《三国演义》。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057837504636854325", "date": "2026-05-22", "impressions": 371, "engagements": 1, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 39, "text": "好消息是:读过的部分会被缓存。同一本书第二次读,花的时间和钱,都远小于第一次。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057837491605176802", "date": "2026-05-22", "impressions": 493, "engagements": 2, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 87, "text": "为什么这么干?因为它没有「昨天」。 它每次睁眼都是失忆的,上一句说完就忘干净。所以你每递一句话,系统都得把它该知道的一切重塞一遍。 它不是在「记得」,是每一次都重新读一遍这本书。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057837476388221095", "date": "2026-05-22", "impressions": 505, "engagements": 4, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 100, "text": "剩下的十三万多,全是它每次随身背着的东西:41 个工具的完整说明书(占六成)、四百多条技能清单、一份记忆、几个插件须知、一段「你是谁、该怎么做事」的身份说明。 你打一个 hello,这一整套跟着一起发过去。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2057837459879395455", "date": "2026-05-22", "impressions": 4767, "engagements": 250, "likes": 12, "replies": 3, "reposts": 1, "new_follows": 3, "char_len": 120, "text": "用 Claude Code,在没有上下文时给它打一个词:hello。 你猜它实际发出去多少? 我把这次请求抓下来一看:137918 个字节,十三万多字符,一本书那么厚。 我打的那个 hello,只占 5 个字符。 https://t.co/RFYUN59oKd", "slug": "2026-05-21-hello-is-a-book", "angle": "thread", "source": "bot", "prompt_sha": null, "age_days": 8, "mature": true}
{"tweet_id": "2058224761252299157", "date": "2026-05-23", "impressions": 21277, "engagements": 192, "likes": 24, "replies": 1, "reposts": 1, "new_follows": 4, "char_len": 149, "text": "2020年11月 — 蚂蚁集团上市叫停 2021年4月 — 阿里反垄断罚款182.28亿元 2021年6-12月 — 滴滴上市、App下架、退市 2021年7月 — 教培\"双减\" 2021年8月 — 游戏防沉迷新规 2021年10月 — 美团反垄断罚款34.42亿元 2021年12月 — HFCAA落地与\"预摘牌名单\" 2021年是个拐点", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 7, "mature": true}
{"tweet_id": "2058217502656938444", "date": "2026-05-23", "impressions": 4893, "engagements": 396, "likes": 54, "replies": 0, "reposts": 6, "new_follows": 4, "char_len": 136, "text": "写代码的 agent 越来越聪明,也越来越像黑箱。ccglass 就是想给这个黑箱开一扇玻璃窗——现在能看清它每一步是怎么动的了。 0.3.0 新增:flow 时间线(看它怎么挑工具、跑、再喂回模型)、请求全文导出、agent 还能查自己。 https://t.co/99thvxcoqh", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 7, "mature": true}
{"tweet_id": "2058216918486778164", "date": "2026-05-23", "impressions": 2053, "engagements": 3, "likes": 0, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 20, "text": "我肯定还是要保持Claude Code满额", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 7, "mature": true}
{"tweet_id": "2058207511497601311", "date": "2026-05-23", "impressions": 6199, "engagements": 56, "likes": 3, "replies": 4, "reposts": 0, "new_follows": 1, "char_len": 118, "text": "好多人都在问我 Claude Code和Codex到底哪个更强?似乎这是关心AI的人的必问题。我一直说,我不知道。我不是自媒体行业的,工具性能测试不是我的工作。我只用Claude Code,把所有的细节都用清楚,虽然我不能保证它是最好的工具。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 7, "mature": true}
{"tweet_id": "2058206855894294879", "date": "2026-05-23", "impressions": 333474, "engagements": 54837, "likes": 489, "replies": 57, "reposts": 40, "new_follows": 193, "char_len": 79, "text": "过了五年, 才那么清晰的看到2021年,中国到底发生了什么。。。。一件接着一件,一条接着一条,没有一个季度停息过。 https://t.co/15liLOwnLO", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 7, "mature": true}
{"tweet_id": "2058007586872737852", "date": "2026-05-23", "impressions": 1044, "engagements": 37, "likes": 5, "replies": 1, "reposts": 0, "new_follows": 1, "char_len": 34, "text": "@DashHuang 对呀,原来其实就是体力不足,现在有了AI体力满血", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 7, "mature": true}
{"tweet_id": "2058005978147418119", "date": "2026-05-23", "impressions": 21281, "engagements": 1583, "likes": 150, "replies": 12, "reposts": 21, "new_follows": 39, "char_len": 101, "text": "看得见,才谈得上理解。我写了个 ccglass,npm 装上就能看到 Claude Code 背地里给大模型发了什么——系统提示词、48 个工具、token 账,全摊开。https://t.co/99thvxcoqh", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 7, "mature": true}
{"tweet_id": "2058489572057813283", "date": "2026-05-24", "impressions": 3278, "engagements": 100, "likes": 21, "replies": 2, "reposts": 0, "new_follows": 1, "char_len": 91, "text": "大家都算错了一笔账:需求不变,除以提高的效率,得出需要更少的人。可历史上每一次都反过来——煤更高效了,用量涨了十几倍;程序员从五十万涨到八百万。被 AI 冲击的行业,才是该冲进去的行业。", "slug": "efficiency-demand-explosion", "angle": "queue", "source": "bot", "prompt_sha": null, "age_days": 6, "mature": true}
{"tweet_id": "2058482836391522381", "date": "2026-05-24", "impressions": 1471, "engagements": 29, "likes": 3, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 38, "text": "现在北京每一辆公交车上都有一个安全员,他们的工作比电梯工还少 - 他们只负责坐车", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 6, "mature": true}
{"tweet_id": "2058471444255064292", "date": "2026-05-24", "impressions": 1583, "engagements": 7, "likes": 3, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 23, "text": "的确任何东西如果有可能,从黑盒变成白盒都有价值", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 6, "mature": true}
{"tweet_id": "2058471220803436619", "date": "2026-05-24", "impressions": 218, "engagements": 2, "likes": 0, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 69, "text": "@BeauJohnson89 Thanks for picking it up. Yes. It started just to solve my own needs", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 6, "mature": true}
{"tweet_id": "2058425538352714042", "date": "2026-05-24", "impressions": 1702, "engagements": 6, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 34, "text": "Claude Trace 新版本claude code 没法用了,这个可以用", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 6, "mature": true}
{"tweet_id": "2059037913661485319", "date": "2026-05-25", "impressions": 4193, "engagements": 160, "likes": 10, "replies": 5, "reposts": 0, "new_follows": 5, "char_len": 87, "text": "我心里有个挺糙、但挺好用的标准:你每天自己手写、发给大语言模型的东西,如果还没到 1000 行,那真的,还没进入这个时代。这不是门槛,是个温度计——量的是你有没有真的把手伸进去。", "slug": "2026-05-24-relearn-to-walk", "angle": "A", "source": "bot", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2058941837159837928", "date": "2026-05-25", "impressions": 3167, "engagements": 25, "likes": 9, "replies": 1, "reposts": 0, "new_follows": 1, "char_len": 4, "text": "多谢分享", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2058941757501714758", "date": "2026-05-25", "impressions": 960, "engagements": 24, "likes": 8, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 94, "text": "不是「有用」,是「好像有用」。 大多数 AI 应用栽在这儿:像算命先生说的话,句句像在说你,其实对谁都成立——这就是伯纳姆效应。真有用才是市场价值,而能不能被验证,决定一个 AI 业务到底成不成立。", "slug": "2026-05-24-useful-vs-seems-useful", "angle": "B", "source": "bot", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2058916592541639037", "date": "2026-05-25", "impressions": 456, "engagements": 32, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 1, "char_len": 24, "text": "@DashHuang @caozlog 你在新西兰?", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2058906366878925213", "date": "2026-05-25", "impressions": 1285, "engagements": 26, "likes": 7, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 129, "text": "Karpathy 看 AI 写的代码,原话是「我会心脏病发作」。下一句却是:「但它能跑。」 汇编代码、前端编译出的 JS、Word 生成的 HTML、人类的 DNA、绕远路的喉返神经——全一个样:丑得要死,但能跑,还跑了几亿年。 AI 写的丑代码,其实挺符合自然规律。 —— 王建硕", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2058895299729605079", "date": "2026-05-25", "impressions": 560, "engagements": 26, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 227, "text": "Just open-sourced polysync 🎬🎧 Audio-sync multi-cam footage, then auto-cut to whoever's talking. • cross-correlation, handles mismatched mics • reversible sidecars, never re-encodes your originals • auto-edit + PiP pip install polysync https://t.co/Hgiut2zyfa", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2058878794455482569", "date": "2026-05-25", "impressions": 1576, "engagements": 46, "likes": 7, "replies": 4, "reposts": 0, "new_follows": 1, "char_len": 107, "text": "和 AI 交互,最该偷懒的不是打字快,而是根本不用打对。拼错、语法乱、顺序颠倒,大语言模型都不在乎——它会自己理解你的大致意思。所以语音输入和 Claude Code 几乎是绝配:一边想一边说,一股脑全告诉它,剩下的它意会。", "slug": "2026-05-25-holdspeak", "angle": "B", "source": "bot", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2058875125983039757", "date": "2026-05-25", "impressions": 2091, "engagements": 40, "likes": 11, "replies": 2, "reposts": 1, "new_follows": 0, "char_len": 109, "text": "有人问我未来是不是都是一人公司。我不信。 技术让个体变强,从来不是让协作消失,而是把协作推到一个新的量级上。一人公司不是终点,它只是路过的一个状态。要么壮大(小概率),要么死掉(大概率),不会一直维持在这种状态太长时间。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2058831922273648689", "date": "2026-05-25", "impressions": 6972, "engagements": 98, "likes": 9, "replies": 2, "reposts": 0, "new_follows": 1, "char_len": 106, "text": "我看任何可以跑的程序的汇编语言代码时,我看前端框架编译出来 Javascript 时,我看Microsoft Word生成的文档的 HTML 源代码时,都看到了: 臃肿、大量复制粘贴、笨拙的抽象、脆弱、真的恶心。 但它们都能跑", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2058790835005882751", "date": "2026-05-25", "impressions": 31305, "engagements": 536, "likes": 10, "replies": 2, "reposts": 0, "new_follows": 13, "char_len": 16, "text": "为啥大家都在这个时候活跃了起来?", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2058790709235527743", "date": "2026-05-25", "impressions": 906, "engagements": 14, "likes": 1, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 24, "text": "程序员现在看起来是第一大,因为结果太容易验证质量", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 5, "mature": true}
{"tweet_id": "2059394329316397191", "date": "2026-05-26", "impressions": 596, "engagements": 12, "likes": 5, "replies": 2, "reposts": 0, "new_follows": 0, "char_len": 73, "text": "不是 AI 把人吃掉了。是会用 AI 的人,把不会用 AI 的人吃掉了。 骑狼的方法不复杂:先去试一下骑。骑过之后你会发现,狼其实很听话,它只是没人骑过它而已。", "slug": "2026-05-18-ten-thousand-wolves", "angle": "A", "source": "bot", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059312528866578930", "date": "2026-05-26", "impressions": 415, "engagements": 4, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 71, "text": "我在尝试 Google Gemini Flow,得到了一段奇怪的视频,似乎这个驾驭层无法驾驭自己的身体 https://t.co/VO3geF4wAy", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059311594648310196", "date": "2026-05-26", "impressions": 840, "engagements": 9, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 146, "text": "确定性不是必然要求,其实让大语言模型有确定性分分钟的事,temp=0, seed确定,然后采样策略固定下来,绝对是啥输入绝对焊死出啥输出,分毫不差,但这已经不是大语言模型的优势了。大语言模型核心是一个有边界的行为,是在一个概率空间里面的。至于UI等等,没人说不能用图呀,Swift也得配JPG呀", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059303843708702924", "date": "2026-05-26", "impressions": 6036, "engagements": 119, "likes": 6, "replies": 1, "reposts": 1, "new_follows": 5, "char_len": 84, "text": "用 Python、C 写代码这一层,已经瞬间完结了,不是慢慢被替代。但它上面冒出了全新的一层:拿自然语言编程。措手不及的是老程序员,抓住机会的,反而是那些从不会写代码的人。", "slug": "programmers-instantly-unemployed", "angle": "queue", "source": "bot", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059275098901918115", "date": "2026-05-26", "impressions": 689, "engagements": 4, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 1, "char_len": 13, "text": "这倒是一个有趣的现象。。。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059273716950036582", "date": "2026-05-26", "impressions": 609, "engagements": 40, "likes": 1, "replies": 0, "reposts": 1, "new_follows": 0, "char_len": 23, "text": "https://t.co/axhQN8eWJH", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059272541894811838", "date": "2026-05-26", "impressions": 4511, "engagements": 59, "likes": 6, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 81, "text": "互联网早期老一代人特别鄙夷年轻人:就看到你画上网费也没见你挣到啥钱。 不过严肃的说起来,整个互联网都搞了七八年了,要不是 SP 业务,都还没有人能说明白这东西怎么挣钱。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059270434265428145", "date": "2026-05-26", "impressions": 1823, "engagements": 187, "likes": 8, "replies": 1, "reposts": 0, "new_follows": 1, "char_len": 23, "text": "https://t.co/fBdfHBw5Lr", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059263975267995753", "date": "2026-05-26", "impressions": 556, "engagements": 2, "likes": 0, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 15, "text": "至少得是个名词,而不是形容词吧", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059242757848490314", "date": "2026-05-26", "impressions": 147, "engagements": 1, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 142, "text": "@DashHuang @geekwu 以前在一个药物的研讨会上,有人问医生,这个药会不会有依赖性?医生是真么回答的:怎么说呢?你每天早上刷牙,刷了就舒服,不刷牙就不舒服,你说你对刷牙时有依赖性了吗?任何让我们的生活更好的事情,我们都应该会依赖的。我现在就很依赖碳水,依赖电,也很依赖衣服。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059228675854508466", "date": "2026-05-26", "impressions": 2346, "engagements": 26, "likes": 1, "replies": 5, "reposts": 0, "new_follows": 0, "char_len": 71, "text": "AI 到底怎么翻译比较好?本来电脑是个好词,但是已经被用过了,要么 「想器」,如同「计算器」一样,或者「模脑」?或者类似「手机」一样,叫「灵机」?", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059216926539419895", "date": "2026-05-26", "impressions": 842, "engagements": 17, "likes": 4, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 119, "text": "想换工作、想买新工具、想给产品推个新功能,都先过一关:过去 30 天里,有没有一个具体的人,被这件事真的卡住?这一关过不了,多半是你脑补了一个需求,然后给自己讲了个故事,讲得太顺,连自己都信了。我们要的是被卡住的人,不是被趋势带跑的我们自己。", "slug": "2026-05-18-trend-not-need", "angle": "B", "source": "bot", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059210587310371208", "date": "2026-05-26", "impressions": 671, "engagements": 58, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 23, "text": "https://t.co/d6RjyQU3Bj", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059197567456666012", "date": "2026-05-26", "impressions": 66983, "engagements": 1632, "likes": 18, "replies": 3, "reposts": 2, "new_follows": 15, "char_len": 23, "text": "https://t.co/vFwXyWMqc7", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059151884871663926", "date": "2026-05-26", "impressions": 1089, "engagements": 15, "likes": 5, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 131, "text": "肯定算动词呀,一个人学会vibe coding厉害死了,至于用claude code,codex,kimi code还是别的啥,没有那么重要。要是把这些名词都学一遍,也不见得vibe coding能做的多好。核心是把不同的东西后面不变的那个,而不是学会各个东西之间区别。", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059149367232397804", "date": "2026-05-26", "impressions": 12822, "engagements": 64, "likes": 5, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 80, "text": "人人都更喜欢学名词,因为名词太容易打勾了。可真正恒久、不被时间淘汰的,恰恰是那些永远打不了勾的动词。名词是潮水,一茬接一茬地涨落。动词是河床。AI 时代,学动词。", "slug": "2026-05-26-learn-verbs-not-nouns", "angle": "A", "source": "bot", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059100714962714638", "date": "2026-05-26", "impressions": 881, "engagements": 98, "likes": 7, "replies": 1, "reposts": 0, "new_follows": 1, "char_len": 23, "text": "https://t.co/jPdvkw0Qxf", "slug": null, "angle": null, "source": "manual", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059087498404569503", "date": "2026-05-26", "impressions": 1095, "engagements": 34, "likes": 4, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 129, "text": "丑得心脏病发作,但它能跑 Karpathy 自己看 AI 写的代码时,原话是:「我会心脏病发作。」 他给的形容词:臃肿、大量复制粘贴、笨拙的抽象、脆弱、真的恶心。 但他下一句紧接着是:「但它能跑。」 这就是 2026 年所有用 AI https://t.co/VspkzciMwX", "slug": "2026-05-25-ugly-but-runs", "angle": "fulltext", "source": "bot", "prompt_sha": null, "age_days": 4, "mature": true}
{"tweet_id": "2059572152626745739", "date": "2026-05-27", "impressions": 379, "engagements": 7, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 136, "text": "我本来想用「驾驶」替代「驾驭」这个词,「Drive」替代「Harness」,仔细品味了一下,恰恰是「驾驭」更形象的表达了AI的自主性。「驾驶」的宏观方向感来自人类,微观方向感也来自于人类;而「驾驭」,宏观的方向感来自于人类,微观的方向感来自于AI。这是这两个词的核心区别。", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059570658703773703", "date": "2026-05-27", "impressions": 785, "engagements": 8, "likes": 1, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 128, "text": "这个我还得好好想想 - 我原来学习画画的时候,一直记着一句话,「每个人身体里面的都有1000张烂画。赶紧把这1000张画完,剩下的都是好画了」。现在的skills,文档,约束,提示词,这些技巧都很初级,生成的的确不咋地,但希望先把这1000篇烂文章给输出完了吧", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059570002861392004", "date": "2026-05-27", "impressions": 331, "engagements": 13, "likes": 0, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 22, "text": "刚想填充一下profile,这下尴尬了。。。", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059545752633016511", "date": "2026-05-27", "impressions": 487, "engagements": 11, "likes": 1, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 58, "text": "表达清楚是需要有信息量的,就是需要有足够多的“字数”。一个大型软件工程可能需要几十万句话(行代码)才能讲清楚需求的。", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059514566607020455", "date": "2026-05-27", "impressions": 392, "engagements": 5, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 88, "text": "不是 AI 没法用,是你没告诉它「烤鸭」。你说「我饿了」,它只能给你一张所有人都适用的食物清单——那是这个世界对一个陌生提问者的平均答复。你要的是定制答案,那你就得先给它定制信息。", "slug": "2026-05-18-hungry-roast-duck", "angle": "C", "source": "bot", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059713195120492905", "date": "2026-05-27", "impressions": 318, "engagements": 3, "likes": 2, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 47, "text": "你今天打开 AI,是又开了一个空白文档,写完就关掉?还是在往一份已经几百页的手册里,再添上一页?", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059660168682762724", "date": "2026-05-27", "impressions": 1104, "engagements": 87, "likes": 0, "replies": 2, "reposts": 0, "new_follows": 0, "char_len": 55, "text": "苹果日历里面,1582年10月4号的第二天真的是10月15号 🤔🤔 https://t.co/2xzPofyQgy", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059641994465722571", "date": "2026-05-27", "impressions": 906, "engagements": 1, "likes": 1, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 7, "text": "这个交互好玩的", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059641080971399416", "date": "2026-05-27", "impressions": 29, "engagements": 2, "likes": 0, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 55, "text": "@JsonsP22174 的确有一些特定的模式,比如“我用不卖关子的方式给你讲”, “这样肯定稳”,瞬间就会被发现", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059630190750556615", "date": "2026-05-27", "impressions": 484, "engagements": 5, "likes": 2, "replies": 0, "reposts": 1, "new_follows": 0, "char_len": 38, "text": "凡是有用的做不出来,凡是做出来的没啥用。这不是谁没本事,是路本身还没走出来。", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059623621287657911", "date": "2026-05-27", "impressions": 717, "engagements": 9, "likes": 1, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 163, "text": "I see. Still use the browser use approach. It works but not as reliable. I never have a solid answer to whether content platform like x, substack should open post API to agent. Let's see.... Thanks.", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059622154271080585", "date": "2026-05-27", "impressions": 709, "engagements": 24, "likes": 7, "replies": 0, "reposts": 0, "new_follows": 1, "char_len": 31, "text": "你浪费掉的那点钱,远远不如你那颗不肯多想一步的脑子,浪费得多。", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059618682658287962", "date": "2026-05-27", "impressions": 671, "engagements": 12, "likes": 1, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 59, "text": "Substack don't have a reliable API. How can you do that? Manually post?", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059618565876240758", "date": "2026-05-27", "impressions": 481, "engagements": 4, "likes": 2, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 64, "text": "I don' t know whether the AI writing's quality problem is my illusion or real.", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059602633569325310", "date": "2026-05-27", "impressions": 748, "engagements": 5, "likes": 0, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 22, "text": "blog还好吧,没有所谓的blog味这一说呀", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059600799785980026", "date": "2026-05-27", "impressions": 337, "engagements": 8, "likes": 1, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 69, "text": "@Fenng 也不是,给我一篇文章,我还真的能从文字中间读出来那种AI感。我的一种解释是,好的内容本来就稀缺,原来找到就不容易,现在更不容易", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059590187911672000", "date": "2026-05-27", "impressions": 9226, "engagements": 92, "likes": 10, "replies": 11, "reposts": 0, "new_follows": 1, "char_len": 41, "text": "为啥AI写的文章就是读不下去呢?到底差别在哪里?人怎么就能区分出来人味和AI味呢?", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059589415765520841", "date": "2026-05-27", "impressions": 20455, "engagements": 866, "likes": 0, "replies": 1, "reposts": 0, "new_follows": 9, "char_len": 37, "text": "你离深刻只差 11 个字和一个句号了,加油!哈哈 但不得不承认我觉得你说的是对的", "slug": null, "angle": null, "source": "manual", "prompt_sha": "f6b47a3", "age_days": 3, "mature": true}
{"tweet_id": "2059805924303286343", "date": "2026-05-28", "impressions": 248, "engagements": 9, "likes": 2, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 43, "text": "学校禁 AI,和当年大学禁 Google 一模一样。如果这个技能不重要,那就让它不重要好了。", "slug": null, "angle": null, "source": "manual", "prompt_sha": "0f4be99", "age_days": 2, "mature": false}
{"tweet_id": "2059791174131655120", "date": "2026-05-28", "impressions": 322, "engagements": 7, "likes": 4, "replies": 1, "reposts": 0, "new_follows": 0, "char_len": 122, "text": "AI味可能就是太完美的味道:没有错别字,没有让读者不舒服,产生碰撞的棱角。太过于顺滑了,就没有阅读时候的碰撞感。几乎每一步都踩在合理的范围以内。如同看到一个过于完美的美女图我总怀疑是AI生成的:完美世界到处都是一样无聊的美,现实世界各有各的瑕疵", "slug": null, "angle": null, "source": "manual", "prompt_sha": "0f4be99", "age_days": 2, "mature": false}
{"tweet_id": "2059790330741113017", "date": "2026-05-28", "impressions": 50, "engagements": 0, "likes": 0, "replies": 0, "reposts": 0, "new_follows": 0, "char_len": 112, "text": "@duojiexi0224 可能就是太完美了,没有错别字,没有让自己不舒服,产生碰撞的棱角。太过于顺滑了,就没有阅读时候的碰撞感。如同看到一个过于完美的美女图我总怀疑是AI生成的:完美世界都是一样无聊的美,现实世界各有各的瑕疵", "slug": null, "angle": null, "source": "manual", "prompt_sha": "0f4be99", "age_days": 2, "mature": false}
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |