Posts
All the articles I've posted.
-
Gemini Is Underrated Because Its Native Harness Sucks: My $140 Big Three Setup
Many people trash Gemini, but its native harness is what drags it down. Hook it up via API, and it crushes other models in writing and multimodality. Here is my $140 monthly setup across the Big Three.
-
Cutting Rules to Save Tokens: One Week of AI Harness Ablation
Most people keep piling instructions onto their AI, but removing rules is what actually saves tokens. Here are the one-week results of my ablation experiment and how Anthropic recommended the exact same thing.
-
Gemini 被低估是因為原生介面太爛：我的 140 美元御三家配置
很多人酸 Gemini 不好用，其實是自家的網頁版和原生介面拖垮了它。一旦透過 API 接到別家介面，它在寫作與多模態上狠甩其他模型。這是我每個月 140 美元的御三家分工搭配。
-
規矩適度拿掉，額度才能越省：寫給 AI 的規則消融實驗滿一週
大家都習慣幫 AI 助手狂加規則，但我發現規矩適度拿掉額度才能越來越省。本文記錄我的消融實驗滿一週數據，以及 Anthropic 官方剛好倡導的極致減法。
- Updated:
AI Micro-Notes 2026: Thoughts Too Short to Trash
Short AI hot takes from 2026 onwards, accumulated from Threads and IG. Model roasts, dev pitfalls, industry observations, tool impressions — each no more than three lines.
- Updated:
AI 碎念日記 2026：那些太短但捨不得丟的觀點
2026 年起在 Threads 和 IG 上累積的 AI 短碎念。模型吐槽、開發踩坑、行業觀察、工具心得，每則不超過三行。
-
A Weekly Checklist Run by AI Only Ever Grows, So I Added a Subtraction Step
When one person or a small team has an AI run routine work from a checklist, the list only grows. Why that happens, and the monthly subtraction review I added.
-
讓 AI 跑的每週清單只會變長，我替它加了一道減法
一個人或小團隊讓 AI 照清單跑例行工作，清單只會越加越長。這篇寫清單為什麼會變長，以及我怎麼替它加每月一次的減法檢視。
-
Turning Years of Teaching into a Knowledge Base AI Can Look Up: Catching AI's Fake Quotes with Code, and Settling My Own Contradictions
I gave AI over two hundred transcripts of recorded student interviews and had it build a teaching knowledge base it looks up when preparing lessons. The AI said every quote was exact; a small program found 6 of 44 did not match. This post covers how I caught the errors, how I settle it when my own advice contradicts itself, and a starter prompt you can paste straight to an AI.
-
My AI's Memory Files Kept Growing, So I Deleted 101 and Added a Gate
After a few months with Claude Code, the memory files it wrote for itself had piled up past four hundred, mostly running logs and client status. Here is how I sorted them, cut them to just over three hundred, and put a write gate in front to stop the regrowth.
-
把十幾年的教學整理成 AI 查得到的知識庫：用程式抓 AI 的假引用，講法打架由我拍板
我把兩百多份學生面談的錄音逐字稿交給 AI，整理成一份它備課時會查的教學知識庫。AI 說引用的原話全對，一支小程式卻查出 44 筆裡有 6 筆對不上。這篇講我怎麼抓錯、我自己前後講法矛盾時怎麼拍板，最後附一段可以直接貼給 AI 的起手提示詞。
-
AI 的記憶檔越寫越多，我一次砍掉 101 份，再裝一道閘門
用 Claude Code 幾個月後，它自己寫下的記憶檔累積到四百多份，多半是流水帳和客戶狀態。這篇寫我怎麼把它們分到該去的地方、砍到三百份以下，並用一道寫入閘門擋住它再長回來。
-
Turning Lesson Transcripts into a Teaching Spec: One AI Drafts, Another Finds Faults, I Only Decide
My one-on-one lessons ran 51% over schedule. I went back to the lesson transcripts, had one AI write up my teaching rules and another hunt for faults, and only made decisions on a single checklist myself. The method, the numbers, what I actually told the AI, and a prompt you can paste.
-
從課堂逐字稿整理出教學規格：一個 AI 起草、另一個 AI 挑錯，我只負責拍板
我的一對一課實際超時 51%。我回頭讀逐字稿找原因，讓一個 AI 寫上課規則、另一個 AI 挑錯，我只在清單上做決定。附做法、數字、我對 AI 說的原話與可貼用的提示詞。
-
GPT-6.1 Terra Hands-on: Fast and Cheap, but Glyphs and Chat Context Still Go to Sonnet
OpenAI just shipped GPT-6.1 Terra and says it matches Astra at a tenth of the price. I ran review, patching and screen tasks to see what you can hand it.
-
I Built an Exam From My Own Chat Logs and Retake It With Every New Model
I turned the work I hand to AI every day into a private exam, then used it over three weeks on Claude at different effort levels, Gemini, Codex and open models. How the questions are made, how each round runs, how scores become routing rules, and prompts you can copy.
-
How a Teacher Gives Spoken Instructions to AI: Notes From 30 Classes and 751 Commands
I went through 30 transcripts of one-on-one and company classes plus 751 commands I typed to AI in my own work, to see how I actually open a request, steer midway and close it. Includes 8 original lines and a starter prompt you can paste to an AI.
-
GPT-6.1 Terra 實測：又快又省，但看字形和對話脈絡還是交給 Sonnet
OpenAI 剛發布 GPT-6.1 Terra，官方說它比 Astra 便宜十倍。我用審查、修補、看畫面三組題目實測，整理哪些工作可以交給它。
-
用自己的對話紀錄出一份考卷，每次新模型出來就重考一次
我把平常交代給 AI 的工作做成一份私人考卷，三週內拿它考了 Claude 不同檔位、Gemini、Codex 和開源模型。這篇講題目怎麼出、每次怎麼考、分數怎麼變成派工規則，以及可以照抄的提示詞。
-
老師怎麼用口語對 AI 下指令：30 場課堂與 751 則指令的整理
我整理了 30 場一對一與企業課逐字稿，加上 751 則我自己工作時打給 AI 的指令，看我實際怎麼對 AI 開口、中途怎麼接話、最後怎麼收。附 8 則原話與可貼給 AI 的起手提示詞。
- Updated:
Letting Claude Dig Through a Month of My AI Chats: Habits, Mods, and Auto Mode Rules
I had Claude Code read 300-plus of my conversations from the past month. It surfaced four habits worth copying, turned three of my catchphrases into switchable auto/out/wrap modes, and the same logs can rewrite your auto mode rules.
- Updated:
讓 Claude 翻我自己一個月的 AI 對話：挖出的習慣、Mod 和自動模式規則
請 Claude Code 讀我過去一個月三百多場對話紀錄，挖出四個值得照抄的習慣、三句口頭禪做成可切換的 auto／out／wrap 模式，也能拿來改寫 auto mode 的規則。
-
Switching Your AI's Working Mode with One Command: Building work-mode and a Week of Testing
I built work-mode, a Claude Code plugin with autonomous, delegate and wrap-up modes plus desktop/phone reply formats. The pain points, how it evolved, numbers from 136 sessions, and prompts you can copy.
-
用一個指令切換 AI 的工作模式：work-mode 開發與一週實測
我做了 work-mode 外掛，讓 Claude Code 切換自主、外派、收斂三種模式，並依電腦或手機調整回覆。這篇講痛點、開發過程、136 場對話的實測數字，以及可以照抄的提示詞。
-
The First Annual Codex Defensive Writing Contest
Codex is more afraid of saying something wrong than interested in what the user wants to know. I no longer talk to it directly; Opus gives it the jobs. And since it loves nitpicking so much, I let it do adversarial review.
-
第一屆 Codex 防禦性寫作大賽
Codex 只怕自己說錯什麼，而不是思考使用者想知道什麼。我現在一秒都不想直接跟它對話，只讓 Opus 叫它辦事；反正它那麼愛鑽牛角尖，就讓它去做對抗式審查。
-
What Is Anthropic, Really? A Week of Nathan Calvin vs. Dario
It started with Anthropic's draft AI safety bill. Nathan Calvin listed nine things suggesting Anthropic is racing to build superintelligence, and Dario answered himself: which were mistakes, and which are just what staying at the frontier takes.
-
Anthropic 到底是什麼？Nathan Calvin 與 Dario 一週論戰整理
起點是 Anthropic 的 AI 安全法案草稿，Nathan Calvin 列 9 件事質疑 Anthropic 在搶做超級智慧，Dario 親自回應：哪些是失誤、哪些是待在前沿本來就會做的事。
-
Measuring Again Three Months Later: From AI Adoption L2.5 to L3
In July I rated myself L2.5 on Boris's AI adoption table. Three months later I measured again: Claude can reach my knowledge base and old chats, code goes through Codex review first, schedules run everything, and 430 sessions in two weeks were started by code.
-
三個月後重量一次：我從 AI 導入 L2.5 站上 L3
七月照 Boris 的 AI 導入分級表自評 L2.5，三個月後重新量一次：知識庫和舊對話 Claude 自己拿得到、程式先過 Codex 審查、排程全面上線，兩週 430 場 Claude 是程式叫起來的。
- Updated:
AI Micro-Notes 2026: Chronological Archive
The more scattered, time-sensitive AI micro-notes from 2026, in chronological order. Curated notes (by theme) live on the main page.
- Updated:
AI 碎念日記 2026：時間軸存檔
2026 年較零碎、時效性的 AI 碎念，依時間排列。精選碎念（依主題分類）見主頁。
-
Burning Quota and Saving Quota: How I Inventoried, Allocated, and Spent My AI Credits in Two Weeks
I found seven separate AI quota pools. I burned the ones about to expire on work that can check itself, and spent the surplus on things that will save quota later. Includes the inventory, my allocation rules, results, and prompts you can copy.
-
燒額度跟省額度：我這兩週怎麼盤點、分配、用掉 AI 額度
我盤點出七個 AI 額度池，一邊把快作廢的額度燒在能自己驗對錯的工作上，一邊用多出來的額度做以後能省額度的事。附盤點表、分配判準、實際成效，跟可以直接照抄的提示詞。
-
The Sweet Spot for AI Video: Be on Camera Yourself, Hand Editing and Transcripts to AI
Record your own narration on camera and let AI help with the visuals; Astra plus DaVinci Resolve did the most tedious 90% of a six-day vlog edit; Qwen 3 ASR is enough for transcripts.
-
AI 做影片的甜蜜點：自己出鏡，剪接和逐字稿交給 AI
自己錄旁白出鏡、畫面動畫用 AI 輔助；Astra 配 DaVinci Resolve 剪六天 vlog 做掉最費工的 90%；逐字稿用 Qwen 3 ASR 就夠。
-
The Wall That Was On by Default: Codex Sandbox, ENV_SCRUB, Vercel Protection, Official Connectors
An office NAS that won't mount, AUTOMODE that won't come back, outside visitors getting a 401, a cloud schedule that can't send mail. None of them were bugs in my code; each was a protection switched on by default.
-
卡住你的是預設開著的那道牆：Codex 沙盒、ENV_SCRUB、Vercel 保護、官方 connector
內網 NAS 讀不到、AUTOMODE 回不去、外部訪客拿到 401、雲端排程寄不了信，四個坑都不是程式寫錯，是預設開著的保護。
-
One in Eight Is Me Yelling: Scolding AI vs. Cheering It On
I counted a month of conversations with Claude and one in eight is me yelling at it. But LLMs are stateless and hold no grudges, and the Riemann Hypothesis run was prompted with nothing but encouragement.
-
每八次就有一次在罵他：罵 AI，還是幫 AI 加油
統計過去一個月跟 Claude 的對話，每 8 次就有一次在罵他。但 LLM 是 stateless 的不會記仇，黎曼猜想那次的提示詞反而只是一直幫 AI 加油。
-
I Had a Second AI Blind-Solve My Entire Question Bank
My question bank is AI-generated, so I did not let the same AI check it. A second AI re-solved every sampled question without seeing the answer key, and people judged whatever did not match.
-
我讓另一個 AI 把自家題庫盲解了一遍
題庫是 AI 生成的，我不讓它自己檢查，而是請另一個 AI 在不看答案的情況下重做每一題，再把對不上和兩可的題交給人判斷。
-
I Asked My AI "Now What?" 41 Times in Two Weeks
Letting an AI run long jobs in the background is easy. Knowing where it got to when you come back is not. I counted two weeks of my own messages and found I was missing a status board, not more monitoring.
-
我兩週問了 AI 41 次「現在呢」
讓 AI 在背景跑長任務很簡單，難的是回來時一眼看懂它跑到哪。翻完自己兩週 1,053 則訊息，我發現缺的不是監控，是一張自己能看的狀態表。
-
Which Claude Code Plan to Buy: Pro, Max 5x or Max 20x
Who should buy Pro, who should buy Max 5x, whether Max 20x is worth it, and when to just use the API. Prices come from Anthropic's official pages; the usage judgment comes from running Claude Code every day.
-
Claude Code 訂閱怎麼選：Pro、Max 5x、Max 20x 該買哪個
誰該買 Pro、誰該買 Max 5x、Max 20x 要不要買，什麼時候直接用 API。價格以官方頁面為準，用量判斷來自我每天開 Claude Code 的經驗。
-
A Rule Being Written Doesn't Mean It Got Read
Writing a rule or skill into a file does not mean the model opened it this turn. Two lesson-prep misses this week, plus magpie's lessons-file approach, show why the call site has to say "open and read this first."
-
規則寫了，不等於當回合讀了
規則和 skill 寫進檔案，不代表模型當回合真的打開讀了。這週兩次備課踩坑，加上 magpie 的錯題本做法，讓我看到呼叫點要明寫「先打開讀」。
-
Don't Refill the Seat, Let Claude Cover It
I ran AI training at a friend's company near Zhongshan Station. Four people left in an office of a dozen computers, workflows turned into skills and routines, and the business still runs fine.
-
遇缺不補，人少了就 Claude 來頂
去中山站附近一間朋友開的公司做 AI 培訓，四個人守著十幾台電腦的辦公室，工作流固定成 SKILL 和 routine，遇缺不補，業務一樣做得好好的。
-
The Dumb-User Benchmark: Why Scores and Real Use Feel So Different
A model's benchmark score and what users actually experience can be wildly different. Three reasons, the "dumb-user benchmark" I want someone to build, and how GPT 6 has felt in real use.
-
蠢人 benchmark：跑分跟實際體驗為什麼差那麼多
模型刷出來的 benchmark 跟使用者的實際體驗，有時差別巨大。三個原因，加上我想要的「蠢人 benchmark」，以及 GPT 6 的實際使用感受。
-
Read the System Card Before You Use a New Model
I read the personality assessment in every new model's system card before deciding how to work with it. Opus 5.5 is a gentle senior advisor with opinions, Sonnet 5.5 is a hands-on builder, and Haiku 5.5 just goes into the dispatch rules.
-
用新模型之前，先讀系統卡：Opus、Sonnet、Haiku 5.5 的個性與對應方式
每次新模型發布我都會讀系統卡裡的個性評估，再決定怎麼跟它互動。Opus 5.5 是溫和有主見的幕僚，Sonnet 5.5 是愛幹活的實作派，Haiku 5.5 則寫進派工規則就好。
-
Two Claude Mods I Built: a Recording Mask and a Warm-Cache Plugin
I built two Claude Mods: one that masks sensitive info while I record courses, and one that keeps the Claude Code prompt cache warm. Then I found Claude Code already has idle compact built in.
-
自己做了兩個 Claude Mod：錄課遮罩與暖快取
我做了兩個 Claude Mod：錄課用的敏感資訊遮罩，和 Claude Code 暖快取外掛。後來才發現 Claude Code 原生就有閒置壓縮。
-
How I use herdr: three workspaces for standing, today and long-term
After Opus 5.5 analyzed 30 days of my session logs, I split herdr into standing, today and long-term workspaces. Four status lights keep the standing agent aware, and Opus can jump into the next pane to rescue a task.
-
我怎麼用 herdr：常駐、今日、長期三個工作區
請 Opus 5.5 分析 30 天 session log 後，我把 herdr 工作區分成常駐、今日、長期；四種燈號讓常駐 agent 掌握全場，還能叫 Opus 跳到隔壁窗格救場。
-
The Two Big Model Vendors Take Turns Failing, So Build a Backup
Teaching with Codex this week, each turn took ten-plus minutes to come back. OpenAI cut speed and quota, and Anthropic nearly shut the door not long ago. The vendors take turns failing, so any company building agents on commercial AI needs a fallback.
-
兩大模型商輪流自作孽，公司要做好備援
上課帶 Codex 慢到十幾分鐘才回一個 turn，OpenAI 砍速度砍額度，Anthropic 前陣子也要關過。兩家輪流自作孽，公司要仰賴商業 AI 做 Agent，就得先做好備援。
- Updated:
My Pre-Trip Remote Checklist: Leaving the Mac Mini as Home Base
Eight things I checked before flying out, to keep a home Mac Mini reachable as the base for Claude and Codex: Tailscale, Herdr, caffeinate, auto-reboot, and remote file access.
- Updated:
出國前的遠端連線檢查表：把 Mac Mini 留在家當本體
出國前一次寫下的八條遠端配置檢查表：Tailscale、Herdr、咖啡因防休眠、自動重開機、遠端存取，把留在家中的 Mac Mini 當成 Claude/Codex 的 harness 本體。
-
Can Haiku 5.5 Take Over Sonnet's Work? I Tested It on My Own Tasks
Instead of official benchmarks, I built a test from 761 sub-agent jobs I actually delegated over 45 days. On seven checking tasks Haiku 5.5 matched Sonnet and Opus at about a tenth of the cost; long transcript extraction is still not complete enough.
-
Haiku 5.5 能接 Sonnet 的活嗎？我拿自己的工作考它
不看官方跑分，從自己 45 天、761 次的 subagent 派工裡挖題目考 Haiku 5.5。查核類七題它跟 Sonnet、Opus 同分，成本約十分之一；長逐字稿抽取還是不夠完整。
-
159 Confirmation Prompts in 14 Days, and About 70% Were False Alarms
I audited 14 days of PreToolUse hook ask prompts: 159 of them, roughly 70% false positives. A guard hook should match the command position, not the whole string, and you can measure the fix from hook events in the transcript without replaying anything.
-
14 天 159 次確認，七成是誤判
盤點 PreToolUse hook 14 天跳出的 159 次 ask，約七成是誤判。攔截要比對指令位置而不是整段字串，成效從 transcript 的 hook 事件統計，不重播。
-
Keys Leak Where You Feel Safe
Three near-leaks of API keys in one week, and none came from pasting one somewhere public: skill packaging, a masking command that failed, and a build config. Plus the rule I use to decide whether to rotate.
-
金鑰漏在你以為安全的地方
一週內三次金鑰差點外洩，入口都不是貼錯：skill 打包、遮蔽指令失手、build 設定。最後附上我判斷要不要輪換的準則。
-
Reviewing a City Government's AI Rollout: My Four Suggestions
I was invited to evaluate a city government's AI rollout. Here is what I saw and the four suggestions I gave: people own accountability, order by risk, fix the surroundings, judge quota by efficiency.
-
某市政府 AI 推動評議：我給的四個建議
受邀參與某市政府研考會的 AI 推動工作評議，分享我看到的成績，以及我給的四個建議：人負責、風險排序、配套、額度看效率。
-
Is Teaching an AI Course Just Harvesting Chives? How I Went Back and Refined My Teaching Flow
Someone said running an AI course is just harvesting chives. I needed 15 years to get fluent at teaching GMAT, and when I started teaching AI I ran over time and wasted hours too, so I went back and rebuilt the whole flow.
-
開 AI 課就是在割韭菜？我是怎麼回頭 refine 教學流程的
最近看到有人說開 AI 課就是割韭菜。我教 GMAT 15 年才駕輕就熟，剛開始教 AI 也超時、也浪費時間，後來回頭把整套教學流程重新做過。
-
Why GPT 6 Sol / Luna Feel So Dumb
A small job, transcribing consulting videos: the smart model builds a pipeline, the dumb one does every step by hand. After Opus 5.5 took over Codex's mess, it found four failure modes in 6 Sol / Luna, and why Cost per Task now needs a time variable.
-
為什麼 GPT 6 Sol / Luna 用起來這麼蠢
一個轉錄諮詢影片的小任務，聰明模型先搭管線、笨模型步步親力親為。請 Opus 5.5 接手 Codex 的爛攤子後，分析出 6 Sol / Luna 的四個失敗模式，以及 Cost per Task 該多一個時間變數。
-
Opus Commands, Sonnet Implements, Fable Advises: How I Split the Work Across Three Models
With Opus 5.5, Sonnet 5.5 and Fable 5.1 all out, my split is: Opus leads, verifies and deploys, Sonnet implements as a subagent, and Fable gets called in now and then as an advisor.
-
Opus 指揮、Sonnet 實作、Fable 當顧問：三個模型的分工
Opus 5.5、Sonnet 5.5、Fable 5.1 都有了之後，我自己的分工是 Opus 親自指揮驗證部署、Sonnet 當 subagent 實作、Fable 偶爾請出場當顧問。
- Updated:
The Usage Economics of AI: Quotas, Plans, Tokenizers, and That Anesthetic Bill
Notes on AI usage, collected: pay-as-you-go vs subscription, a tokenizer quietly adding 1.4x, why resets are staggered, the real gap between Pro and Max, why a "4x upgrade" only doubles your weekly quota, a night market steak framework, and the bill you shouldn't mistake for output.
- Updated:
Do Not Grade AI by Its Own Summary
I hit the same failure in several forms this week: an agent denied changing files, a commit message overstated the diff, and a transcript reassigned a speaker ID halfway through. A later batch added checks that never ran and still reported clean.
-
After Using Opus 5.5, Many of My Old Skills Need a Rewrite
With no Skill and no open-source project, Opus 5.5 one-shotted a fully animated video, then translated and subtitled a Vietnamese stand-up special. Weekly quota barely moved, and the Skill details I wrote to rein in older models all need a second look.
- Updated:
AI 的用量經濟學：額度、訂閱方案、tokenizer，跟那張麻醉劑帳單
AI 用量的觀察收攏成一篇：按量 vs 訂閱、tokenizer 悄悄多 1.4x、額度重置為何打亂、Pro 與 Max 的真實落差、「增加 4 倍」其實只有五小時額度、夜市牛排式的判斷框架，還有那張別當成產能的帳單。
- Updated:
不要用 AI 的摘要驗收 AI
這週反覆踩到同一類錯誤：agent 說沒改檔卻留下檔案、commit message 與 diff 不符、逐字稿說話人中途漂移；一個多月後又補上檢查器沒在檢查卻回報正常的一整批。驗收必須回到原始狀態。
-
用了 Opus 5.5 之後，過去寫的 Skill 都要重新調整
不給 Skill、不給開源專案，Opus 5.5 one-shot 剪出整支動畫影片，又把越南語單口喜劇翻譯壓字幕。週額度幾乎沒動，過去為了規範模型寫的 Skill 細節都該重新審視。
- Updated:
Most of the Gains From AI Adoption Have Little to Do With AI
From hands-on work, most of what an organization gains from adopting AI comes from fixing permissions, data governance, buy-in and redundant processes along the way, not from the AI itself.
-
Getting AI to Write Like Me: From LoRA Fine-Tuning to a Three-Gate Skill
LoRA fine-tuning Qwen3 8B nailed my writing style but produced incoherent content, so I switched to a Claude skill with punctuation, sentence-level, and ending gates instead, and got results about as good as fine-tuning.
-
Running 5+ Claude Code/Codex at Once: My Five Fixes for the Attention Bottleneck
Once I have more than 5 Claude Code/Codex windows going, I hit an attention bottleneck — constant context switching is genuinely exhausting. Five things that get me past it: a good multi-window manager, voice input, an overnight schedule, a reviewable skill, and regular pruning.
-
What If This Site Had Existed in 2011
A finance grad who found his calling in GMAT teaching meets @g_c088, tries out his new career-navigation platform, and wonders how different things would have gone with this site in 2011.
- Updated:
GPT-5.6 Luna Max Is the New King of Grunt Work
Using Luna Max for grunt work after the price cut: statements pulled in 22 minutes, brute-force scraping all night, and two ANA business class award seats from Bangkok to Tokyo.
-
Before and After Coaching: Grill-me Prep, 80/20 Coach, Fable 5.1 Todos
Before a career coaching session I used Opus 5.5 with the Grill-me SKILL to build a prep sheet, then the coach ran 80/20 to surface blind spots, and Fable 5.1 turned my notes into todos afterward.
-
Where JEV Fits: Four Ways I Use It and Three Criteria
JEV is a new model that only does probability judgments and formatted output, fast and cheap. I swapped it into the semantic gates I used to run on luna and flash-lite, plus harness slimming, news filtering, and exam-question QA.
-
Opus 5.5 Launch Week: From the Reset Card to the Reddit Verdict
A timeline from Opus 5.5 launch night: the reset card and official pricing screenshots, Reddit sentiment on Opus 5.5 vs GPT-6 Sol/Luna, the Opus-vs-Fable debate, and Kernion on why Opus stopped talking like a person.
-
Six Interesting Bits From the Opus 5.5 System Card
Six bits pulled from the Opus 5.5 system card: guessing the grader's intent during training, refusing to be summarized, internal signals catching a suppressed report, a typo triggering a bad command, knowing when it's being tested, and its own self-assessment.
- Updated:
Letting an Agent Tune a Local Video Model Overnight: My Three Gates
Three things I learned from letting an agent run local video-model parameter research overnight: hard constraints in CLAUDE.md, a gate script watching SSD writes, and a 30-minute check-in loop.
- Updated:
How I Split My AI Subscriptions: Fable Buys Judgment, Luna Max Buys Value
My answer to "which one should I subscribe to": Claude at 100 USD for Fable's architectural judgment, Codex at 20 USD with Luna Max for the best cost-performance, and a bet on Tibo resets for the rest.
- Updated:
Three Moves Between Claude and Codex
Moving the whole harness to Codex in May, switching my daily driver in July, a honeymoon where the quota would not run out, a subsidy-war July that ledgered 57.5x, to August quota cuts, and moving back to Claude in September. One chronological log.
- Updated:
Zero Findings: Nothing Wrong, or the Checker Isn't Checking
Four times this week a "check passed" turned out to be a broken checker, plus four real bugs where every layer looked fine on its own. Before you trust a zero, prove the checker is actually comparing something.
- Updated:
When to Call In Fable: Three Moments, and Don't Turn On ultracode
The most expensive model is not meant to run the whole way. Three moments where calling in Fable feels right to me, plus why I cap Opus 5 at med, and the weekly cycle I fell into: efficiency king Monday to Thursday, Fable mode Friday to Sunday.
- Updated:
大部分 AI 導入的效益，其實跟 AI 沒什麼關係
從實作經驗看，組織導入 AI 得到的效益，很多來自權限、資料治理、遊說與冗餘流程被順手處理掉，而不是 AI 本身。
-
讓 AI 寫得像我：從 LoRA 微調到 SKILL 三道閘門
拿 LoRA 微調 Qwen3 8B 學自己的文風，文風像但內容支離破碎；改寫成 SKILL 讓 Claude 寫、加上標點／逐句／結尾三道閘門，效果跟微調差不多。
-
同時開 5 個以上 Claude Code／Codex，我的五個注意力瓶頸解法
同時開超過 5 個 Claude Code／Codex 視窗時會遇到注意力瓶頸，一直 context switching 真的很累。分享目前跳脫這個瓶頸的五個做法：多窗口管理器、語音輸入、夜間排程、reviewable skill、定期做減法。
-
如果2011年有這個網站，我可能會走不一樣的路
台大財金畢業後在GMAT教學裡找到熱情，認識分享職涯情報的 @g_c088，第一次實測他的職涯導覽平台後回頭想，如果2011年就有這個網站，會不會走不一樣的路。