Tag: harness
All the articles with the tag "harness".
-
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.
-
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。
-
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.
-
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。
- 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。
-
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。
-
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」。
-
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。
-
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.
-
JEV 放哪裡：我的四個用法跟三個判斷標準
新模型 JEV 只做機率判斷與格式化輸出，速度快又省成本。我拿它換掉原本用 luna、flash lite 做的語意判斷閘門，也用在瘦身 harness、篩新聞、把關考題四個地方。
- Updated:
The Skills and Hooks I Use Every Day, Now Open Source
I open-sourced the Claude Code skills and hooks I actually use every day. No technical skill required, pure logic, built for knowledge workers with no coding background.
- Updated:
我每天在用的 Skill 跟 Hook，開源了
把我每天都在用、不需要技術力、純邏輯導向的幾個 Claude Code Skill 跟 Hook 開源出來，最適合跟我一樣沒有編程背景的知識工作者。
-
How to Write a Gate That Actually Blocks
The gate was running and stopped nothing. Four rules: bind it to a field that cannot survive a no-op, counter-test both directions before trusting a zero, exit non-zero or it is not a gate, and never treat wording the prompt never specified as a contract.
-
Proving Absence Is Much Harder Than Proving Presence
Four times this week my agent concluded something did not exist. The worst: it ruled that 16 emails were never sent, when they had all gone out days earlier.
-
怎麼寫一個真的會擋下來的閘門
閘門明明在跑，卻什麼都沒擋住。四條判準：綁 no-op 下活不下來的欄位、回 0 前先兩邊反驗、只印訊息不改離開碼等於沒閘門、別把 prompt 沒規定的字面當契約。
-
證明「沒有」比證明「有」難得多
這週我的 agent 連續四次下了「不存在」的結論，最嚴重的一次是判定 16 封信沒寄，結果信早就全數寄出。負向斷言要窮舉，而它只查了一個地方。
-
Hand the Dharma to Your Agent and Let It Discipline Itself
A SKILL called buddhist-method that constrains agent behavior with six Buddhist concepts: verify before asserting, look for root causes, re-check actual state, throw the draft away, cut the cause instead of hiding it, and don't change position under pressure.
- Updated:
The More Rules You Add, the Less Claude Listens — I Sent a Team of Agents to Trim My Setup and Cut 36% of Always-On Context
Late at night I got Opus 4.8, and my weekly quota happened to reset. The first thing I did was put my harness on a diet. A COO student who had burned through 88% of his 1M context made me face one thing squarely: the more rules you add, the less the model listens.
-
把佛法交給 agent，讓他用佛法約束自己
一個叫 buddhist-method 的 SKILL，用六條佛法概念約束 agent 的行為——查證、找根因、重新確認狀態、丟掉草稿、捨斷成因、不因壓力改立場。
- Updated:
規則越加，Claude 越不聽話——派一隊 AI 重整設定，常態上下文省 36%
半夜拿到 Opus 4.8、額度剛好重置，我做的第一件事是減肥我的 harness。一個 COO 學員把 1M 上下文用到 88% 的案例，讓我認真面對一件事：規則加越多，模型反而越不聽話。
-
My CLAUDE.md Boiled Down to Eight Lines of "Take Pride In…, Take Shame In…"
My CLAUDE.md cut down to eight paired maxims, plus a human-machine collaboration circular written deliberately in Party-document style, plus a sixteen-character guideline.
-
我的 CLAUDE.md 精煉到只剩八句「以…為榮，以…為恥」
把 CLAUDE.md 一路砍到只剩八組對句，外加一篇刻意寫成公文體的人機協同宣導文，還有十六字方針。
-
Six New Context Engineering Rules for Claude 5, and the 1,473 Lines I Cut
After reading Anthropic's context engineering guidance for Claude 5, I turned the key points into nine cards, then cut 1,473 lines from a harness I had built up over four model generations.
-
Fewer Prompts Is Not Less Control
Claude 5 and GPT-5.6 official guidance is converging on the same thing — retiring old-style prompt stacking. But trimming is not letting go; control just moves. And the payoff of maintaining that layering is switching tools without losing a step.
-
Claude 5 時代的情境工程六條新規則，與我砍掉的 1473 行 harness
讀完 Anthropic 對 Claude 5 的情境工程指南後，我把重點做成九張圖卡，然後照著把累積四個世代的 harness 削掉 1473 行。
-
更少的 prompt，不是更少的控制
Claude 5 與 GPT-5.6 的官方指南正在合流，都在淘汰舊式 prompt 堆疊。但精簡不等於放任，控制只是換了位置——而維護好這套分層的紅利，是換哪個工具都能立刻接手。
-
Two Traps Running Claude Code on Local Ollama: Truncated Context, and A3B Buckling Under Heavy Verification
Two things I logged: cco (Claude Code driven by local Ollama) had long been giving off-topic answers, and the root cause was not a weak model but a full harness whose system prompt had ballooned to 30-50k tokens and was being silently truncated; then I put A3B on the Mac mini for two days as a night worker.
-
Instead of Begging the Model Not to Lie, I Wrote a Hook That Stops It
The sequel to the Opus 4.8 confabulation post: I moved "don't make up numbers" from a plea in CLAUDE.md to a pending-guard hook that blocks git commit at PreToolUse.
-
Second Harness Diet: Global Skills From 58 Down to 40
A follow-up to the harness diet series: six months later, a bigger cleanup that cut my global skills from 58 down to 40, start to finish.
-
本機 Ollama 跑 Claude Code 的兩個坑：context 被截斷、A3B 撐不住重驗證
記錄兩件事：cco（本機 Ollama 驅動 Claude Code）長期答非所問，根因不是模型不夠聰明，是完整 harness 的 system prompt 早就膨脹到 3-5 萬 token 被靜默截斷；順手把 Mac mini 換上 A3B 當夜間工人測了兩天。
-
與其拜託模型不要騙我，不如寫一個擋得住的 hook
接續 Opus 4.8 捏造工具輸出那篇：我把「別亂編數字」從 CLAUDE.md 的拜託，升級成一個掛在 PreToolUse 擋 git commit 的 pending-guard hook。
-
第二次 harness 減肥：全域 skills 58 砍到 40
接續 harness 減肥系列，半年後做了第二輪更大規模的整理，把全域 skills 從 58 個砍到 40 個的完整過程。
-
A Student's Seven-Agent Setup That Wouldn't Run — A Cautionary Tale About Harness First
A student built seven agents on an M5 Max and kept hitting timeouts. The problem wasn't a weak model — it was skipping the basics: harness, context management, task routing.
-
學員自建七個 agent 跑不動——一個 harness 為重的反面教材
一個學員在 M5 Max 上自建七個 agent 卻一直 timeout，問題不在模型不夠強，而在沒先把 harness、上下文管理、任務分派這些基本功打好。
-
Two Months of Claude Code Harness Ops: Crashes, Zombies, Cache Misses, and Plan Mode Edges
From March to April, running Claude Code as a long-running harness surfaced a bunch of issues. Survival tactics during the outage, my hook for cleaning up zombie processes, the silent memory bloat fixed in 2.1.86, the resume cache-miss fix in 2.1.90, and the Plan-Mode-question-density problem after 4.7.
-
Claude Code Harness 兩個月運維史：當機、殭屍、快取、Plan Mode 的邊邊角角
從 3 月到 4 月，Claude Code 作為長時間運轉的 harness，踩過的幾個點：當機時的求生策略、我自己寫的 hook 清殭屍進程、2.1.86 的記憶體暴增修復、2.1.90 修好的 resume cache miss、還有 Plan Mode 問題太多的 4.7 癥結。