Home » Categories

AI Systems

Practical notes on LLMs, agents, MCP, Claude Code, skills, token economics, and modern AI engineering.
AI Systems

What Makes an AI Skill Different from a Prompt?

A prompt and a skill can contain the same words. The difference is the machinery around them: a skill is a file the model loads on its own when your request matches its one-line description, and it can declare inputs, tools, and a scope it won't cross. Here's how that loading works and why it lets you build things a prompt can't.

AI Systems

Why AI Agents Fail in Production

Why AI agents fail: most failures trace to governance gaps (phase gates, state handoffs, capability boundaries) more than to the model itself, and the two need completely different fixes. How to tell them apart.

AI Systems

Token 是什麼？LLM 為何只讀 Token？

Token 是什麼？LLM 為何不直接處理完整字詞：解析字級、字元級、子詞級三種 tokenization 方法，包含 BPE 示範程式碼，以及 token 數量對計費和上下文長度的影響。

LLM

See all 47
AI Systems

Claude 的文字浮水印怎麼運作

Anthropic 說明了未來 Claude 的文字浮水印怎麼做：做法是換一個亂數來源，用來決定下一個字挑哪一個，不加字也不多花 token。這篇是來聊聊我們目前獲得的資訊有哪些。

AI Systems

How the KV cache speeds up LLM generation

A transformer redoes the same attention projections for every past token at each decoding step. The KV cache stores those keys and values so they get reused instead of recomputed, and the one cost it adds is memory that grows with the sequence.

Agent

See all 41
AI Systems

Running more AI reviewers on the same code

Fan three or four AI reviewers at the same diff and what comes back is a list of candidates. Agentic OS's own audit records show how many survive checking, and why its default has dispatched subagents return evidence while one primary owns the write.

AI Systems

subagent 分派：第二個 agent 該拿到什麼

開一個 agent 去看另一個 agent 做完的東西很容易，難的是後面那個看得見前面漏掉了什麼。從 Agentic OS 的 review 規定看下去：要讓第二個 agent 看得出東西，靠的是刻意不給它 session、對話記錄和實作理由。

Governance

See all 24
AI Systems

Claude 的文字浮水印怎麼運作

Anthropic 說明了未來 Claude 的文字浮水印怎麼做：做法是換一個亂數來源，用來決定下一個字挑哪一個，不加字也不多花 token。這篇是來聊聊我們目前獲得的資訊有哪些。

AI Systems

Running more AI reviewers on the same code

Fan three or four AI reviewers at the same diff and what comes back is a list of candidates. Agentic OS's own audit records show how many survive checking, and why its default has dispatched subagents return evidence while one primary owns the write.

AI Systems

subagent 分派：第二個 agent 該拿到什麼

開一個 agent 去看另一個 agent 做完的東西很容易，難的是後面那個看得見前面漏掉了什麼。從 Agentic OS 的 review 規定看下去：要讓第二個 agent 看得出東西，靠的是刻意不給它 session、對話記錄和實作理由。

AI Systems看中文版

What Are GPT-5.6's Sol, Terra, and Luna?

OpenAI shipped GPT-5.6 as Sol, Terra, and Luna on July 9, 2026, and quietly ended the dedicated Codex checkpoint. What replaced it is a reasoning-effort dial whose top notch spawns subagents. This post lays out the three models, the six effort levels, and what the independent benchmarks measured.

Architecture

See all 23
AI Systems

Running more AI reviewers on the same code

Fan three or four AI reviewers at the same diff and what comes back is a list of candidates. Agentic OS's own audit records show how many survive checking, and why its default has dispatched subagents return evidence while one primary owns the write.

AI Systems

subagent 分派：第二個 agent 該拿到什麼

開一個 agent 去看另一個 agent 做完的東西很容易，難的是後面那個看得見前面漏掉了什麼。從 Agentic OS 的 review 規定看下去：要讓第二個 agent 看得出東西，靠的是刻意不給它 session、對話記錄和實作理由。

Claude Code

See all 22
AI Systems

Running more AI reviewers on the same code

Fan three or four AI reviewers at the same diff and what comes back is a list of candidates. Agentic OS's own audit records show how many survive checking, and why its default has dispatched subagents return evidence while one primary owns the write.

AI Systems

subagent 分派：第二個 agent 該拿到什麼

開一個 agent 去看另一個 agent 做完的東西很容易，難的是後面那個看得見前面漏掉了什麼。從 Agentic OS 的 review 規定看下去：要讓第二個 agent 看得出東西，靠的是刻意不給它 session、對話記錄和實作理由。

AI SystemsRead in English

AI 寫 code 為什麼又搬回終端機了

這兩年 AI 寫 code 的重心，悄悄從嵌在 IDE 裡的補全，搬回了終端機（Claude Code、Codex CLI 那一類）。我覺得這不是復古，是因為 agent 變成一個「你不盯著看的 process」，而終端機本來就是為這種東西設計的。

AI Systems看中文版

Why coding agents are moving back to the terminal

The coding agents developers reach for now — Claude Code, Codex CLI, Aider — are terminal programs, not IDE plugins. An IDE is built around a human at the keyboard, the terminal around processes you don't babysit. When AI coding became a job instead of a keystroke, it moved home.

Agentic OS

See all 21
AI Systems

Running more AI reviewers on the same code

Fan three or four AI reviewers at the same diff and what comes back is a list of candidates. Agentic OS's own audit records show how many survive checking, and why its default has dispatched subagents return evidence while one primary owns the write.

AI Systems

subagent 分派：第二個 agent 該拿到什麼

開一個 agent 去看另一個 agent 做完的東西很容易，難的是後面那個看得見前面漏掉了什麼。從 Agentic OS 的 review 規定看下去：要讓第二個 agent 看得出東西，靠的是刻意不給它 session、對話記錄和實作理由。

Token Economics

See all 18
AI Systems

The Same GLM 5.2 Has Different Prices Across Providers

GLM 5.2 shipped with open weights under an MIT license, which generally permits third parties to host and commercialize inference. As of writing, OpenRouter lists twenty-five providers offering it, at input prices from $0.93 to $3.00 per million tokens.

Prompt Engineering

See all 17
AI Systems

What Makes an AI Skill Different from a Prompt?

A prompt and a skill can contain the same words. The difference is the machinery around them: a skill is a file the model loads on its own when your request matches its one-line description, and it can declare inputs, tools, and a scope it won't cross. Here's how that loading works and why it lets you build things a prompt can't.

All posts (73)

Support My Work

Buy Me a Coffee