SYS: NOMINAL · OP: A.CHEN --:--:-- PT

// 機房 · FLOOR ONLINE

生產線。
系統就在這裡打造。

這裡是 xCraft AI 的建置端——驅動產線的工具鏈、進行中的生產管線, 以及一份記載交付與故障的日期日誌。這裡沒有履歷;履歷在展示廳。 這個房間就是機器本身,公開運轉。

Floor status · live $ status --floor
build
XC_001 "AI Silent Data Loss" · 視覺 2/6 ▲ in prod
pipeline
概念 3 · 視覺 1 · 查看看板
floor
6 款運轉中 · 6 款在測試台 · 檢視
lab_log
35 則紀錄 · 最新 2026.07.05
since
2026.05.12 · 公開記錄中
channel
youtube 開通中 · 2026 上線前

> 產線一切正常。捲動以檢視各模組

▸ 進入 展示廳

徽章:由程式生成——先生成,再手動微調。一如這裡的其他一切。

// 系統檢查

本頁的每一項聲明,都附上真實狀態。點任一行即可跳轉。

// 02 · 技術堆疊

商業層,以及底下支撐它的堆疊

這套實務以一根主柱領軍——把商業需求變成真正交付的 AI 解決方案。 底下的五項技術專業,是讓它成真的交付堆疊。狀態只會在 實驗日誌說變的時候才變。

主柱 · 商業層

商業層

開放接案

整個堆疊存在的目的:把真實的商業需求,變成能交付上線的解決方案。 理解營運、找出痛點、提出對策、動手建置並導入—— 下方五項專業,正是讓它可交付的憑藉。

UNDERSTANDDIAGNOSEPROPOSEIMPLEMENT

商用不動產 AI

草擬交易與行銷文件,把租約與 CC&R 摘要成白話重點,並讓業主報告隨時保持最新——讓你的團隊把時間花在成交,而不是整理資料。

牙科診所 AI

先從前台與行銷工作流程著手——溝通、內容、影像前置作業——凡涉及病患紀錄的部分,一律先簽妥相應協議,並以本地部署為前提。

租賃管理 AI

把申請人文件整理成乾淨一致、便於審閱的格式,讓合規文件準時到位,並核對租金與維護紀錄——租客決定仍由人做。

這是服務項目,不是案例研究:開放接案中,尚無交付紀錄。看接下來交付什麼

↓ 由交付堆疊支撐 ↓

P01

基本功與提示詞工藝

◐ IN PRACTICE

規格先行的提示詞。先寫需求,再寫提示詞;模型依任務挑選,而不是憑習慣。

› 佐證:每天用於下方每一格的實際產出。

P02

情境與記憶

◐ IN PRACTICE

懂你世界的 AI:常駐助理、結構化記憶、MCP 連線。

› 佐證:個人 AI 作業系統——實驗日誌 2026.05.122026.05.23

P03

AI 媒體製作

▲ IN PRODUCTION

堆疊最深的一端:把影像、影片與語音模型當成生產管線來跑,而不是拉霸機。分鏡 → 生成 → 剪輯 → 交付。xCraft AI 工作室就靠它運轉。

› 佐證:實驗日誌 2026.05.31 · 第一支影片製作中,工作室見下方。

P04

代理與自動化

◐ IN PRACTICE

是委派,不是下提示:限定範圍的自主工作流程,加上人在迴圈的護欄。排程自動化層已在運轉;代理層建置中。

› 佐證:實驗日誌 2026.05.15 · 2026.05.31 · 網格上的機器人(是模擬——而且清楚標示)。

P05

AI 優先建置

◐ IN PRACTICE

由 AI 擔任主要作者、工程師擔任編輯的軟體。頭號例證:這個網站——AI 起草、人工審校的 HTML,由靜態建置流程從共用片段組裝而成。無框架。歡迎檢視原始碼,一切都在裡面。

› 佐證:檢視網頁原始碼。

◐ IN PRACTICE:正用於真實工作 · ▲ IN PRODUCTION:正在上線管線產出成果。

// 03 · 工作室

媒體支柱在這裡交付

X xCraft//AI

xCraft AI 是我的工作室——這套架構實際上場操練的生產端。 AI 製作的影片,用製造產線的方式來跑:可重複的流程、成本紀律、 品質關卡。工作室負責發布;這個房間,展示的是它背後的操作者。

STORYBOARDGENERATECUTSHIP

XC_001 · IN PRODUCTION VID

> "AI Silent Data Loss"——一則 48 秒的現場筆記

> 階段:視覺——6 個節拍已生成 2 個

> 頻道:開通中 · 2026 上線前

> 管線:查看下方看板

> 上線時通知我

// 生產管線

產線上現在跑著什麼

(生產看板為原始工作紀錄,依英文原文保留。)

Concept3

XC_002

The personal-OS thesis

greenlit: why sitting out the tool is the real loss

XC-LF1

The Cowork / CLI journey

working with an imperfect, powerful AI

P2 · idea

Memory that survives the session

durable AI memory without context bloat

Script + Voice0

· empty ·

Visuals1

XC_001

AI Silent Data Loss

48s field note on a public AI-tool failure · 2 of 6 beats generated

Assembly0

· empty ·

Published0

· empty ·

BOARD STATE: 2026.07.07 · baked from a data file. Empty columns are the real state, not a hidden one.

// 04 · 產線

產線靠什麼裝備

工具要靠實際做事才能在這條產線上佔一席之地,而不是靠名氣。兩個層級: 一邊是此刻正在跑產線的,另一邊是在測試台上的——已購入、 測試中,在通過驗證之前,不會託付任何要交付的工作。

▲ 已裝備 · 正在跑產線6

▲ IN PRODUCTION

Claude · Anthropic

核心操作員

每台工作站上的代理與 CLI——下方一切的那雙手。規格進,成果出。

▲ IN PRODUCTION

Midjourney

影像生成

供給影片產線的分鏡畫格與靜態影像——在動起來之前的視覺節拍。

▲ IN PRODUCTION

ElevenLabs

語音

旁白與聲音複製——每一段剪輯底下的敘事音軌。

▲ IN PRODUCTION

Google Workspace · Gemini

基礎設施

郵件、雲端硬碟,加上接進日常工作的空間內建模型。這條產線的水電管路。

▲ IN PRODUCTION

Custom GPT

客製化助理

建立在 OpenAI 上、限定任務範圍的助理——為一件重複發生的工作,打包好指令與參考資料。

▲ IN PRODUCTION

Gemini Gem

客製化助理

同一套模式搬進 Google 的體系——為單一重複任務調校的 Gem,就放在 Workspace 原本的所在。

◐ 測試台上 · 評估中6

HeyGen

虛擬人像/主持人影片

VAPI

語音代理 API

Phot.AI

影像編輯與強化

Lyzr

代理框架

Fireflies

會議擷取與逐字稿

n8n

工作流程自動化

另有 +10 款收在儲備區,等它們爭取到一份工作再上場

▲ IN PRODUCTION:正在跑實際工作 · ◐ EVALUATING:已購入,測試中。Magic Hour AI 曾試用於影片製作,跑了一輪就被剔除——測試台就是工具證明自己(或證明不行)的地方。

// 05 · 實驗日誌

我建了什麼、壞了什麼、學到什麼

一份從零打造個人 AI 作業系統、逐條標註日期的工程日誌。沒有精華剪輯:失敗都留在紀錄裡,因為日誌本該如此。 // 捲動查看完整歷史(日誌為原始工作紀錄,依英文原文保留。)

ENTRIES: 35 · FIRST: 2026.05.12 · LATEST: 2026.07.05 · LOGGING CONTINUES

  1. REBUILD

    Site v2.0 shipped—one hand-written file, rebuilt in place. New job: public portfolio for an AI Generalist practice layered on an engineering career. Every status on the page is now baked from this log.

  2. IDENTITY

    Anonymity layer retired—the log gains a byline. Alex Chen: mechanical engineer, nearly two decades in hardware—the last decade in defense aerospace—operator of everything below. The entries stand as written.

  3. SKILL-AUDIT

    Turned the audit on the auditor—a research panel reviewed the OS’s own memory-capture protocol, patched its real gaps, then a second interrogation cut forty percent of the new text as dead weight. Every rule must earn its lines.

  4. VISION-INGEST

    Two years of records that survived only as scanned images—recovered by direct vision transcription, each batch gated by a closing-balance roll-forward so nothing landed until it reconciled to the cent.

  5. SILENT-FAILURE

    An overnight test caught a nightly pipeline silently declining all its work over one cosmetic flag—while still reporting all-clear. Gate narrowed to real damage; freshness tripwire added. Silent staleness is worse than a loud crash.

  6. LEDGER

    A data store rebuilt as a lossless, append-only ledger—every derived table a pure projection, and every record corroborated against the source documents’ own closing balances.

  7. REMEDIATION

    Nine-scanner audit of the whole OS, every finding adversarially verified—34 confirmed, 2 refuted—then the full remediation shipped the same session, with cleanup rules that now enforce themselves.

  8. SKEPTICS

    A skeptic-verifier agent panel overturned its own migrate-everything plan—most scheduled bots proved redundant and were retired in favor of one always-visible countdown.

  9. AUDIT

    Forty-six Claude subagent review passes across six modules; findings fixed in waves. Sharpest lesson: several flagged “duplicates” were intentional cross-checks. Verify before fixing.

  10. SELF-DOCS

    Every module now exposes an under-the-hood engine diagram, reachable from one central hub—and a single nightly routine quietly keeps every dashboard current.

  11. GUARDRAIL

    Codified when a surface is generator-built versus hand-authored—then got bitten by a stale reference to a component already retired, and turned the bite into a standing guardrail.

  12. DATA-BRIDGE

    Agent-as-bridge pattern codified—the agent pulls from tool-locked sources, writes a plain data file, and an offline generator rebuilds the page from it. One source of truth, no hand-edited HTML.

  13. AI-REVIEW

    Multi-agent adversarial review caught five real bugs—and one confident, high-severity false positive. Verdict: AI review surfaces; the engineer verifies.

  14. SELF-HEALING

    Self-healing backup gate online—every automated routine now reports real health from its output artifacts, not the scheduler’s “it ran” flag.

  15. SELF-PORTRAIT

    The OS learned to render itself—one generated command deck mapping module dependencies, rule cascades, live memory health, and the whole knowledge corpus as a navigable galaxy.

  16. READ-INTEGRITY

    Integrity watchdog hardened against cloud-sync phantoms—no byte is trusted until two independent reads agree.

  17. ARCHITECTURE

    Data pipeline refactored to a clean backend/frontend split—every filesystem path centralized into one source-of-truth module, and “backend store first” codified as a standing convention.

  18. DEPLOY

    Studio site shipped to production—source compiled to a three-file static bundle, the log baked to static markup, DNS repointed behind auto-renewing TLS with the mail records untouched.

  19. AGENTS

    Decorative grid bots shipped across the site interface: an animation, labeled a simulation on the page. Set dressing for pillar P04, not automation.

  20. TELEMETRY

    Autonomous workspace logging driven by agentic AI—a curated, anonymized stream broadcasting real-time development notes from the lab.

  21. ▸ load earlier entries (15)
    • SCALE

      Scaled to thirteen domain modules—full-system migration and integrity audit executed in a single pass.

    • PIPELINE

      Agentic media pipeline online—script to synthetic voice to generated visuals, fully orchestrated; assembly stage next.

    • WATCHDOG

      Autonomous integrity watchdog deployed—nightly scans guard every file against silent corruption.

    • RENDER

      Cinematic command deck shipped—a navigable bridge rendered over the entire system.

    • MAINTENANCE

      Self-maintaining changelog—a scheduled engine compresses and curates the OS’s own development history.

    • SCHEDULER

      Event-derivation engine—a database-driven timeline computes its own forward schedule.

    • AI_LOGIC

      Prompt-as-specification methodology adopted—constrained inputs, defined acceptance criteria, verifiable outputs.

    • MEMORY

      Memory architecture overhauled—tag-indexed recall, layered protocols, and self-pruning logs.

    • DESIGN

      Tokenized design system shipped—one-switch theming across every interface in the OS.

    • COMMAND

      Unified command center built—a single aggregation layer surfacing every module at a glance.

    • ARCHITECTURE

      Modular domain architecture established—isolated knowledge modules, each with dedicated memory and rules.

    • AUTOMATION

      Autonomous triage module deployed—agentic classification over a live, high-volume message stream.

    • INTEGRITY

      Systemic write-integrity fault traced to root cause—verification guardrails hardened around every file operation.

    • INTERFACE

      First data dashboard online—a bento-grid view over a live structured-data layer.

    • KERNEL

      Personal OS kernel initialized—persistent memory architecture and a constitution of operating rules.

// 06 · 接下來交付

路線圖公開,所以空檔也公開

YT-01 YouTube 頻道 ——第一支影片製作中;頻道開通中。 IN BUILD
NL-01 電子報寄送 ——每則實驗日誌配一封簡短的通知信。需要外部郵件服務;評估中。 EVALUATING
FW-04 架構深化:代理 ——當一條限定範圍的自主工作流程撐過一個月的實際使用,P04 就升為 IN PRODUCTION。 IN BUILD

UPDATED: 2026.07.10 ——部署時烘焙。如果這個日期已經很舊,那本身也是資訊。

// 07 · 聯絡

兩條直達線路

關於系統、架構或日誌的問題,走這裡。想聘用或合作?那在展示廳——切換轉盤就在頁面上方。

DIRECT · ACTIVE

contact@xcraftai.com。關於系統、架構或日誌的問題都歡迎。每一封我都會讀。

▸ 傳送訊息

PROFESSIONAL · ACTIVE

LinkedIn——以標準格式呈現的職涯:操作者背後的工程紀錄。

▸ 開啟 LinkedIn