Value · 應用價值

The same file is engineering manual, spec, disaster-recovery playbook, and an executable command sequence Claude Code runs directly. Infrastructure-as-Code is common. Infrastructure-as-Prose — where Bicep, az CLI, Docker, npm, SSML are woven into human-readable narrative — is uncommon.

同一份 build.md 是工程手冊 + 規格書 + 災難復原 playbook + Claude Code 直接執行的指令集。Infrastructure-as-Code 常見;Infrastructure-as-Prose 少見。

Not "I ran it once, here." Two iterations, 48 hours apart, 805 seconds each — same second. Strict rules: working tree clean throughout, §E 6/6 programmatic checks green, fail-fix-restart loop for source bugs. Qualified iterations are the only ones counted.

不是「跑一次秀給你看」,是嚴格規則下兩次、隔 48 小時、各 805 秒完全同秒數(2026-04-22-b = 2026-04-24-b)。合格條件:working tree 全程未改 source、§E 六項全綠。

When build.md breaks (Azure CLI upgraded, API behavior drifted), Claude Code automatically classifies the failure, tears down the RG, snapshots build.md, fixes the bug, commits + pushes, and restarts with the next letter. No human in the loop.

當 build.md 壞了(Azure CLI 升級、API 行為漂移),Claude Code 自動:辨識 source bug → 砍 RG → snapshot → 修 build.md → commit + push → 換 letter 重跑。不問人。

After finding a bug, Claude writes the rule "don't step into this again" (spec/001/build-md-authoring.md) and follows it itself: snapshot before revise, positive confirmation, no grep -q <magic-string>, idempotent blocks. Each spec entry points to a specific iteration that triggered it.

AI 跑完發現 bug,寫下「以後別再踩這坑」的規則 (spec/001/build-md-authoring.md)並自我遵守:snapshot before revise、positive confirmation、禁止 grep -q <magic-string>。Spec 觸發事件可追回具體 iteration letter。

One 14-minute iteration passes through: AAD role grants, KV RBAC, SQL Server AAD→SQL auth toggle, Azure Speech SSML, Prisma schema push, Docker multi-stage, MSI AcrPull, CORS, Blob public-access, Next.js standalone, ISR. Claude Code doesn't stall on "I don't know this one."

一次 iteration 穿越:AAD role 授權、KV RBAC、SQL Server AAD↔SQL auth 切換、Speech SSML、Prisma schema push、 Docker multi-stage、MSI AcrPull、CORS、Blob public-access、Next.js standalone、ISR。Claude Code 沒卡在「我不熟這個」。

Every iteration leaves traces (checkpoints-*.md, iteration-*.md) — both qualified and disqualified runs are committed. You can trace iter-a disqualified because §7 grep-q-blob bug → commit 94ce474 fix → iter-b qualified 805s in full detail. Not cherry-picked demo.

每次 iteration 留痕(checkpoints-*.md、iteration-*.md),合格 / 失格都進 git。 評審直接看到「iter a 失敗 → commit 94ce474 修 → iter b 合格 805s」的完整鏈條。不是 cherry-picked demo。