Domain Upgrade

Upgrade Log 格式（中斷時自動建立）：

| Item | Type | Dispatch To | Status | Notes |
|------|------|------------|--------|-------|
| ... | A/B/C | /skill-edit | pending/done | |

Phase 0: Context Discovery

State

• Reads: ~/.prismstack/projects/{slug}/.prismstack/check-results.json, edit-log.jsonl, convert-log.jsonl (what happened since last upgrade)
• Writes: upgrade-log.jsonl (append: timestamp, action, dispatched-to, result)
• Updates: domain-config.json → last_upgrade timestamp

自動搜尋上游產出和先前執行紀錄：

_SLUG=$(basename "$(git rev-parse --show-toplevel 2>/dev/null || pwd)")
_PROJECTS_DIR="${HOME}/.prismstack/projects/${_SLUG}"

# Search for recent /skill-check results
ls "${_PROJECTS_DIR}"/skill-check-*.md 2>/dev/null

# Search for recent /skill-edit commits
git log --oneline -10 --grep="skill-edit" 2>/dev/null

# Search for recent /source-convert additions
git log --oneline -10 --grep="source-convert" 2>/dev/null

# Search for prior upgrade logs
ls upgrade-log.md 2>/dev/null

如果找到先前的 skill-check 結果 → 摘要呈現，作為升級的起點參考。 如果找到先前的 upgrade-log → 讀取，告知用戶先前的升級進度。

Entry: Mode Routing

Check if user provided a mode argument. If not, ask:

你今天帶了什麼？

1. feedback — 我有測試回饋要處理
2. upgrade — 我想改進某個 skill
3. listen — 我有些想法但還不確定

Route to the matching mode below.

Mode 1: Feedback

Step 1 — Collect

讀取 shared/methodology/functional-test-guide.md — 用真實任務測試 skill 效果的方法（結構 vs 功能的差異、測試迴圈、記錄格式）。

Ask the user:

• 什麼 worked？（繼續保持的）
• 什麼 didn't work？（需要修的）
• 什麼 was surprising？（意料之外的，好壞都算）

Step 2 — Classify

For each feedback item, classify:

• A類（自動修）： 詞彙不對、artifact 格式不接、trigger 描述不準
  • These are mechanical fixes. Route to /skill-edit directly.
• B類（討論後修）： 判斷維度缺了、判斷方向錯了、互動節奏不對
  • These need discussion first. Present analysis, get user alignment, then dispatch.
• C類（需要升級材料）： gotchas 太淺、benchmark 不對、缺領域深度
  • These can't be fixed without new material. Add to upgrade backlog.

Step 3 — Present & Confirm

STOP GATE: Present the classified list to user before acting.

回饋分類結果：

A類（自動修）：
  - [item] → /skill-edit [target]

B類（討論後修）：
  - [item] → 需要先討論 [what]

C類（需要升級材料）：
  - [item] → 加入升級待辦，需要 [what material]

確認後我開始處理 A類。B類我們逐項討論。

Step 4 — Execute

• A items → dispatch to /skill-edit with specific instructions
• B items → discuss one by one → dispatch to appropriate skill
• C items → log to upgrade backlog, remind user what materials are needed

Step 5 — Verify

Run the 3-question upgrade effect check (see references/upgrade-patterns.md).

Mode 2: Upgrade

Step 1 — Scope

Ask:

• 哪個 skill？
• 想改進什麼？

Step 2 — Pattern Match

Read references/upgrade-patterns.md. Match user's need to a pattern:

If the need is ambiguous, use the Need Clarification Pattern from the reference.

Step 3 — Dispatch

Route to the matched skill with clear instructions:

• What to change
• Why (user's stated reason)
• Constraints (don't break X, keep Y)

Step 4 — Verify

STOP GATE: After dispatch completes, run the 3-question upgrade effect check:

Upgrade Impact Score (3 questions, scored)

Score range: -3 to +3.

• +2 to +3: Excellent upgrade
• +1: Meaningful improvement
• 0: Neutral (consider if worth the change)
• Negative: Revert immediately

Present results to user.

Mode 3: Listen

Step 1 — Open Conversation

Help the user articulate what they need:

• 「你覺得哪裡最卡？」
• 「哪個 skill 最常用但最不滿意？」
• 「上次用的時候有沒有哪裡覺得怪怪的？」
• 「有沒有什麼事情你想做但現在做不到？」

Step 2 — Translate

Turn vague feelings into specific observations:

• "覺得慢" → which step is slow? → specific skill or workflow gap
• "不夠好" → compared to what? → missing benchmark or gotcha
• "用不順" → where does it break? → workflow connection issue

Step 3 — Route

Once the need is clear, route to either:

• Feedback mode — if it's about existing skill quality
• Upgrade mode — if it's about adding/changing capability

Gotchas

1. Claude tries to do everything itself instead of dispatching. Always route to the specialized skill. You are a coordinator, not an implementer. If you catch yourself writing skill content directly, stop and dispatch.

2. Claude loses track of multiple upgrades. Maintain an upgrade log. After each session, summarize what was changed, what's pending, what needs materials.

3. Builder 不主動說「你的 skill 不夠好」. The user decides when to upgrade. You respond to their needs, you don't push upgrades on them. Provide perspective when asked, not unsolicited criticism.

4. Don't confuse upgrade with rebuild. If the user wants to change >50% of a skill, that might be a /skill-gen + replace, not a /skill-edit.

5. Always check downstream impact. A change to one skill can break the next skill in the workflow. The 3-question check exists for this reason.

Completion

Completion 萃取

報告 STATUS 前，回顧用戶在升級過程中的輸入。 萃取 4 種信號（expertise / correction / preference / benchmark）到 domain-config.json。 詳見 shared/methodology/context-accumulation-guide.md。 大部分 session 不需要萃取。

When done, report:

STATUS: DONE

Changes made:
- [list of changes with skill names]

Upgrade effect check:
1. Workflow: [result]
2. Judgment accuracy: [result]
3. Cross-skill connection: [result]

Pending (if any):
- [C-class items or deferred work]

建議下一步：
1. `/skill-check review {changed-skill}` — 驗證修改效果