diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
index 3baf0b9..23c65ab 100644
--- a/.claude-plugin/marketplace.json
+++ b/.claude-plugin/marketplace.json
@@ -1,6 +1,6 @@
 {
   "name": "wisdom",
-  "version": "3.0.3",
+  "version": "3.1.0",
   "owner": {
     "name": "cmj",
     "url": "https://github.com/cmj0121"
@@ -9,7 +9,7 @@
     {
       "name": "wisdom",
       "description": "Single-purpose skills to plan, build, review and ship software, led by smith.",
-      "version": "3.0.3",
+      "version": "3.1.0",
       "source": "./plugins/wisdom"
     }
   ]
diff --git a/README.md b/README.md
index a7dd5f4..75b49f5 100644
--- a/README.md
+++ b/README.md
@@ -75,6 +75,7 @@ other skills in turn. Every skill also works on its own.
 | Skill      | Purpose                                                | Magic word       |
 | ---------- | ------------------------------------------------------ | ---------------- |
 | `diagram`  | Draw an ASCII diagram of a decided structure           | `draw a diagram` |
+| `assets`   | Build media assets one confirmed change per round      | `build assets`   |
 | `compact`  | Re-render the previous answer as a dense table         | `compact it`     |
 | `brief`    | Keep answers short and discuss one topic at a time     | `brief me`       |
 | `lingua`   | Answer in your chosen language, remembered per project | `reply in`       |
diff --git a/README.zh-TW.md b/README.zh-TW.md
index 6bb399c..6c45ce7 100644
--- a/README.zh-TW.md
+++ b/README.zh-TW.md
@@ -72,6 +72,7 @@ v3 從零重寫，圍繞三條規則。
 | Skill      | 目的                                            | Magic word       |
 | ---------- | ----------------------------------------------- | ---------------- |
 | `diagram`  | 為已定案的結構畫 ASCII 圖                       | `draw a diagram` |
+| `assets`   | 一輪一個確認過的修改，逐步產出媒體素材          | `build assets`   |
 | `compact`  | 把上一個回答重新整理成精簡表格                  | `compact it`     |
 | `brief`    | 讓回答保持簡短，一次只討論一個主題              | `brief me`       |
 | `lingua`   | 以你選定的語言回答，並依專案記住                | `reply in`       |
diff --git a/plugins/wisdom/.claude-plugin/plugin.json b/plugins/wisdom/.claude-plugin/plugin.json
index ee9ca0c..95bf650 100644
--- a/plugins/wisdom/.claude-plugin/plugin.json
+++ b/plugins/wisdom/.claude-plugin/plugin.json
@@ -1,7 +1,7 @@
 {
   "name": "wisdom",
   "description": "Single-purpose skills to plan, build, review and ship software, led by smith.",
-  "version": "3.0.3",
+  "version": "3.1.0",
   "license": "MIT",
   "author": {
     "name": "cmj",
diff --git a/plugins/wisdom/skills/assets/SKILL.md b/plugins/wisdom/skills/assets/SKILL.md
new file mode 100644
index 0000000..0414288
--- /dev/null
+++ b/plugins/wisdom/skills/assets/SKILL.md
@@ -0,0 +1,105 @@
+---
+name: assets
+description: Builds image, video or audio assets one confirmed change per round against locked concepts. Use when a media asset must converge instead of drifting.
+license: MIT
+allowed-tools:
+  - Read
+  - Write
+  - Edit
+metadata:
+  shortcut: "build assets"
+---
+
+# Assets
+
+Build a media asset — image, video, audio — in rounds. Each round changes one thing against
+a fixed base, so the asset converges instead of drifting. The skill owns the loop and the
+lock files; generating belongs to whatever backend the environment already has.
+
+This version is written against images. Video and audio run the same loop as far as the
+backend allows; there are no timeline, clip or track fields.
+
+## Inputs
+
+Ask for the missing ones in a single question; never pick one.
+
+- **Session directory** (`$SESSION`) — where the lock files and candidates live. The caller
+  decides it before this skill runs: a project or a session directory, in git or not.
+- **Set and asset** — the set names the lock file (`<set>_assets.lock`); the asset names
+  the entry in it.
+- **Shared concepts** — on a first run with no global lock, which concepts every asset
+  shares; none is an answer.
+- **Language** — the `lingua` preference in context, else the user's language. The round
+  spec's labels are written in it too.
+
+## Lock files
+
+YAML, hand-editable, and only confirmed state — a candidate never enters a lock.
+
+`$SESSION/ASSETS.global.lock` — concepts every asset in the project shares:
+
+```yaml
+concepts:
+  style: hand-drawn watercolour, low saturation
+  character: a stuffed bear with a notched left ear
+```
+
+`$SESSION/<set>_assets.lock` — one asset set:
+
+```yaml
+candidates: 3 # the count last confirmed for this set
+assets:
+  bear:
+    current: candidates/bear-r1-2.jpg # the last pick: the next round's base
+    frozen: [face, pose, crop, background]
+    local: # confirmed concepts of this asset alone
+      belly: short hemp threads round the tear
+    history:
+      - base: images/38.jpg
+        change: add a ring of short hemp threads round the belly tear
+        scope: the edge of the tear only
+        picked: candidates/bear-r1-2.jpg
+```
+
+Global, `frozen` and `local` concepts all constrain a round, except the one concept the
+round changes. A local concept may add to the global ones, never override one: a change
+that contradicts a global concept stops the round — ask whether the global concept itself
+is to change, since that reaches every asset. On a yes the change is written with the pick
+to the global lock, not under `local`.
+
+## Round
+
+1. **Load** both lock files when they exist. Neither existing is a first run, not an error.
+2. **Base.** The asset's `current`; with no entry, the file the user names. With no base at
+   all, design one first: a round whose spec has no base and whose change is the asset's
+   description.
+3. **Frozen and change.** Frozen defaults to the lock's `frozen`. The change comes from the
+   user, one per round — two changes are two rounds; ask which goes first.
+4. **Scope.** The region or aspect the change may touch. Given the change but no scope,
+   propose the smallest scope that can hold it rather than asking an open question.
+5. **Candidates.** The count the user last gave, else the lock's `candidates`, else 3.
+6. **Confirm.** Show the round spec and stop — the only confirmation stop in a round.
+   Generate nothing before the user confirms; an edited line is applied and the spec shown
+   again. In English the labels are `Base`, `Frozen`, `Change`, `Scope`, `Candidates`.
+
+   ```text
+   基準: images/38.jpg
+   凍結: 臉、姿勢、剪裁、背景
+   改: 腹部破口邊緣加一圈短麻線
+   範圍: 只動破口邊緣
+   候選: 3
+   ```
+
+7. **Generate** with the image, video or audio tool available in the session (an MCP
+   server or a CLI). The prompt carries the base, every constraining concept, the change
+   and the scope. Save the candidates as `$SESSION/candidates/<asset>-r<round>-<n>.<ext>`,
+   `<round>` being one past the highest already on disk for the asset, and show their
+   paths. With no such tool, output the spec and the prompt, say that nothing was
+   generated, and wait for the user's files.
+8. **Pick.** On a pick, write the asset's entry to `<set>_assets.lock`, creating both lock
+   files on the first pick, the global one even with no concepts: set `current` to the
+   pick, record the change under `local`, keep the confirmed `frozen`, append the round to
+   `history`, store the candidate count. Shared concepts go to `ASSETS.global.lock`. No
+   pick writes nothing, and the next round keeps the same base.
+
+Never delete a candidate: the unpicked ones are the user's to clean up.