diff --git a/README.md b/README.md
index e0f371f6c..0d6fc24c2 100644
--- a/README.md
+++ b/README.md
@@ -233,6 +233,24 @@ lark-cli api POST /open-apis/im/v1/messages --params '{"receive_id_type":"chat_i
 --format csv       # Comma-separated values
 ```
 
+### JSON Output Contract
+
+With `--format json` (the default), success and error envelopes are distinct.
+
+Success goes to **stdout**, exit code `0`:
+
+```json
+{ "ok": true, "identity": "user", "data": { "guid": "..." }, "meta": { "count": 1 } }
+```
+
+Errors go to **stderr**, non-zero exit code:
+
+```json
+{ "ok": false, "identity": "user", "error": { "type": "api", "subtype": "...", "code": 99991679, "message": "...", "hint": "..." } }
+```
+
+To check whether a command succeeded, test `ok == true` (or the exit code) — **not** `code == 0`. Unlike raw OpenAPI responses (`{"code": 0, "msg": "ok", ...}`), the success envelope carries no `code` or `msg` field; `code` appears only inside `error` as the upstream OpenAPI code. See [errs/ERROR_CONTRACT.md](errs/ERROR_CONTRACT.md) for the full error taxonomy.
+
 ### Pagination
 
 ```bash
diff --git a/README.zh.md b/README.zh.md
index 6d8221f2a..74b706405 100644
--- a/README.zh.md
+++ b/README.zh.md
@@ -234,6 +234,24 @@ lark-cli api POST /open-apis/im/v1/messages --params '{"receive_id_type":"chat_i
 --format csv       # 逗号分隔值
 ```
 
+### JSON 输出契约
+
+`--format json`（默认）下，成功与错误的信封结构不同。
+
+成功信封写入 **stdout**，退出码 0：
+
+```json
+{ "ok": true, "identity": "user", "data": { "guid": "..." }, "meta": { "count": 1 } }
+```
+
+错误信封写入 **stderr**，退出码非 0：
+
+```json
+{ "ok": false, "identity": "user", "error": { "type": "api", "subtype": "...", "code": 99991679, "message": "...", "hint": "..." } }
+```
+
+判断命令是否成功，请检查 `ok == true`（或进程退出码），**不要用 `code == 0`**。与原始 OpenAPI 响应（`{"code": 0, "msg": "ok", ...}`）不同，成功信封没有 `code` 和 `msg` 字段；`code` 只出现在错误信封的 `error` 内，含义是上游 OpenAPI 的 numeric code。完整错误分类见 [errs/ERROR_CONTRACT.md](errs/ERROR_CONTRACT.md)。
+
 ### 分页
 
 ```bash
diff --git a/errs/ERROR_CONTRACT.md b/errs/ERROR_CONTRACT.md
index 99dce8948..5262ce58d 100644
--- a/errs/ERROR_CONTRACT.md
+++ b/errs/ERROR_CONTRACT.md
@@ -72,6 +72,28 @@ other category. `error.type` is `"policy"`, `error.subtype` is one of
 `challenge_required` / `access_denied`, and process exit is `6` via
 `CategoryPolicy`.
 
+### Success envelope (stdout)
+
+For contrast: success responses render to **stdout** as an
+`output.Envelope` (`internal/output/envelope.go`), exit code `0`:
+
+```json
+{
+  "ok": true,
+  "identity": "user",
+  "data": { "guid": "e297d3d0-..." },
+  "meta": { "count": 1 }
+}
+```
+
+Consumers must branch on `ok` (or the process exit code). The success
+envelope has **no top-level `code` or `msg` field** — `code` exists only
+inside `error`, where it is the upstream numeric code (invariant 4).
+Wrappers that follow the raw OpenAPI convention and test `code == 0`
+will misclassify every successful call as a failure, which is
+especially dangerous around write commands (e.g. retrying a create that
+already succeeded).
+
 ## Categories
 
 | Category | When | Exit | Typed struct |
diff --git a/skills/lark-shared/SKILL.md b/skills/lark-shared/SKILL.md
index 272b2bae4..e78ddd496 100644
--- a/skills/lark-shared/SKILL.md
+++ b/skills/lark-shared/SKILL.md
@@ -146,6 +146,24 @@ lark-cli update
 
 **重要**：始终使用 `lark-cli update` 更新，它会同时更新 CLI 和 AI Skills。
 
+## JSON 输出契约
+
+`--format json`（默认）下，成功与错误的信封结构不同：
+
+成功信封写入 **stdout**（退出码 0）：
+
+```json
+{ "ok": true, "identity": "user", "data": { "guid": "..." }, "meta": { "count": 1 } }
+```
+
+错误信封写入 **stderr**（退出码非 0）：
+
+```json
+{ "ok": false, "identity": "user", "error": { "type": "api", "subtype": "...", "code": 99991679, "message": "...", "hint": "..." } }
+```
+
+**判断成功必须用 `ok == true`（或进程退出码 0），不要用 `code == 0`**：成功信封没有顶层 `code` / `msg` 字段，`code` 只出现在错误信封的 `error` 内，含义是上游 OpenAPI 的 numeric code。按 OpenAPI 老格式 `{"code": 0, "msg": "ok"}` 判断会把所有成功调用误判为失败；封装写入类命令（如 `task +create`）时尤其危险，误判会绕过幂等逻辑导致重复创建。
+
 ## 安全规则
 
 - **禁止输出密钥**（appSecret、accessToken）到终端明文。
diff --git a/skills/lark-task/references/lark-task-create.md b/skills/lark-task/references/lark-task-create.md
index 7c946df4d..f8bb5fc23 100644
--- a/skills/lark-task/references/lark-task-create.md
+++ b/skills/lark-task/references/lark-task-create.md
@@ -46,7 +46,20 @@ lark-cli task +create --summary "Test Task" --dry-run
 1. Confirm with the user: task summary, due date, assignee, and tasklist if necessary.
    - **Crucial Rule for Assignee**: If the user explicitly or implicitly says "create a task for me" (给我创建一个任务), or "help me create a task" (帮我新建/创建一个任务), you MUST assign the task to the current logged-in user. You can get the current user's `open_id` by executing `lark-cli auth status` (it already outputs JSON by default, so do not add `--json`) or `lark-cli contact +get-user` first, extracting `.identities.user.openId` (from `auth status`) or `.data.user.open_id` (from `contact +get-user`), and then passing it to the `--assignee` parameter.
 2. Execute `lark-cli task +create --summary "..." ...`
-3. Report the result: task ID and summary.
+3. Judge success by `ok == true` in the stdout JSON (the success envelope has no `code` field — do not test `code == 0`), then report the result: task ID (`data.guid`) and summary.
+
+Example success response:
+
+```json
+{
+  "ok": true,
+  "identity": "user",
+  "data": {
+    "guid": "e297d3d0-4b60-4a5f-a4d4-xxxxxxxxxxxx",
+    "url": "https://applink.larkoffice.com/client/todo/detail?guid=e297d3d0-4b60-4a5f-a4d4-xxxxxxxxxxxx"
+  }
+}
+```
 
 > [!CAUTION]
 > This is a **Write Operation** -- You must confirm the user's intent before executing.