diff --git a/.codex-plugin/plugin.json b/.codex-plugin/plugin.json
index 0f8c39d..e0ae4c7 100644
--- a/.codex-plugin/plugin.json
+++ b/.codex-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "computer-usage-summary-skill",
-  "version": "0.2.2",
+  "version": "0.2.3",
   "description": "Privacy-first, cross-platform local ActivityWatch summaries for Codex.",
   "author": {
     "name": "liuyewang",
diff --git a/docs/CHANGELOG.md b/docs/CHANGELOG.md
index c4df5fc..ed5dd39 100644
--- a/docs/CHANGELOG.md
+++ b/docs/CHANGELOG.md
@@ -4,6 +4,14 @@ This is a privacy-safe product and validation log. It records decisions and
 feedback themes, never ActivityWatch events, window titles, URLs, account
 names, local paths, or behavioral telemetry.
 
+## 2026-07-27 - v0.2.3 Standardized Conversational Output
+
+- Standardized conversational summaries around `Today at a glance`, `What you
+  did`, `App table`, and `Caveats`.
+- Kept the four-section structure for unavailable ActivityWatch data and made
+  reduced output opt-in when the user explicitly requests it.
+- Added a regression test for the default conversational output contract.
+
 ## 2026-07-27 - v0.2.2 Review Fixes
 
 - Convert invalid IANA time-zone input into a clean CLI validation error.
diff --git a/skills/computer-usage-summary/SKILL.md b/skills/computer-usage-summary/SKILL.md
index 7d22de8..b90434b 100644
--- a/skills/computer-usage-summary/SKILL.md
+++ b/skills/computer-usage-summary/SKILL.md
@@ -80,6 +80,41 @@ When the app, local API, buckets, or requested-range events are unavailable:
 7. When ActivityWatch is unavailable, return its structured reason. Markdown,
    TSV, and CSV outputs provide `status`, `reason`, and `source` columns.
 
+## Default Conversational Output
+
+For conversational summaries, always use the following four sections, in this
+order, unless the user explicitly asks for a simpler or shorter output:
+
+1. `Today at a glance`
+   - State the local date and time zone.
+   - Report confirmed active time, away time, and evidence coverage.
+   - In fallback mode, state that unavailable values are `unavailable`; do not
+     estimate them.
+2. `What you did`
+   - Give a chronological list of meaningful activity windows.
+   - Include the time window, activity, supporting app or sanitized title,
+     confidence, and evidence label (`confirmed ActivityWatch`, `Screen Time
+     reported`, `live snapshot`, or `metadata clue`).
+3. `App table`
+   - Include app name, observed foreground sessions, active time, first seen,
+     last active, and source note when ActivityWatch data is available.
+   - If data is unavailable, keep this section and show the structured reason
+     with `unavailable` metrics rather than omitting the table.
+4. `Caveats`
+   - State the relevant privacy and evidence limits, including that foreground
+     sessions are observed intervals rather than process-launch counts and that
+     tracking starts only when ActivityWatch starts.
+
+Keep these sections even when one contains little or no data. A request such as
+"简单一点", "简要输出", "只要结论", or "只给表格" is an explicit request
+for a reduced format; then omit or compress only the sections the user asks to
+remove. Do not infer a reduced format merely because the request is short.
+
+For a requested export (`TSV`, `CSV`, or a saved report), follow the export
+format the user requested instead of wrapping it in these conversational
+headings. If the user asks for both a summary and an export, provide the four
+conversational sections first and then the requested export.
+
 ## Privacy Rules
 
 - Treat window titles as sensitive. Remove URLs, truncate long titles, and
diff --git a/tests/test_activitywatch_summary.py b/tests/test_activitywatch_summary.py
index 9b3b3ae..71b6770 100644
--- a/tests/test_activitywatch_summary.py
+++ b/tests/test_activitywatch_summary.py
@@ -175,6 +175,19 @@ def test_skill_frontmatter_is_present(self):
         self.assertTrue(skill.startswith("---\nname: computer-usage-summary\n"))
         self.assertIn("description:", skill.split("---", 2)[1])
 
+    def test_skill_default_conversational_output_contract(self):
+        skill = (SKILL_ROOT / "SKILL.md").read_text(encoding="utf-8")
+        headings = [
+            "`Today at a glance`",
+            "`What you did`",
+            "`App table`",
+            "`Caveats`",
+        ]
+        positions = [skill.index(heading) for heading in headings]
+        self.assertEqual(positions, sorted(positions))
+        self.assertIn("unless the user explicitly asks for a simpler or shorter output", skill)
+        self.assertIn("Do not infer a reduced format merely because the request is short.", skill)
+
 
 if __name__ == "__main__":
     unittest.main()