diff --git a/claude/.claude-plugin/plugin.json b/claude/.claude-plugin/plugin.json
index 395d78e..33c11ef 100644
--- a/claude/.claude-plugin/plugin.json
+++ b/claude/.claude-plugin/plugin.json
@@ -1,7 +1,7 @@
 {
   "name": "chatcut",
   "displayName": "ChatCut",
-  "version": "1.10.23",
+  "version": "1.10.24",
   "description": "ChatCut plugin for editable video editing and creation workflows from Claude Code.",
   "author": {
     "name": "ChatCut",
@@ -37,7 +37,7 @@
       "headers": {
         "x-chatcut-mcp-client": "claude_code",
         "x-chatcut-mcp-surface": "embedded-preview",
-        "x-chatcut-plugin-version": "1.10.23"
+        "x-chatcut-plugin-version": "1.10.24"
       }
     }
   },
diff --git a/claude/skills/visual-analysis/SKILL.md b/claude/skills/visual-analysis/SKILL.md
index d27b61c..b1d7bf3 100644
--- a/claude/skills/visual-analysis/SKILL.md
+++ b/claude/skills/visual-analysis/SKILL.md
@@ -1,6 +1,6 @@
 ---
 name: visual-analysis
-description: Verify composed ChatCut edits, compare source frames, locate visual changes, and build labeled contact sheets using the hosted plugin's actual frame delivery.
+description: Analyze videos and references scene by scene, verify composed ChatCut edits, compare source frames, locate visual changes, and build labeled comparisons using the hosted plugin's actual frame delivery.
 ---
 
 # Visual Analysis
@@ -9,6 +9,16 @@ Use fresh pixels to decide visual questions: whether an edit appears, when an
 animation changes, how a crop or overlay composes, or which source shot fits.
 The Codex/Claude host filesystem and ChatCut's backend workspace are separate.
 
+## Whole-video or granular scene analysis
+
+For a video/reference breakdown, read
+[Scene-by-scene analysis](references/scene-breakdown.md). It adds chronological
+visual observations, each scene's words from the existing ChatCut transcript,
+shot/screen-state detail and optional scene JSON.
+Choose this path from the user's question; a focused edit check keeps the
+inspection workflow below. No new provider, queue or user mode parameter is
+required. Use the current host's source/timeline delivery and clock contracts.
+
 ## Keep coordinates and subjects explicit
 
 | Subject                             | Tool and coordinates                                                                          | What the image proves                                                      |
diff --git a/claude/skills/visual-analysis/references/scene-breakdown.md b/claude/skills/visual-analysis/references/scene-breakdown.md
new file mode 100644
index 0000000..13c653a
--- /dev/null
+++ b/claude/skills/visual-analysis/references/scene-breakdown.md
@@ -0,0 +1,127 @@
+# Scene-by-scene analysis
+
+Use this for understanding a video, studying a reference, or producing a granular
+scene/shot breakdown. Keep a focused visual check focused; a subtitle or transition
+question need not generate a report about the whole video.
+
+## Establish the subject and evidence
+
+Resolve the actual source or composed timeline, requested range and useful outcome.
+Reuse readable media, the existing ChatCut transcript and frames. Analyzing a local
+file does not require placing it on a timeline or uploading it. For a URL, use an
+available acquisition route; a thumbnail, page description or successful import
+alone does not establish its content. Report inaccessible media as a source gap.
+
+Read duration and the relevant clock before assigning times. Original footage
+uses source time; a composed edit uses its timeline clock. Follow the entrypoint's
+host-specific field names and units. Keep placement, trim and rate mappings when
+relating the two. The same numeric value can mean different moments on those
+clocks.
+
+## Inspect at the level the question needs
+
+Begin with a bounded overview spanning the selected range, including its opening
+and closing content. Use the current host's source-frame or timeline-viewer tools;
+batch useful samples and divide long videos into meaningful ranges. Chapters,
+transcripts and detected changes can locate candidates, but they do not prove
+that the relevant visuals were inspected.
+
+Expand around consequential changes: a new shot, a changed software screen,
+prompt or setting, an inserted example, an action's start/result, or a transition.
+Inspect ordered before/during/after moments when motion matters. Read small text
+from a useful full-size frame or crop. A few overview images cannot establish
+every cut or a frame-exact boundary; use the available cut/adjacent-frame route
+when that precision changes the answer. Do not inspect every frame by ritual.
+
+## Read speech from the ChatCut transcript
+
+This analysis assumes the media already has a ChatCut transcript. Read each
+scene's timestamped words from it through the host's transcript field or view,
+and put the words actually spoken in that interval into the scene's `audio`. Use
+the transcript of the media's own speech; a translated or reworded caption track
+does not record what was said. Do not transcribe again or reconstruct speech by
+listening. If the transcript is missing, still processing or unreadable for the
+range, report that gap and leave the speech unresolved; creating a transcript is
+the transcription workflow's job.
+
+Keep the transcript's words, including uncertain or misrecognised terms. When
+on-screen text shows the intended spelling, note the correction instead of
+silently rewriting the quote. A sentence can cross a visual cut; assign its words
+to the scene where they are spoken without inventing a word boundary. Summarise
+only when the user asks for less detail, and label the summary as one.
+
+Listen to the audio only for tone, music, sound effects or synchronization, when
+the question depends on them. Without that listening, say so and do not invent
+acoustic descriptions. A transcript's music tag is evidence of the tag, not a
+listening result.
+
+## Describe meaningful scenes
+
+Let visible content and the user's purpose determine segmentation. Separate
+material shot/state changes; retain a coherent continuing action across transcript
+sentences. A tutorial chapter can contain a presenter, a request, a software
+operation and its output—those are different observations when they matter.
+Group an unresolved rapid montage explicitly rather than guessing its cuts.
+There is no target scene count or fixed sampling density.
+
+For each useful interval capture:
+
+- Start/end on the declared clock, at the supported precision.
+- Visible subjects, setting, layout, text, action and camera/graphic movement.
+- The transcript words spoken in the interval, plus any tone, music or sound
+  effects actually listened for.
+- Framing or shot type, and the scene's role in the video when useful.
+
+Prefer specific observations to a repeated generic presenter description. For a
+software demonstration, inspect the important screen, selected inputs, readable
+settings and resulting state. Keep a creator's request separate from its result.
+An output does not reveal its hidden prompt or internal production algorithm.
+
+Keep illegible text, inferred settings and unverified claims uncertain. Reconcile
+material disagreements between narration, visible controls and results. Current
+documentation can explain a tool but cannot prove the creator used it for this
+clip. Recovering the full production method or creating a new video continues in
+the corresponding workflow; a scene list alone does not complete either task.
+
+## Return useful prose or structured scenes
+
+Use a concise narrative/table for an ordinary question. When the user or a
+downstream workflow needs JSON, the following scene fields retain the useful
+granularity of Higgsfield's exposed Video Analysis results without requiring that
+provider, its queue or its private prompt:
+
+```json
+{
+  "time_domain": "source",
+  "analyzed_range": { "start_seconds": 0, "end_seconds": 4 },
+  "scenes": [
+    {
+      "scene_number": 1,
+      "timestamp_start": "00:00:00.000",
+      "timestamp_end": "00:00:04.000",
+      "visual": "A close view shows a cup on a table; a title enters above it.",
+      "audio": "\"Here is the new cup.\" Music and effects not listened to.",
+      "shot_type": "Close-Up",
+      "label": "Product introduction"
+    }
+  ],
+  "evidence": {
+    "timing_precision": "estimated from sampled frames",
+    "audio_basis": "chatcut_transcript",
+    "coverage": "Sampled frames and transcript words across this selected range."
+  }
+}
+```
+
+This is an illustrative shape, not an observation of a supplied video or a claim
+about a supplied video. The seven scene keys are `scene_number`,
+`timestamp_start`, `timestamp_end`, `visual`, `audio`, `shot_type` and `label`.
+Keep source/asset identity, evidence locators, precision and gaps alongside the
+scenes when needed. Choose useful labels for the actual content; advertising
+categories should not force a tutorial or film into a marketing interpretation.
+
+Check chronological intervals against the actual duration and clock, including
+the tail. Internally continuous times can still drift from the source. Cite the
+actual frames/ranges supporting important claims. State partial coverage rather
+than stretching an uninspected final scene. A completed provider job, more scenes,
+valid JSON or distinct wording does not establish correct analysis.
diff --git a/codex/.codex-plugin/plugin.json b/codex/.codex-plugin/plugin.json
index 5d170b4..8ca44e5 100644
--- a/codex/.codex-plugin/plugin.json
+++ b/codex/.codex-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "chatcut",
-  "version": "1.10.21",
+  "version": "1.10.22",
   "description": "ChatCut Codex plugin for editable video editing and creation workflows from Codex.",
   "author": {
     "name": "ChatCut",
diff --git a/codex/.mcp.json b/codex/.mcp.json
index c02fca1..fc685d3 100644
--- a/codex/.mcp.json
+++ b/codex/.mcp.json
@@ -3,7 +3,7 @@
     "chatcut": {
       "http_headers": {
         "x-chatcut-mcp-surface": "codex",
-        "x-chatcut-plugin-version": "1.10.21"
+        "x-chatcut-plugin-version": "1.10.22"
       },
       "oauth_resource": "https://api.chatcut.io/api/external-mcp/mcp",
       "url": "https://api.chatcut.io/api/external-mcp/mcp"
diff --git a/codex/skills/visual-analysis/SKILL.md b/codex/skills/visual-analysis/SKILL.md
index d27b61c..b1d7bf3 100644
--- a/codex/skills/visual-analysis/SKILL.md
+++ b/codex/skills/visual-analysis/SKILL.md
@@ -1,6 +1,6 @@
 ---
 name: visual-analysis
-description: Verify composed ChatCut edits, compare source frames, locate visual changes, and build labeled contact sheets using the hosted plugin's actual frame delivery.
+description: Analyze videos and references scene by scene, verify composed ChatCut edits, compare source frames, locate visual changes, and build labeled comparisons using the hosted plugin's actual frame delivery.
 ---
 
 # Visual Analysis
@@ -9,6 +9,16 @@ Use fresh pixels to decide visual questions: whether an edit appears, when an
 animation changes, how a crop or overlay composes, or which source shot fits.
 The Codex/Claude host filesystem and ChatCut's backend workspace are separate.
 
+## Whole-video or granular scene analysis
+
+For a video/reference breakdown, read
+[Scene-by-scene analysis](references/scene-breakdown.md). It adds chronological
+visual observations, each scene's words from the existing ChatCut transcript,
+shot/screen-state detail and optional scene JSON.
+Choose this path from the user's question; a focused edit check keeps the
+inspection workflow below. No new provider, queue or user mode parameter is
+required. Use the current host's source/timeline delivery and clock contracts.
+
 ## Keep coordinates and subjects explicit
 
 | Subject                             | Tool and coordinates                                                                          | What the image proves                                                      |
diff --git a/codex/skills/visual-analysis/references/scene-breakdown.md b/codex/skills/visual-analysis/references/scene-breakdown.md
new file mode 100644
index 0000000..13c653a
--- /dev/null
+++ b/codex/skills/visual-analysis/references/scene-breakdown.md
@@ -0,0 +1,127 @@
+# Scene-by-scene analysis
+
+Use this for understanding a video, studying a reference, or producing a granular
+scene/shot breakdown. Keep a focused visual check focused; a subtitle or transition
+question need not generate a report about the whole video.
+
+## Establish the subject and evidence
+
+Resolve the actual source or composed timeline, requested range and useful outcome.
+Reuse readable media, the existing ChatCut transcript and frames. Analyzing a local
+file does not require placing it on a timeline or uploading it. For a URL, use an
+available acquisition route; a thumbnail, page description or successful import
+alone does not establish its content. Report inaccessible media as a source gap.
+
+Read duration and the relevant clock before assigning times. Original footage
+uses source time; a composed edit uses its timeline clock. Follow the entrypoint's
+host-specific field names and units. Keep placement, trim and rate mappings when
+relating the two. The same numeric value can mean different moments on those
+clocks.
+
+## Inspect at the level the question needs
+
+Begin with a bounded overview spanning the selected range, including its opening
+and closing content. Use the current host's source-frame or timeline-viewer tools;
+batch useful samples and divide long videos into meaningful ranges. Chapters,
+transcripts and detected changes can locate candidates, but they do not prove
+that the relevant visuals were inspected.
+
+Expand around consequential changes: a new shot, a changed software screen,
+prompt or setting, an inserted example, an action's start/result, or a transition.
+Inspect ordered before/during/after moments when motion matters. Read small text
+from a useful full-size frame or crop. A few overview images cannot establish
+every cut or a frame-exact boundary; use the available cut/adjacent-frame route
+when that precision changes the answer. Do not inspect every frame by ritual.
+
+## Read speech from the ChatCut transcript
+
+This analysis assumes the media already has a ChatCut transcript. Read each
+scene's timestamped words from it through the host's transcript field or view,
+and put the words actually spoken in that interval into the scene's `audio`. Use
+the transcript of the media's own speech; a translated or reworded caption track
+does not record what was said. Do not transcribe again or reconstruct speech by
+listening. If the transcript is missing, still processing or unreadable for the
+range, report that gap and leave the speech unresolved; creating a transcript is
+the transcription workflow's job.
+
+Keep the transcript's words, including uncertain or misrecognised terms. When
+on-screen text shows the intended spelling, note the correction instead of
+silently rewriting the quote. A sentence can cross a visual cut; assign its words
+to the scene where they are spoken without inventing a word boundary. Summarise
+only when the user asks for less detail, and label the summary as one.
+
+Listen to the audio only for tone, music, sound effects or synchronization, when
+the question depends on them. Without that listening, say so and do not invent
+acoustic descriptions. A transcript's music tag is evidence of the tag, not a
+listening result.
+
+## Describe meaningful scenes
+
+Let visible content and the user's purpose determine segmentation. Separate
+material shot/state changes; retain a coherent continuing action across transcript
+sentences. A tutorial chapter can contain a presenter, a request, a software
+operation and its output—those are different observations when they matter.
+Group an unresolved rapid montage explicitly rather than guessing its cuts.
+There is no target scene count or fixed sampling density.
+
+For each useful interval capture:
+
+- Start/end on the declared clock, at the supported precision.
+- Visible subjects, setting, layout, text, action and camera/graphic movement.
+- The transcript words spoken in the interval, plus any tone, music or sound
+  effects actually listened for.
+- Framing or shot type, and the scene's role in the video when useful.
+
+Prefer specific observations to a repeated generic presenter description. For a
+software demonstration, inspect the important screen, selected inputs, readable
+settings and resulting state. Keep a creator's request separate from its result.
+An output does not reveal its hidden prompt or internal production algorithm.
+
+Keep illegible text, inferred settings and unverified claims uncertain. Reconcile
+material disagreements between narration, visible controls and results. Current
+documentation can explain a tool but cannot prove the creator used it for this
+clip. Recovering the full production method or creating a new video continues in
+the corresponding workflow; a scene list alone does not complete either task.
+
+## Return useful prose or structured scenes
+
+Use a concise narrative/table for an ordinary question. When the user or a
+downstream workflow needs JSON, the following scene fields retain the useful
+granularity of Higgsfield's exposed Video Analysis results without requiring that
+provider, its queue or its private prompt:
+
+```json
+{
+  "time_domain": "source",
+  "analyzed_range": { "start_seconds": 0, "end_seconds": 4 },
+  "scenes": [
+    {
+      "scene_number": 1,
+      "timestamp_start": "00:00:00.000",
+      "timestamp_end": "00:00:04.000",
+      "visual": "A close view shows a cup on a table; a title enters above it.",
+      "audio": "\"Here is the new cup.\" Music and effects not listened to.",
+      "shot_type": "Close-Up",
+      "label": "Product introduction"
+    }
+  ],
+  "evidence": {
+    "timing_precision": "estimated from sampled frames",
+    "audio_basis": "chatcut_transcript",
+    "coverage": "Sampled frames and transcript words across this selected range."
+  }
+}
+```
+
+This is an illustrative shape, not an observation of a supplied video or a claim
+about a supplied video. The seven scene keys are `scene_number`,
+`timestamp_start`, `timestamp_end`, `visual`, `audio`, `shot_type` and `label`.
+Keep source/asset identity, evidence locators, precision and gaps alongside the
+scenes when needed. Choose useful labels for the actual content; advertising
+categories should not force a tutorial or film into a marketing interpretation.
+
+Check chronological intervals against the actual duration and clock, including
+the tail. Internally continuous times can still drift from the source. Cite the
+actual frames/ranges supporting important claims. State partial coverage rather
+than stretching an uninspected final scene. A completed provider job, more scenes,
+valid JSON or distinct wording does not establish correct analysis.
diff --git a/grok/.cursor-plugin/plugin.json b/grok/.cursor-plugin/plugin.json
index 418952e..2a48198 100644
--- a/grok/.cursor-plugin/plugin.json
+++ b/grok/.cursor-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "chatcut",
-  "version": "1.10.17",
+  "version": "1.10.18",
   "description": "Let Grok Bot agents create and edit videos in ChatCut through natural-language requests, with editable timelines, captions, motion graphics, media generation, and export.",
   "author": {
     "name": "ChatCut",
diff --git a/grok/mcp.json b/grok/mcp.json
index 9f00808..f01d470 100644
--- a/grok/mcp.json
+++ b/grok/mcp.json
@@ -4,7 +4,7 @@
       "url": "https://api.chatcut.io/api/external-mcp/mcp",
       "headers": {
         "x-chatcut-mcp-surface": "embedded-preview",
-        "x-chatcut-plugin-version": "1.10.17"
+        "x-chatcut-plugin-version": "1.10.18"
       }
     }
   }
diff --git a/grok/skills/verification/SKILL.md b/grok/skills/verification/SKILL.md
index 18b7273..0e67361 100644
--- a/grok/skills/verification/SKILL.md
+++ b/grok/skills/verification/SKILL.md
@@ -5,6 +5,10 @@ description: Hosted ChatCut plugin sessions only (the `chatcut` MCP server). If
 
 # Verification
 
+For a video/reference breakdown or detailed frame analysis, use the bundled
+[visual-analysis skill](../visual-analysis/SKILL.md), then continue the checks
+below when verifying an edit.
+
 Use both structural and visual evidence when the requested result is visible:
 
 1. Re-read the affected scope with `read_project`, `preview_timeline`, `inspect_item`, `browse_assets`, or `inspect_asset` as appropriate. Use the latest response for ids, tracks, timing, and readiness.
diff --git a/grok/skills/visual-analysis/SKILL.md b/grok/skills/visual-analysis/SKILL.md
new file mode 100644
index 0000000..8c4084a
--- /dev/null
+++ b/grok/skills/visual-analysis/SKILL.md
@@ -0,0 +1,106 @@
+---
+name: visual-analysis
+description: Analyze videos and references scene by scene, verify composed ChatCut edits, compare source frames, locate visual changes, and build labeled comparisons using the hosted plugin's actual frame delivery.
+---
+
+# Visual Analysis
+
+Use fresh pixels to decide visual questions: whether an edit appears, when an
+animation changes, how a crop or overlay composes, or which source shot fits.
+The agent host filesystem and ChatCut's backend workspace are separate.
+
+## Whole-video or granular scene analysis
+
+For a video/reference breakdown, read
+[Scene-by-scene analysis](references/scene-breakdown.md). It adds chronological
+visual observations, each scene's words from the existing ChatCut transcript,
+shot/screen-state detail and optional scene JSON.
+Choose this path from the user's question; a focused edit check keeps the
+inspection workflow below. No new provider, queue or user mode parameter is
+required. Use the current host's source/timeline delivery and clock contracts.
+
+## Keep coordinates and subjects explicit
+
+| Subject                             | Tool and coordinates                                                                          | What the image proves                                                      |
+| ----------------------------------- | --------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------- |
+| Placed timeline composition         | `preview_timeline`, `views:["viewer"]`, `viewerFrames` in timeline frames                     | Clips with their placement, crop, captions, overlays, effects, and framing |
+| Original video / Motion Graphic     | `inspect_asset`, `sourceFrameCount` for an overview or `sourceTimesMs` in source milliseconds | The source media at those positions                                        |
+| Original image / SVG                | `inspect_asset` with its default image preview                                                | The source image, without timeline placement                               |
+| Readable original file in this host | The host's local inspection tools                                                             | Source bytes; no temporary timeline or cloud upload is needed              |
+
+Use `timelineId` when scope could be ambiguous. Viewer metadata reports the
+actual `frame`, `timelineTimeMs`, and FPS. Source metadata reports `sourceTimeMs`.
+Do not label a source timestamp as timeline time. Read a placed item's start,
+source offset, playback rate and timeline FPS before mapping between them.
+
+Timeline ranges are half-open `[fromFrame,toFrame)`. Frames at `toFrame` or the
+duration are outside the renderable range. The hosted tool may skip invalid
+requested moments and report adjustments; it fails when none are renderable.
+Check the returned frames rather than assuming every request was rendered.
+`sourceTimesMs` and `sourceFrameCount` are alternatives; bounded transcript
+context uses the separate `transcriptRangesMs` field.
+
+## Read the actual delivery
+
+Preview tools can return an inline image, temporary per-frame `resource_link`
+URIs, or written workspace paths on a workspace-backed runtime. Hosted plugins
+have no backend workspace. Inspect inline images directly when the host shows
+them. For image links, use the full returned URI with a host-supported image
+reader, or an allowed download into the host's own analysis directory. Never
+infer local files from names such as `timeline-frames/...`, and never treat a
+backend `workspacePath` as a path on the user's computer.
+
+If rendering is unavailable, read its diagnostic. Source-frame requests on
+this connector need the target project open in a capable editor; there is no
+automatic cloud fallback. Inspect a readable original locally when that is the
+subject, or surface the returned editor URL for composed proof. A metadata read
+or successful tool call without visible pixels does not establish appearance.
+
+## Narrow a change
+
+Start with a frame before and after the suspected change. Sample a few moments
+between them, then refine the adjacent pair that brackets the difference.
+Stop at adjacent timeline frames when exact onset matters. For source footage,
+refine source milliseconds instead. An overview proves only sampled moments.
+
+Read full frames before cropping. Use the same crop rectangle and scale for
+comparisons, and label the returned frame/time, timeline or source identity.
+Keep chronological order left-to-right, then top-to-bottom. Render fresh
+evidence after editing; an old contact sheet describes the old state.
+
+## Straightening across shots
+
+A clip can contain different camera angles. Choose a reliable level reference
+for each affected shot; a doorway viewed in perspective is not enough to infer
+camera roll. Apply one rotation to the whole clip only when it suits all shots;
+otherwise limit the correction to the intended shot or range and preserve the
+others. If the intended reference is unclear, clarify that choice before editing.
+
+Compare matching timeline frames before and after the correction, including
+shots that should stay unchanged. Check level, framing and cropped edges together.
+Await the edit before requesting its preview, and inspect fresh pixels after an
+undo too. Report straightening or restoration only to the extent those frames
+verify it.
+
+## Contact sheets from host-local images
+
+Use the tool's returned labeled sheet when available. For custom comparisons,
+the shared [concat-images helper](scripts/concat-images.mjs) accepts images that
+already exist on the host. It needs Node and an available ImageMagick command;
+if those are absent, use the host's existing image tools or inspect frames
+individually. This skill does not require a new package installation.
+
+```bash
+node <this-skill-dir>/scripts/concat-images.mjs \
+  --output <host-analysis-dir>/title-change.jpg --columns 3 \
+  --caption "timeline frame 120" --caption "timeline frame 123" \
+  --caption "timeline frame 126" \
+  <host-analysis-dir>/frame-120.jpg <host-analysis-dir>/frame-123.jpg \
+  <host-analysis-dir>/frame-126.jpg
+```
+
+`--crop WIDTHxHEIGHT+X+Y` applies a common pixel crop; `--tile-width` limits tile
+size. Read the helper's JSON and inspect the resulting `outputPath` using the
+host's image viewer. Label strips are analysis annotations, not video content.
+Analysis images stay separate from project media; import one through
+[asset-import](../asset-import/SKILL.md) only if the user wants to keep or use it.
diff --git a/grok/skills/visual-analysis/references/scene-breakdown.md b/grok/skills/visual-analysis/references/scene-breakdown.md
new file mode 100644
index 0000000..13c653a
--- /dev/null
+++ b/grok/skills/visual-analysis/references/scene-breakdown.md
@@ -0,0 +1,127 @@
+# Scene-by-scene analysis
+
+Use this for understanding a video, studying a reference, or producing a granular
+scene/shot breakdown. Keep a focused visual check focused; a subtitle or transition
+question need not generate a report about the whole video.
+
+## Establish the subject and evidence
+
+Resolve the actual source or composed timeline, requested range and useful outcome.
+Reuse readable media, the existing ChatCut transcript and frames. Analyzing a local
+file does not require placing it on a timeline or uploading it. For a URL, use an
+available acquisition route; a thumbnail, page description or successful import
+alone does not establish its content. Report inaccessible media as a source gap.
+
+Read duration and the relevant clock before assigning times. Original footage
+uses source time; a composed edit uses its timeline clock. Follow the entrypoint's
+host-specific field names and units. Keep placement, trim and rate mappings when
+relating the two. The same numeric value can mean different moments on those
+clocks.
+
+## Inspect at the level the question needs
+
+Begin with a bounded overview spanning the selected range, including its opening
+and closing content. Use the current host's source-frame or timeline-viewer tools;
+batch useful samples and divide long videos into meaningful ranges. Chapters,
+transcripts and detected changes can locate candidates, but they do not prove
+that the relevant visuals were inspected.
+
+Expand around consequential changes: a new shot, a changed software screen,
+prompt or setting, an inserted example, an action's start/result, or a transition.
+Inspect ordered before/during/after moments when motion matters. Read small text
+from a useful full-size frame or crop. A few overview images cannot establish
+every cut or a frame-exact boundary; use the available cut/adjacent-frame route
+when that precision changes the answer. Do not inspect every frame by ritual.
+
+## Read speech from the ChatCut transcript
+
+This analysis assumes the media already has a ChatCut transcript. Read each
+scene's timestamped words from it through the host's transcript field or view,
+and put the words actually spoken in that interval into the scene's `audio`. Use
+the transcript of the media's own speech; a translated or reworded caption track
+does not record what was said. Do not transcribe again or reconstruct speech by
+listening. If the transcript is missing, still processing or unreadable for the
+range, report that gap and leave the speech unresolved; creating a transcript is
+the transcription workflow's job.
+
+Keep the transcript's words, including uncertain or misrecognised terms. When
+on-screen text shows the intended spelling, note the correction instead of
+silently rewriting the quote. A sentence can cross a visual cut; assign its words
+to the scene where they are spoken without inventing a word boundary. Summarise
+only when the user asks for less detail, and label the summary as one.
+
+Listen to the audio only for tone, music, sound effects or synchronization, when
+the question depends on them. Without that listening, say so and do not invent
+acoustic descriptions. A transcript's music tag is evidence of the tag, not a
+listening result.
+
+## Describe meaningful scenes
+
+Let visible content and the user's purpose determine segmentation. Separate
+material shot/state changes; retain a coherent continuing action across transcript
+sentences. A tutorial chapter can contain a presenter, a request, a software
+operation and its output—those are different observations when they matter.
+Group an unresolved rapid montage explicitly rather than guessing its cuts.
+There is no target scene count or fixed sampling density.
+
+For each useful interval capture:
+
+- Start/end on the declared clock, at the supported precision.
+- Visible subjects, setting, layout, text, action and camera/graphic movement.
+- The transcript words spoken in the interval, plus any tone, music or sound
+  effects actually listened for.
+- Framing or shot type, and the scene's role in the video when useful.
+
+Prefer specific observations to a repeated generic presenter description. For a
+software demonstration, inspect the important screen, selected inputs, readable
+settings and resulting state. Keep a creator's request separate from its result.
+An output does not reveal its hidden prompt or internal production algorithm.
+
+Keep illegible text, inferred settings and unverified claims uncertain. Reconcile
+material disagreements between narration, visible controls and results. Current
+documentation can explain a tool but cannot prove the creator used it for this
+clip. Recovering the full production method or creating a new video continues in
+the corresponding workflow; a scene list alone does not complete either task.
+
+## Return useful prose or structured scenes
+
+Use a concise narrative/table for an ordinary question. When the user or a
+downstream workflow needs JSON, the following scene fields retain the useful
+granularity of Higgsfield's exposed Video Analysis results without requiring that
+provider, its queue or its private prompt:
+
+```json
+{
+  "time_domain": "source",
+  "analyzed_range": { "start_seconds": 0, "end_seconds": 4 },
+  "scenes": [
+    {
+      "scene_number": 1,
+      "timestamp_start": "00:00:00.000",
+      "timestamp_end": "00:00:04.000",
+      "visual": "A close view shows a cup on a table; a title enters above it.",
+      "audio": "\"Here is the new cup.\" Music and effects not listened to.",
+      "shot_type": "Close-Up",
+      "label": "Product introduction"
+    }
+  ],
+  "evidence": {
+    "timing_precision": "estimated from sampled frames",
+    "audio_basis": "chatcut_transcript",
+    "coverage": "Sampled frames and transcript words across this selected range."
+  }
+}
+```
+
+This is an illustrative shape, not an observation of a supplied video or a claim
+about a supplied video. The seven scene keys are `scene_number`,
+`timestamp_start`, `timestamp_end`, `visual`, `audio`, `shot_type` and `label`.
+Keep source/asset identity, evidence locators, precision and gaps alongside the
+scenes when needed. Choose useful labels for the actual content; advertising
+categories should not force a tutorial or film into a marketing interpretation.
+
+Check chronological intervals against the actual duration and clock, including
+the tail. Internally continuous times can still drift from the source. Cite the
+actual frames/ranges supporting important claims. State partial coverage rather
+than stretching an uninspected final scene. A completed provider job, more scenes,
+valid JSON or distinct wording does not establish correct analysis.
diff --git a/grok/skills/visual-analysis/scripts/concat-images.mjs b/grok/skills/visual-analysis/scripts/concat-images.mjs
new file mode 100644
index 0000000..56be49b
--- /dev/null
+++ b/grok/skills/visual-analysis/scripts/concat-images.mjs
@@ -0,0 +1,305 @@
+#!/usr/bin/env node
+
+import { existsSync, mkdirSync, mkdtempSync, rmSync } from "node:fs";
+import { basename, dirname, extname, join, resolve } from "node:path";
+import { spawnSync } from "node:child_process";
+
+const parsed = parseArgs(process.argv.slice(2));
+const outputPath = resolve(parsed.output);
+const outputExtension = extname(outputPath).toLowerCase();
+
+if (![".jpg", ".jpeg", ".png", ".webp"].includes(outputExtension)) {
+  fail("--output must end in .jpg, .jpeg, .png, or .webp");
+}
+if (parsed.inputs.length === 0 || parsed.inputs.length > 25) {
+  fail("Pass between 1 and 25 input images");
+}
+if (parsed.captions.length && parsed.captions.length !== parsed.inputs.length) {
+  fail("Pass exactly one --caption per input image, or omit captions entirely");
+}
+
+const inputs = parsed.inputs.map((input) => resolve(input));
+for (const input of inputs) {
+  if (!existsSync(input)) fail(`Input image does not exist: ${input}`);
+}
+
+mkdirSync(dirname(outputPath), { recursive: true });
+const temporaryDirectory = mkdtempSync(
+  join(dirname(outputPath), ".concat-images-"),
+);
+const imageMagick = resolveImageMagick();
+// Some ImageMagick builds cannot discover installed fonts without fontconfig.
+const captions = parsed.captions.length
+  ? parsed.captions
+  : inputs.map(defaultCaption);
+const captionFont = [
+  ...(captions.every((caption) => /^[\x00-\x7F]*$/.test(caption))
+    ? [
+        "/System/Library/Fonts/Menlo.ttc",
+        "/usr/share/fonts/truetype/dejavu/DejaVuSansMono.ttf",
+      ]
+    : []),
+  "/System/Library/Fonts/Supplemental/Arial Unicode.ttf",
+  "/System/Library/Fonts/Supplemental/Arial.ttf",
+  "/usr/share/fonts/opentype/noto/NotoSansCJK-Regular.ttc",
+  "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",
+  join(process.env.WINDIR ?? "C:\\Windows", "Fonts", "arial.ttf"),
+].find((font) => existsSync(font));
+
+try {
+  const prepared = inputs.map((input, index) => {
+    const preparedPath = join(temporaryDirectory, `${index}.png`);
+    const args = imageMagick.prefix.concat([
+      input,
+      ...(parsed.crop ? ["-crop", parsed.crop, "+repage"] : []),
+      "-auto-orient",
+      "+profile",
+      "icc",
+      "-resize",
+      `${parsed.tileWidth}x${parsed.tileWidth}>`,
+      preparedPath,
+    ]);
+    run(imageMagick.command, args, `prepare ${input}`);
+    return preparedPath;
+  });
+
+  const dimensions = run(
+    imageMagick.command,
+    imageMagick.prefix.concat(prepared, ["-format", "%w %h\n", "info:"]),
+    "read image dimensions",
+  )
+    .trim()
+    .split("\n")
+    .map((line) => line.split(" ").map(Number));
+  const aspect =
+    Math.max(...dimensions.map(([width]) => width)) /
+    Math.max(...dimensions.map(([, height]) => height));
+  const captionHeight = parsed.noCaptions ? 0 : 28;
+  let columns = parsed.columns ?? 1;
+  let tileWidth = 0;
+  for (
+    let candidate = 1;
+    candidate <= Math.max(inputs.length, parsed.columns ?? 1);
+    candidate++
+  ) {
+    if (parsed.columns && candidate !== parsed.columns) continue;
+    const rows = Math.ceil(inputs.length / candidate);
+    const width = Math.floor(
+      Math.min(
+        parsed.tileWidth,
+        1568 / candidate - 2 * parsed.gap,
+        (Math.floor(1568 / rows) - 2 * parsed.gap - captionHeight) * aspect,
+      ),
+    );
+    if (width > tileWidth) {
+      tileWidth = width;
+      columns = candidate;
+    }
+  }
+  if (tileWidth < 1) fail("The requested columns/gap leave no room for images");
+  const tileHeight = Math.max(1, Math.floor(tileWidth / aspect));
+  const captioned = prepared.map((input, index) => {
+    const output = join(temporaryDirectory, `captioned-${index}.png`);
+    run(
+      imageMagick.command,
+      imageMagick.prefix.concat([
+        input,
+        "-resize",
+        `${tileWidth}x${tileHeight}`,
+        "-background",
+        "#000000",
+        "-gravity",
+        "center",
+        "-extent",
+        `${tileWidth}x${tileHeight}`,
+        ...(parsed.noCaptions
+          ? []
+          : [
+              "(",
+              "-size",
+              `${tileWidth}x${captionHeight}`,
+              "xc:#eeeeee",
+              ...(captionFont ? ["-font", captionFont] : []),
+              "-fill",
+              "#111111",
+              "-stroke",
+              "none",
+              "-pointsize",
+              "18",
+              "-gravity",
+              "west",
+              "-annotate",
+              "+5+0",
+              captions[index],
+              ")",
+              "-append",
+            ]),
+        output,
+      ]),
+      "add caption below image",
+    );
+    return output;
+  });
+  const montageArgs = imageMagick.montagePrefix.concat([
+    ...(captionFont ? ["-font", captionFont] : []),
+    ...captioned,
+    "-tile",
+    `${columns}x`,
+    "-geometry",
+    `+${parsed.gap}+${parsed.gap}`,
+    "-background",
+    "#eeeeee",
+    outputPath,
+  ]);
+  run(imageMagick.command, montageArgs, "compose contact sheet");
+
+  process.stdout.write(
+    `${JSON.stringify(
+      {
+        columns,
+        crop: parsed.crop ?? null,
+        inputs: inputs.map((input, index) => ({
+          caption: captions[index],
+          path: input,
+        })),
+        ...(parsed.noCaptions
+          ? {}
+          : {
+              note: "The light-gray strips below the images contain tool-added position labels, not video content.",
+            }),
+        width: columns * (tileWidth + 2 * parsed.gap),
+        height:
+          Math.ceil(inputs.length / columns) *
+          (tileHeight + captionHeight + 2 * parsed.gap),
+        outputPath,
+        rows: Math.ceil(inputs.length / columns),
+      },
+      null,
+      2,
+    )}\n`,
+  );
+} finally {
+  rmSync(temporaryDirectory, { force: true, recursive: true });
+}
+
+// Frame filenames carry the exact coordinate returned by the preview tools.
+function defaultCaption(input) {
+  const name = basename(input, extname(input));
+  const timeline =
+    name.match(/^timeline-(\d+)f-\d+ms-\d+$/) ??
+    name.match(/^timeline-frame-(\d+)$/);
+  if (timeline) return `f${timeline[1]}`;
+  const source =
+    name.match(/^source-(\d+)ms-\d+$/) ?? name.match(/^asset-frame-(\d+)ms$/);
+  if (source) return `${(Number(source[1]) / 1000).toFixed(3)}s`;
+  return basename(input);
+}
+
+function parseArgs(args) {
+  const result = {
+    noCaptions: false,
+    captions: [],
+    columns: undefined,
+    crop: undefined,
+    gap: 0,
+    inputs: [],
+    output: undefined,
+    tileWidth: 720,
+  };
+  for (let index = 0; index < args.length; index += 1) {
+    const arg = args[index];
+    if (arg === "--no-captions") result.noCaptions = true;
+    else if (arg === "--caption")
+      result.captions.push(requiredValue(args, ++index, arg));
+    else if (arg === "--columns")
+      result.columns = boundedInteger(
+        requiredValue(args, ++index, arg),
+        arg,
+        1,
+        9,
+      );
+    else if (arg === "--crop") result.crop = requiredValue(args, ++index, arg);
+    else if (arg === "--gap")
+      result.gap = boundedInteger(
+        requiredValue(args, ++index, arg),
+        arg,
+        0,
+        100,
+      );
+    else if (arg === "--output")
+      result.output = requiredValue(args, ++index, arg);
+    else if (arg === "--tile-width")
+      result.tileWidth = boundedInteger(
+        requiredValue(args, ++index, arg),
+        arg,
+        100,
+        2000,
+      );
+    else if (arg === "--help" || arg === "-h") usage(0);
+    else if (arg.startsWith("-")) fail(`Unknown option: ${arg}`);
+    else result.inputs.push(arg);
+  }
+  if (!result.output) fail("--output is required");
+  if (result.crop && !/^\d+x\d+\+\d+\+\d+$/.test(result.crop)) {
+    fail("--crop must use WIDTHxHEIGHT+X+Y pixel geometry");
+  }
+  return result;
+}
+
+function resolveImageMagick() {
+  if (commandExists("magick")) {
+    return { command: "magick", montagePrefix: ["montage"], prefix: [] };
+  }
+  if (commandExists("convert") && commandExists("montage")) {
+    return { command: "convert", montagePrefix: [], prefix: [] };
+  }
+  fail("ImageMagick is unavailable; expected magick or convert + montage");
+}
+
+function commandExists(command) {
+  return (
+    spawnSync("sh", ["-c", `command -v ${command}`], {
+      stdio: "ignore",
+    }).status === 0
+  );
+}
+
+function run(command, args, step) {
+  const executable =
+    step === "compose contact sheet" && command === "convert"
+      ? "montage"
+      : command;
+  const result = spawnSync(executable, args, { encoding: "utf8" });
+  if (result.status !== 0) {
+    fail(
+      `${step} failed: ${(result.stderr || result.stdout || "unknown error").trim()}`,
+    );
+  }
+  return result.stdout;
+}
+
+function requiredValue(args, index, option) {
+  const value = args[index];
+  if (!value) fail(`${option} requires a value`);
+  return value;
+}
+
+function boundedInteger(value, option, minimum, maximum) {
+  const parsed = Number(value);
+  if (!Number.isInteger(parsed) || parsed < minimum || parsed > maximum) {
+    fail(`${option} must be an integer from ${minimum} to ${maximum}`);
+  }
+  return parsed;
+}
+
+function fail(message) {
+  process.stderr.write(`concat-images: ${message}\n`);
+  process.exit(1);
+}
+
+function usage(exitCode) {
+  process.stdout.write(
+    "Usage: concat-images.mjs --output <image> [--columns N] [--crop WxH+X+Y] [--caption text ...] <images...>\n",
+  );
+  process.exit(exitCode);
+}
diff --git a/kimi/.mcp.json b/kimi/.mcp.json
index c0c2ee2..c7e48c8 100644
--- a/kimi/.mcp.json
+++ b/kimi/.mcp.json
@@ -8,7 +8,7 @@
       "http_headers": {
         "x-chatcut-mcp-client": "kimi_work",
         "x-chatcut-mcp-surface": "embedded-preview",
-        "x-chatcut-plugin-version": "1.10.17"
+        "x-chatcut-plugin-version": "1.10.18"
       },
       "note": "Official ChatCut remote MCP endpoint. OAuth 2.0 with dynamic client registration (DCR) + PKCE; the server derives userId from the token.",
       "authorization": "OAuth 2.0 with dynamic client registration",
diff --git a/kimi/kimi.plugin.json b/kimi/kimi.plugin.json
index a8d16b2..52a9c46 100644
--- a/kimi/kimi.plugin.json
+++ b/kimi/kimi.plugin.json
@@ -1,7 +1,7 @@
 {
   "$schema": "https://catalog.msh.team/misc/kimi.plugin.schema.json",
   "name": "chatcut-kimi-work",
-  "version": "1.10.17",
+  "version": "1.10.18",
   "description": "ChatCut plugin for Kimi Work: let the agent edit speech, create motion graphics and media, and deliver editable video projects in the live in-app editor.",
   "author": "ChatCut",
   "homepage": "https://chatcut.io",
@@ -40,7 +40,7 @@
       "headers": {
         "x-chatcut-mcp-client": "kimi_work",
         "x-chatcut-mcp-surface": "embedded-preview",
-        "x-chatcut-plugin-version": "1.10.17"
+        "x-chatcut-plugin-version": "1.10.18"
       }
     }
   }
diff --git a/kimi/skills/chatcut-verification-guide/SKILL.md b/kimi/skills/chatcut-verification-guide/SKILL.md
index efe5533..d412985 100644
--- a/kimi/skills/chatcut-verification-guide/SKILL.md
+++ b/kimi/skills/chatcut-verification-guide/SKILL.md
@@ -5,6 +5,10 @@ description: Use after ChatCut edits, imports, captions, generation, or exports
 
 # ChatCut Verification Guide
 
+For a video/reference breakdown or detailed frame analysis, use the bundled
+[visual-analysis skill](../visual-analysis/SKILL.md), then continue the checks
+below when verifying an edit.
+
 Verify through ChatCut project state first and the in-app browser second.
 
 ## Checks
diff --git a/kimi/skills/visual-analysis/SKILL.md b/kimi/skills/visual-analysis/SKILL.md
new file mode 100644
index 0000000..346c2f6
--- /dev/null
+++ b/kimi/skills/visual-analysis/SKILL.md
@@ -0,0 +1,106 @@
+---
+name: visual-analysis
+description: Analyze videos and references scene by scene, verify composed ChatCut edits, compare source frames, locate visual changes, and build labeled comparisons using the hosted plugin's actual frame delivery.
+---
+
+# Visual Analysis
+
+Use fresh pixels to decide visual questions: whether an edit appears, when an
+animation changes, how a crop or overlay composes, or which source shot fits.
+The agent host filesystem and ChatCut's backend workspace are separate.
+
+## Whole-video or granular scene analysis
+
+For a video/reference breakdown, read
+[Scene-by-scene analysis](references/scene-breakdown.md). It adds chronological
+visual observations, each scene's words from the existing ChatCut transcript,
+shot/screen-state detail and optional scene JSON.
+Choose this path from the user's question; a focused edit check keeps the
+inspection workflow below. No new provider, queue or user mode parameter is
+required. Use the current host's source/timeline delivery and clock contracts.
+
+## Keep coordinates and subjects explicit
+
+| Subject                             | Tool and coordinates                                                                          | What the image proves                                                      |
+| ----------------------------------- | --------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------- |
+| Placed timeline composition         | `preview_timeline`, `views:["viewer"]`, `viewerFrames` in timeline frames                     | Clips with their placement, crop, captions, overlays, effects, and framing |
+| Original video / Motion Graphic     | `inspect_asset`, `sourceFrameCount` for an overview or `sourceTimesMs` in source milliseconds | The source media at those positions                                        |
+| Original image / SVG                | `inspect_asset` with its default image preview                                                | The source image, without timeline placement                               |
+| Readable original file in this host | The host's local inspection tools                                                             | Source bytes; no temporary timeline or cloud upload is needed              |
+
+Use `timelineId` when scope could be ambiguous. Viewer metadata reports the
+actual `frame`, `timelineTimeMs`, and FPS. Source metadata reports `sourceTimeMs`.
+Do not label a source timestamp as timeline time. Read a placed item's start,
+source offset, playback rate and timeline FPS before mapping between them.
+
+Timeline ranges are half-open `[fromFrame,toFrame)`. Frames at `toFrame` or the
+duration are outside the renderable range. The hosted tool may skip invalid
+requested moments and report adjustments; it fails when none are renderable.
+Check the returned frames rather than assuming every request was rendered.
+`sourceTimesMs` and `sourceFrameCount` are alternatives; bounded transcript
+context uses the separate `transcriptRangesMs` field.
+
+## Read the actual delivery
+
+Preview tools can return an inline image, temporary per-frame `resource_link`
+URIs, or written workspace paths on a workspace-backed runtime. Hosted plugins
+have no backend workspace. Inspect inline images directly when the host shows
+them. For image links, use the full returned URI with a host-supported image
+reader, or an allowed download into the host's own analysis directory. Never
+infer local files from names such as `timeline-frames/...`, and never treat a
+backend `workspacePath` as a path on the user's computer.
+
+If rendering is unavailable, read its diagnostic. Source-frame requests on
+this connector need the target project open in a capable editor; there is no
+automatic cloud fallback. Inspect a readable original locally when that is the
+subject, or surface the returned editor URL for composed proof. A metadata read
+or successful tool call without visible pixels does not establish appearance.
+
+## Narrow a change
+
+Start with a frame before and after the suspected change. Sample a few moments
+between them, then refine the adjacent pair that brackets the difference.
+Stop at adjacent timeline frames when exact onset matters. For source footage,
+refine source milliseconds instead. An overview proves only sampled moments.
+
+Read full frames before cropping. Use the same crop rectangle and scale for
+comparisons, and label the returned frame/time, timeline or source identity.
+Keep chronological order left-to-right, then top-to-bottom. Render fresh
+evidence after editing; an old contact sheet describes the old state.
+
+## Straightening across shots
+
+A clip can contain different camera angles. Choose a reliable level reference
+for each affected shot; a doorway viewed in perspective is not enough to infer
+camera roll. Apply one rotation to the whole clip only when it suits all shots;
+otherwise limit the correction to the intended shot or range and preserve the
+others. If the intended reference is unclear, clarify that choice before editing.
+
+Compare matching timeline frames before and after the correction, including
+shots that should stay unchanged. Check level, framing and cropped edges together.
+Await the edit before requesting its preview, and inspect fresh pixels after an
+undo too. Report straightening or restoration only to the extent those frames
+verify it.
+
+## Contact sheets from host-local images
+
+Use the tool's returned labeled sheet when available. For custom comparisons,
+the shared [concat-images helper](scripts/concat-images.mjs) accepts images that
+already exist on the host. It needs Node and an available ImageMagick command;
+if those are absent, use the host's existing image tools or inspect frames
+individually. This skill does not require a new package installation.
+
+```bash
+node <this-skill-dir>/scripts/concat-images.mjs \
+  --output <host-analysis-dir>/title-change.jpg --columns 3 \
+  --caption "timeline frame 120" --caption "timeline frame 123" \
+  --caption "timeline frame 126" \
+  <host-analysis-dir>/frame-120.jpg <host-analysis-dir>/frame-123.jpg \
+  <host-analysis-dir>/frame-126.jpg
+```
+
+`--crop WIDTHxHEIGHT+X+Y` applies a common pixel crop; `--tile-width` limits tile
+size. Read the helper's JSON and inspect the resulting `outputPath` using the
+host's image viewer. Label strips are analysis annotations, not video content.
+Analysis images stay separate from project media; import one through
+[media guide](../chatcut-media-guide/SKILL.md) only if the user wants to keep or use it.
diff --git a/kimi/skills/visual-analysis/references/scene-breakdown.md b/kimi/skills/visual-analysis/references/scene-breakdown.md
new file mode 100644
index 0000000..13c653a
--- /dev/null
+++ b/kimi/skills/visual-analysis/references/scene-breakdown.md
@@ -0,0 +1,127 @@
+# Scene-by-scene analysis
+
+Use this for understanding a video, studying a reference, or producing a granular
+scene/shot breakdown. Keep a focused visual check focused; a subtitle or transition
+question need not generate a report about the whole video.
+
+## Establish the subject and evidence
+
+Resolve the actual source or composed timeline, requested range and useful outcome.
+Reuse readable media, the existing ChatCut transcript and frames. Analyzing a local
+file does not require placing it on a timeline or uploading it. For a URL, use an
+available acquisition route; a thumbnail, page description or successful import
+alone does not establish its content. Report inaccessible media as a source gap.
+
+Read duration and the relevant clock before assigning times. Original footage
+uses source time; a composed edit uses its timeline clock. Follow the entrypoint's
+host-specific field names and units. Keep placement, trim and rate mappings when
+relating the two. The same numeric value can mean different moments on those
+clocks.
+
+## Inspect at the level the question needs
+
+Begin with a bounded overview spanning the selected range, including its opening
+and closing content. Use the current host's source-frame or timeline-viewer tools;
+batch useful samples and divide long videos into meaningful ranges. Chapters,
+transcripts and detected changes can locate candidates, but they do not prove
+that the relevant visuals were inspected.
+
+Expand around consequential changes: a new shot, a changed software screen,
+prompt or setting, an inserted example, an action's start/result, or a transition.
+Inspect ordered before/during/after moments when motion matters. Read small text
+from a useful full-size frame or crop. A few overview images cannot establish
+every cut or a frame-exact boundary; use the available cut/adjacent-frame route
+when that precision changes the answer. Do not inspect every frame by ritual.
+
+## Read speech from the ChatCut transcript
+
+This analysis assumes the media already has a ChatCut transcript. Read each
+scene's timestamped words from it through the host's transcript field or view,
+and put the words actually spoken in that interval into the scene's `audio`. Use
+the transcript of the media's own speech; a translated or reworded caption track
+does not record what was said. Do not transcribe again or reconstruct speech by
+listening. If the transcript is missing, still processing or unreadable for the
+range, report that gap and leave the speech unresolved; creating a transcript is
+the transcription workflow's job.
+
+Keep the transcript's words, including uncertain or misrecognised terms. When
+on-screen text shows the intended spelling, note the correction instead of
+silently rewriting the quote. A sentence can cross a visual cut; assign its words
+to the scene where they are spoken without inventing a word boundary. Summarise
+only when the user asks for less detail, and label the summary as one.
+
+Listen to the audio only for tone, music, sound effects or synchronization, when
+the question depends on them. Without that listening, say so and do not invent
+acoustic descriptions. A transcript's music tag is evidence of the tag, not a
+listening result.
+
+## Describe meaningful scenes
+
+Let visible content and the user's purpose determine segmentation. Separate
+material shot/state changes; retain a coherent continuing action across transcript
+sentences. A tutorial chapter can contain a presenter, a request, a software
+operation and its output—those are different observations when they matter.
+Group an unresolved rapid montage explicitly rather than guessing its cuts.
+There is no target scene count or fixed sampling density.
+
+For each useful interval capture:
+
+- Start/end on the declared clock, at the supported precision.
+- Visible subjects, setting, layout, text, action and camera/graphic movement.
+- The transcript words spoken in the interval, plus any tone, music or sound
+  effects actually listened for.
+- Framing or shot type, and the scene's role in the video when useful.
+
+Prefer specific observations to a repeated generic presenter description. For a
+software demonstration, inspect the important screen, selected inputs, readable
+settings and resulting state. Keep a creator's request separate from its result.
+An output does not reveal its hidden prompt or internal production algorithm.
+
+Keep illegible text, inferred settings and unverified claims uncertain. Reconcile
+material disagreements between narration, visible controls and results. Current
+documentation can explain a tool but cannot prove the creator used it for this
+clip. Recovering the full production method or creating a new video continues in
+the corresponding workflow; a scene list alone does not complete either task.
+
+## Return useful prose or structured scenes
+
+Use a concise narrative/table for an ordinary question. When the user or a
+downstream workflow needs JSON, the following scene fields retain the useful
+granularity of Higgsfield's exposed Video Analysis results without requiring that
+provider, its queue or its private prompt:
+
+```json
+{
+  "time_domain": "source",
+  "analyzed_range": { "start_seconds": 0, "end_seconds": 4 },
+  "scenes": [
+    {
+      "scene_number": 1,
+      "timestamp_start": "00:00:00.000",
+      "timestamp_end": "00:00:04.000",
+      "visual": "A close view shows a cup on a table; a title enters above it.",
+      "audio": "\"Here is the new cup.\" Music and effects not listened to.",
+      "shot_type": "Close-Up",
+      "label": "Product introduction"
+    }
+  ],
+  "evidence": {
+    "timing_precision": "estimated from sampled frames",
+    "audio_basis": "chatcut_transcript",
+    "coverage": "Sampled frames and transcript words across this selected range."
+  }
+}
+```
+
+This is an illustrative shape, not an observation of a supplied video or a claim
+about a supplied video. The seven scene keys are `scene_number`,
+`timestamp_start`, `timestamp_end`, `visual`, `audio`, `shot_type` and `label`.
+Keep source/asset identity, evidence locators, precision and gaps alongside the
+scenes when needed. Choose useful labels for the actual content; advertising
+categories should not force a tutorial or film into a marketing interpretation.
+
+Check chronological intervals against the actual duration and clock, including
+the tail. Internally continuous times can still drift from the source. Cite the
+actual frames/ranges supporting important claims. State partial coverage rather
+than stretching an uninspected final scene. A completed provider job, more scenes,
+valid JSON or distinct wording does not establish correct analysis.
diff --git a/kimi/skills/visual-analysis/scripts/concat-images.mjs b/kimi/skills/visual-analysis/scripts/concat-images.mjs
new file mode 100644
index 0000000..56be49b
--- /dev/null
+++ b/kimi/skills/visual-analysis/scripts/concat-images.mjs
@@ -0,0 +1,305 @@
+#!/usr/bin/env node
+
+import { existsSync, mkdirSync, mkdtempSync, rmSync } from "node:fs";
+import { basename, dirname, extname, join, resolve } from "node:path";
+import { spawnSync } from "node:child_process";
+
+const parsed = parseArgs(process.argv.slice(2));
+const outputPath = resolve(parsed.output);
+const outputExtension = extname(outputPath).toLowerCase();
+
+if (![".jpg", ".jpeg", ".png", ".webp"].includes(outputExtension)) {
+  fail("--output must end in .jpg, .jpeg, .png, or .webp");
+}
+if (parsed.inputs.length === 0 || parsed.inputs.length > 25) {
+  fail("Pass between 1 and 25 input images");
+}
+if (parsed.captions.length && parsed.captions.length !== parsed.inputs.length) {
+  fail("Pass exactly one --caption per input image, or omit captions entirely");
+}
+
+const inputs = parsed.inputs.map((input) => resolve(input));
+for (const input of inputs) {
+  if (!existsSync(input)) fail(`Input image does not exist: ${input}`);
+}
+
+mkdirSync(dirname(outputPath), { recursive: true });
+const temporaryDirectory = mkdtempSync(
+  join(dirname(outputPath), ".concat-images-"),
+);
+const imageMagick = resolveImageMagick();
+// Some ImageMagick builds cannot discover installed fonts without fontconfig.
+const captions = parsed.captions.length
+  ? parsed.captions
+  : inputs.map(defaultCaption);
+const captionFont = [
+  ...(captions.every((caption) => /^[\x00-\x7F]*$/.test(caption))
+    ? [
+        "/System/Library/Fonts/Menlo.ttc",
+        "/usr/share/fonts/truetype/dejavu/DejaVuSansMono.ttf",
+      ]
+    : []),
+  "/System/Library/Fonts/Supplemental/Arial Unicode.ttf",
+  "/System/Library/Fonts/Supplemental/Arial.ttf",
+  "/usr/share/fonts/opentype/noto/NotoSansCJK-Regular.ttc",
+  "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",
+  join(process.env.WINDIR ?? "C:\\Windows", "Fonts", "arial.ttf"),
+].find((font) => existsSync(font));
+
+try {
+  const prepared = inputs.map((input, index) => {
+    const preparedPath = join(temporaryDirectory, `${index}.png`);
+    const args = imageMagick.prefix.concat([
+      input,
+      ...(parsed.crop ? ["-crop", parsed.crop, "+repage"] : []),
+      "-auto-orient",
+      "+profile",
+      "icc",
+      "-resize",
+      `${parsed.tileWidth}x${parsed.tileWidth}>`,
+      preparedPath,
+    ]);
+    run(imageMagick.command, args, `prepare ${input}`);
+    return preparedPath;
+  });
+
+  const dimensions = run(
+    imageMagick.command,
+    imageMagick.prefix.concat(prepared, ["-format", "%w %h\n", "info:"]),
+    "read image dimensions",
+  )
+    .trim()
+    .split("\n")
+    .map((line) => line.split(" ").map(Number));
+  const aspect =
+    Math.max(...dimensions.map(([width]) => width)) /
+    Math.max(...dimensions.map(([, height]) => height));
+  const captionHeight = parsed.noCaptions ? 0 : 28;
+  let columns = parsed.columns ?? 1;
+  let tileWidth = 0;
+  for (
+    let candidate = 1;
+    candidate <= Math.max(inputs.length, parsed.columns ?? 1);
+    candidate++
+  ) {
+    if (parsed.columns && candidate !== parsed.columns) continue;
+    const rows = Math.ceil(inputs.length / candidate);
+    const width = Math.floor(
+      Math.min(
+        parsed.tileWidth,
+        1568 / candidate - 2 * parsed.gap,
+        (Math.floor(1568 / rows) - 2 * parsed.gap - captionHeight) * aspect,
+      ),
+    );
+    if (width > tileWidth) {
+      tileWidth = width;
+      columns = candidate;
+    }
+  }
+  if (tileWidth < 1) fail("The requested columns/gap leave no room for images");
+  const tileHeight = Math.max(1, Math.floor(tileWidth / aspect));
+  const captioned = prepared.map((input, index) => {
+    const output = join(temporaryDirectory, `captioned-${index}.png`);
+    run(
+      imageMagick.command,
+      imageMagick.prefix.concat([
+        input,
+        "-resize",
+        `${tileWidth}x${tileHeight}`,
+        "-background",
+        "#000000",
+        "-gravity",
+        "center",
+        "-extent",
+        `${tileWidth}x${tileHeight}`,
+        ...(parsed.noCaptions
+          ? []
+          : [
+              "(",
+              "-size",
+              `${tileWidth}x${captionHeight}`,
+              "xc:#eeeeee",
+              ...(captionFont ? ["-font", captionFont] : []),
+              "-fill",
+              "#111111",
+              "-stroke",
+              "none",
+              "-pointsize",
+              "18",
+              "-gravity",
+              "west",
+              "-annotate",
+              "+5+0",
+              captions[index],
+              ")",
+              "-append",
+            ]),
+        output,
+      ]),
+      "add caption below image",
+    );
+    return output;
+  });
+  const montageArgs = imageMagick.montagePrefix.concat([
+    ...(captionFont ? ["-font", captionFont] : []),
+    ...captioned,
+    "-tile",
+    `${columns}x`,
+    "-geometry",
+    `+${parsed.gap}+${parsed.gap}`,
+    "-background",
+    "#eeeeee",
+    outputPath,
+  ]);
+  run(imageMagick.command, montageArgs, "compose contact sheet");
+
+  process.stdout.write(
+    `${JSON.stringify(
+      {
+        columns,
+        crop: parsed.crop ?? null,
+        inputs: inputs.map((input, index) => ({
+          caption: captions[index],
+          path: input,
+        })),
+        ...(parsed.noCaptions
+          ? {}
+          : {
+              note: "The light-gray strips below the images contain tool-added position labels, not video content.",
+            }),
+        width: columns * (tileWidth + 2 * parsed.gap),
+        height:
+          Math.ceil(inputs.length / columns) *
+          (tileHeight + captionHeight + 2 * parsed.gap),
+        outputPath,
+        rows: Math.ceil(inputs.length / columns),
+      },
+      null,
+      2,
+    )}\n`,
+  );
+} finally {
+  rmSync(temporaryDirectory, { force: true, recursive: true });
+}
+
+// Frame filenames carry the exact coordinate returned by the preview tools.
+function defaultCaption(input) {
+  const name = basename(input, extname(input));
+  const timeline =
+    name.match(/^timeline-(\d+)f-\d+ms-\d+$/) ??
+    name.match(/^timeline-frame-(\d+)$/);
+  if (timeline) return `f${timeline[1]}`;
+  const source =
+    name.match(/^source-(\d+)ms-\d+$/) ?? name.match(/^asset-frame-(\d+)ms$/);
+  if (source) return `${(Number(source[1]) / 1000).toFixed(3)}s`;
+  return basename(input);
+}
+
+function parseArgs(args) {
+  const result = {
+    noCaptions: false,
+    captions: [],
+    columns: undefined,
+    crop: undefined,
+    gap: 0,
+    inputs: [],
+    output: undefined,
+    tileWidth: 720,
+  };
+  for (let index = 0; index < args.length; index += 1) {
+    const arg = args[index];
+    if (arg === "--no-captions") result.noCaptions = true;
+    else if (arg === "--caption")
+      result.captions.push(requiredValue(args, ++index, arg));
+    else if (arg === "--columns")
+      result.columns = boundedInteger(
+        requiredValue(args, ++index, arg),
+        arg,
+        1,
+        9,
+      );
+    else if (arg === "--crop") result.crop = requiredValue(args, ++index, arg);
+    else if (arg === "--gap")
+      result.gap = boundedInteger(
+        requiredValue(args, ++index, arg),
+        arg,
+        0,
+        100,
+      );
+    else if (arg === "--output")
+      result.output = requiredValue(args, ++index, arg);
+    else if (arg === "--tile-width")
+      result.tileWidth = boundedInteger(
+        requiredValue(args, ++index, arg),
+        arg,
+        100,
+        2000,
+      );
+    else if (arg === "--help" || arg === "-h") usage(0);
+    else if (arg.startsWith("-")) fail(`Unknown option: ${arg}`);
+    else result.inputs.push(arg);
+  }
+  if (!result.output) fail("--output is required");
+  if (result.crop && !/^\d+x\d+\+\d+\+\d+$/.test(result.crop)) {
+    fail("--crop must use WIDTHxHEIGHT+X+Y pixel geometry");
+  }
+  return result;
+}
+
+function resolveImageMagick() {
+  if (commandExists("magick")) {
+    return { command: "magick", montagePrefix: ["montage"], prefix: [] };
+  }
+  if (commandExists("convert") && commandExists("montage")) {
+    return { command: "convert", montagePrefix: [], prefix: [] };
+  }
+  fail("ImageMagick is unavailable; expected magick or convert + montage");
+}
+
+function commandExists(command) {
+  return (
+    spawnSync("sh", ["-c", `command -v ${command}`], {
+      stdio: "ignore",
+    }).status === 0
+  );
+}
+
+function run(command, args, step) {
+  const executable =
+    step === "compose contact sheet" && command === "convert"
+      ? "montage"
+      : command;
+  const result = spawnSync(executable, args, { encoding: "utf8" });
+  if (result.status !== 0) {
+    fail(
+      `${step} failed: ${(result.stderr || result.stdout || "unknown error").trim()}`,
+    );
+  }
+  return result.stdout;
+}
+
+function requiredValue(args, index, option) {
+  const value = args[index];
+  if (!value) fail(`${option} requires a value`);
+  return value;
+}
+
+function boundedInteger(value, option, minimum, maximum) {
+  const parsed = Number(value);
+  if (!Number.isInteger(parsed) || parsed < minimum || parsed > maximum) {
+    fail(`${option} must be an integer from ${minimum} to ${maximum}`);
+  }
+  return parsed;
+}
+
+function fail(message) {
+  process.stderr.write(`concat-images: ${message}\n`);
+  process.exit(1);
+}
+
+function usage(exitCode) {
+  process.stdout.write(
+    "Usage: concat-images.mjs --output <image> [--columns N] [--crop WxH+X+Y] [--caption text ...] <images...>\n",
+  );
+  process.exit(exitCode);
+}
diff --git a/tests/visual-analysis.test.mjs b/tests/visual-analysis.test.mjs
new file mode 100644
index 0000000..7895c7d
--- /dev/null
+++ b/tests/visual-analysis.test.mjs
@@ -0,0 +1,59 @@
+import assert from "node:assert/strict";
+import { readFile } from "node:fs/promises";
+import test from "node:test";
+
+const hosts = ["codex", "claude", "grok", "kimi"];
+const reference = "references/scene-breakdown.md";
+const read = (host, file) =>
+  readFile(
+    new URL(`../${host}/skills/visual-analysis/${file}`, import.meta.url),
+    "utf8",
+  );
+
+test("all plugin hosts ship one visual-analysis entrypoint and the same portable scene reference", async () => {
+  const expected = await read("codex", reference);
+  for (const host of hosts) {
+    const skill = await read(host, "SKILL.md");
+    assert.match(skill, /^name: visual-analysis$/m);
+    assert.ok(skill.includes(`](${reference})`));
+    assert.equal(await read(host, reference), expected);
+    await read(host, "scripts/concat-images.mjs");
+  }
+});
+
+test("the portable JSON example preserves source coordinates and the seven scene fields", async () => {
+  const text = await read("codex", reference);
+  const examples = [...text.matchAll(/```json\s*([\s\S]*?)```/g)].map((m) =>
+    JSON.parse(m[1]),
+  );
+  assert.ok(examples.length);
+  for (const example of examples) {
+    assert.equal(example.time_domain, "source");
+    for (const scene of example.scenes) {
+      assert.deepEqual(Object.keys(scene).sort(), [
+        "audio",
+        "label",
+        "scene_number",
+        "shot_type",
+        "timestamp_end",
+        "timestamp_start",
+        "visual",
+      ]);
+    }
+    assert.equal(example.evidence.audio_basis, "chatcut_transcript");
+  }
+});
+
+test("Grok and Kimi verification retain a working route to the shipped analysis skill", async () => {
+  for (const [host, entry] of [
+    ["grok", "verification"],
+    ["kimi", "chatcut-verification-guide"],
+  ]) {
+    const text = await readFile(
+      new URL(`../${host}/skills/${entry}/SKILL.md`, import.meta.url),
+      "utf8",
+    );
+    assert.ok(text.includes("](../visual-analysis/SKILL.md)"));
+    await read(host, "SKILL.md");
+  }
+});