Cowart Image Edit

Cowart Image Edit

Use this skill to turn user-provided Cowart 批注 screenshots into revised AI-generated bitmaps placed next to the corresponding original images.

Preconditions

Ensure the cowart_mcp tools required by this workflow are available. Do not call render_cowart_canvas_widget as a routine prerequisite: annotation requests sent from the Cowart widget already have an open canvas, and the existing widget synchronizes inserted results from MCP-backed storage. If the user separately asks to open, reopen, or explicitly refresh the canvas, handle that request once with the canvas-opening workflow.

Cowart state is read and written through Cowart MCP tools, not through a localhost browser service.

The user is responsible for providing the relevant screenshot(s). Do not auto-capture the current canvas and do not scan the whole canvas to infer edit requests; a canvas may contain many images with different annotations.

Workflow

1. Read the user-provided screenshot(s).

   Treat each screenshot as the authoritative edit brief for one output image unless the user says multiple screenshots belong to the same image.

   If the user provides multiple screenshots, process them independently and keep their generated outputs separate. Do not merge annotations across screenshots unless explicitly requested.

2. Extract the edit requirements from each screenshot.

   Read visible 批注 labels, arrows, and nearby edit notes from the screenshot itself. Use the arrow tip or marked region to understand where each note applies.

   Ignore editor chrome such as toolbars, blue selection outlines, resize handles, cursor icons, and unrelated neighboring images.

3. Choose the source image for generation.

   Use the clean underlying image content visible in the provided screenshot as the visual base whenever possible.

   If the screenshot is too cropped, obstructed, or low-resolution to serve as a good image base, ask the user for the original image export or a cleaner screenshot of that specific image.

   Do not read the current Cowart canvas to discover edit intent. Use the screenshot for the requested changes. Cowart state may be read later only to place the generated result without covering existing content.

4. Prepare image-generation input.

   Use the provided screenshot, plus a cleaner source image if the user supplied one.

   The generation prompt should:

   • apply the 批注 text as edit instructions
   • preserve the original image's subject, composition, aspect ratio, and style unless an annotation asks otherwise
   • remove all annotation artifacts from the output, including red arrows, labels, blue selection outlines, handles, and tool UI
   • output only the revised clean image

5. Generate a new bitmap.

   Use the built-in image generation flow available in the current environment. Do not overwrite the source image file. Save the new bitmap with a timestamped filename, for example:

   annotation-edit-20260620-153012.png
   

   Resolve the actual local output image carefully before inserting it into Cowart. Do not assume the built-in image generation flow always writes a fresh file under $CODEX_HOME/generated_images.

   Preferred resolution order:

   • Use the exact local image path returned by the current image generation tool call when one is available.
   • If no new file path is returned, inspect the current Codex session JSONL for the current request and extract the PNG/base64 payload from the latest image_generation_call.result, then write it to the timestamped output filename.
   • Use $CODEX_HOME/generated_images only when you can prove the file was created by the current request, for example by matching its timestamp after this generation step. Never pick an older image merely because it is the newest file in a stale generated_images directory.

   Before inserting the resolved file into Cowart, visually inspect the local bitmap and confirm it is the newly generated revised image for this screenshot, not a stale generated asset.

6. Insert the revised image beside the original with Cowart MCP.

   Prefer the Cowart MCP insert_cowart_image tool. Do not hand-write tldraw asset / shape records or fractional index keys unless the MCP tool is unavailable. The tool copies the bitmap into the page-local assets folder, creates the tldraw image asset and image shape, generates a valid tldraw fractional index, places the image beside the anchor while avoiding overlaps, and saves through the project-backed Cowart canvas files.

   Add a new tldraw image asset and a new image shape. Do not update, remove, hide, reparent, or reorder the original image, the original AI 图片 frame, or any annotation shapes.

   Prefer a clear placement anchor when one is already available:

   • If the user has selected the original image, use that image as the anchor.
   • If the user has selected the original AI 图片 frame, use that frame as the anchor.
   • If the screenshot clearly shows the original image and there is a unique matching generated/original image or AI 图片 frame on the current Cowart page, use that as the anchor without asking the user to select it.
   • If there are multiple screenshots/outputs and the matching anchors are not uniquely identifiable, ask the user to select each corresponding anchor or provide an explicit placement order.
   • If no anchor is clear and the user has not required a specific side-by-side comparison, place the result in a nearby clear area on the current page where it does not cover, move, hide, or delete the original image or annotations.

   Placement rules:

   • If the source image is inside an AI 图片 frame, use the frame's page-level bounds as the anchor and place the new image as a sibling of that frame.
   • Otherwise use the source image's own bounds and parent.
   • When the annotated source appears to have earlier revision images nearby, prefer placing the new revised image to the right of the currently annotated/source image, because older annotation outputs may already live on the left.
   • Place the new image to the right of the anchor with a margin of about 40 canvas units.
   • Match the displayed width and height of the anchor unless the user asks for a different size.
   • If that position would overlap existing content, keep moving right by anchor width + 40 until the new image is clear.
   • If using a clear-area fallback with no anchor, keep the generated image near the annotated source page, match the likely source image size when known, and choose a position that does not overlap existing shapes.

   Recommended shape metadata:

   {
     "cowartGeneratedFromAnnotationEdit": true,
     "cowartAnnotationSourceShapeId": "<selected source image or frame id>",
     "cowartAnnotationScreenshot": "<source screenshot file name when available>"
   }
   

7. Save through Cowart.

   Only do Cowart state access after the bitmap is generated. Use this access only to insert the new image beside the anchor or in a nearby clear area, not to discover edit intent.

   Preferred MCP call shape:

   {
     "imagePath": "/absolute/path/to/annotation-edit-20260620-153012.png",
     "projectDir": "/absolute/path/to/user/codex-project",
     "anchorShapeId": "<selected source image or frame id>",
     "placement": "right",
     "margin": 40,
     "matchAnchor": true,
     "fileName": "annotation-edit-20260620-153012.png",
     "annotationScreenshot": "<source screenshot file name when available>",
     "shapeMeta": {
       "cowartGeneratedFromAnnotationEdit": true
     },
     "altText": "Revised image generated from Cowart annotation screenshot"
   }
   

   The MCP tool must return the new assetId, shapeId, saved asset path, page id, bounds, and generated index. Confirm that the returned index is a valid tldraw fractional index and not a custom descriptive string.

   Fallback only when insert_cowart_image is unavailable: update the required store snapshot and save it with save_cowart_canvas_state.

   In fallback mode, use page-local image asset URLs:

   /page-assets/<page-dir>/<filename>
   

   Cowart's MCP storage layer will preserve per-page snapshots under:

   canvas/pages/<page-id-without-page-prefix>/cowart-canvas.json
   

8. Verify visually.

   Let the Cowart widget refresh from MCP-backed storage, then confirm:

   • the original image is still in the same place
   • the original 批注 arrows and labels are still visible
   • the new revised image appears beside the original
   • the new image does not include annotation arrows, labels, selections, or UI chrome

Guardrails

• Never replace the original image unless the user explicitly asks for replacement.
• Never delete or move annotation shapes; they are the visible edit brief.
• Never put the revised image inside the original AI 图片 frame, because that can cover the old image and make the before/after comparison harder.
• Never auto-capture or scan the current canvas for edit intent; use the screenshot(s) supplied by the user.
• If the annotations contradict each other, generate the most literal combined interpretation and mention the ambiguity.
• If a supplied screenshot shows selected-state outlines or toolbar UI, treat them as context only, not as content to generate.