What does pic-ppt do?
Primarily uses image generation to faithfully reconstruct user-provided PPT mockups. By default, one image corresponds to one slide, delivering a .pptx file where: the background, each wireframe, small image, and icon are separate image objects, and text is in native editable text boxes. Images can be independently moved, replaced, and scaled; it does not claim their internal lines are editable.
Core Conventions
- Follow the user-specified order: background first, then each wireframe, then each small image/icon, and finally text layout and composition. You may identify text and positions at the start, but complete the background and foreground materials first.
- Maintain the original image's composition, layering, color distribution, element positions, text content, and page aspect ratio. High-definition regeneration improves clarity and edge quality; do not redesign or add decorations on your own.
- Common rectangular boxes, rounded boxes, tech-style borders, wireframe illustrations, etc., are also regenerated via image generation. Do not switch to native PPT shapes, SVGs, or programmatic drawing just because the shape is simple.
- Do not use the entire mockup as the final background with text overlaid on top; the background must genuinely remove the foreground and text that will be independently reconstructed to avoid double images.
- Do not regenerate the entire page as a single image and present it as the final product, nor merge all foreground elements into one transparent image. Each independent element must correspond to its own image object.
- When clear input is provided, start immediately without adding confirmation steps for outlines, styles, or draft pages. Only request clarification for issues affecting content accuracy that cannot be resolved from the input; continue working on other independent parts.
Tools and Files
- Use the current environment's
imagegenskill and built-inimage_gentool for background filling, cropping, and high-definition regeneration. When the user explicitly specifies other available image generation skills, follow their choice. Even if general image generation guidelines recommend vectors for simple wireframes, this task follows the user's requirement to use image generation. - Consult the tool definitions for the current session. For local reference images, first inspect them with an image viewing tool, then pass the actual reference image according to the tool's supported methods. Simply writing a local file path in the prompt does not equate to passing the image.
- Generate each distinct material independently and inspect it. For transparent wireframes and cropped icons, explicitly set
transparent_background: true; for full background images, typically usefalse. Generating multiple materials does not authorize switching to a paid API or CLI. - If image generation fails, retry by correcting the actual issue. If the current method is unavailable, explain the specific obstacle; do not secretly substitute with PPT shapes or full-page screenshots. Fallback to other APIs/CLIs must follow the authorization rules of the corresponding tools.
- Before starting actual PPT layout, read the current
presentationsskill and its required implementation notes. When runtime dependencies are needed, callload_workspace_dependencies. This skill dictates the reconstruction method and does not hardcode layout APIs that change with version. - Keep the original image unchanged. Save intermediate materials and records in the working directory, and place the final PPT and preview in the delivery directory for the current task. Save the selected image generation results to the working directory before embedding them into the PPT; do not keep only temporary links or external references.
0. Observation and Decomposition
Examine the original image to determine the effective slide area, excluding software toolbars and external white space from screenshots. If no other size requirements exist, maintain the aspect ratio of this area; do not default to a forced 16:9.
Create an internal element list, recording: element number, type, original image coordinate box, occlusion order, output material path; for text, also record the original text, line breaks, font style, estimated font size, alignment, color, and local emphasis. Coordinates should be based on the effective slide area uniformly.
Distinguish four categories:
| Category | Handling Method |
|---|---|
| Full-page base color, background texture, background light field | Merge into a single text-free background image |
| Independent wireframes, borders, decorative lines, decorative arrows | Each independent element as a transparent image |
| Small illustrations, photos, icons | Each independent element as a high-definition image; parts requiring cropping should be transparent |
| Titles, body text, text within boxes, figure captions, identifiable labels | Native editable text boxes |
If an image contains labels, count identifiable labels as the text layer, and remove these texts when regenerating the graphics. Do not bake text inside boxes into the wireframe material.
1. First, Extract the Background Image
Use the original image as an editing reference to remove all text, wireframes, small images, and icons that need to be independently reconstructed, filling in the background they occlude. Preserve the original background's hue, gradient direction, texture scale, light and dark areas, and light source position.
Since occluded areas cannot be truly recovered from a single image, fill them in by continuing the surrounding background context; do not claim to have extracted original hidden layers. Do not cover complex textures with large blocks of solid color, nor leave behind text outlines, line residuals, or repeated halos from foreground elements.
Proceed to wireframes only after the background image passes inspection. See references/asset-prompts.md for image generation prompts.
2. Individually High-Definition Regenerate Wireframes
Locate and reference each wireframe in the original image individually, and output a complete independent material. Clearly define the original outline, gaps, corners, relationship of double lines, proportions, edge decorations, and main color. Regenerate only that object, removing text and surrounding elements.
Key controls:
- Uniform Lighting: The brightness and line width of a continuous same segment should be stable, with balanced left/right or top/bottom visual appearance, no unexpected bright spots, local overexposure, or dark breaks. Gradients or key highlights explicitly present in the original image may be retained.
- Even Distribution: Repeated scales, nodes, corner details, and margins should follow a regular pattern, maintaining the original design's symmetrical or asymmetrical relationships; do not forcibly change an asymmetrical composition to symmetrical.
- Integration into Page: Hue, halo intensity, and sharpness should coordinate with the background; avoid white edges, black edges, or a sticker-like appearance. Do not enlarge light effects or add fluorescent colors for the sake of "high definition."
- True Transparency: Save as PNG with a true Alpha channel; areas outside the frame should be transparent. If the original has a hollow interior, it should also be transparent; if the original has semi-transparent panel fills, retain them with reasonable opacity. Preserve soft, semi-transparent halos; do not cut halos with hard edges.
- Leave Space for Halos: The generated canvas should cover the complete outer contour and any outward glow; do not abruptly cut off at the image edge. Transparent white space should not be used as a basis for visible frame positioning.
Process each different wireframe separately. Identical, repeatedly appearing wireframes can reuse the same material that has passed inspection, but they must still be separate image objects in the PPT. Wireframes with different proportions or decorations should be generated separately; do not stretch the same image to fit a different shape.
3. Individually High-Definition Regenerate Small Images and Icons
Using the small image/icon in the original image as a reference, preserve the subject's identity, outline, quantity, direction, perspective, colors, and relative detail positions. Regenerate to a resolution suitable for the final placement size.
- Output isolated icons/objects as transparent PNGs. Keep the original rectangular frame and background for full photos, unless the original image was already a cutout.
- Do not alter meaningful structures such as arrow directions, number of parts, connection relationships, etc. For images containing recognizable text, first extract the text; reconstruct the graphic and text separately.
- For real data charts, experiment images, or brand logos, regeneration must not change data, evidence details, or logo structure. If faithful reproduction cannot be guaranteed, keep the original image and clearly document this exception; do not pass off generated details as original information.
- Set target resolution based on actual placement size; prioritize that materials do not need to be enlarged at their final preview size, and reserve about 2x pixels when conditions permit. Simply enlarging the original image does not count as high-definition regeneration; do not call a resolution not provided by the tool as 4K.
Check the effect on the target background image one by one; after confirmation, place them at the original position on the list.
4. Restore Editable Text
Create native text boxes block by block according to the original text, preserving the original image's text content, hierarchy, line order, and color distribution. Do not polish, delete characters, or change numbers without authorization. OCR is only a draft; you must proofread against the original image.
- Titles, body text, figure captions, numbers, units, and labels inside wireframes should all be directly clickable and editable.
- Use native rich text fragments for text within the same line that has different colors or weights. You may split into multiple text boxes when independent positioning is needed; do not convert text to PNG, SVG, or outlines.
- Prioritize using the original font; if it cannot be determined or is missing, choose an installed approximate font, focusing on matching weight, width, and line height. Proofread superscripts, subscripts, Greek letters, chemical formulas, and mathematical symbols.
- Match font size, line spacing, paragraph spacing, text box margins, and alignment to the original image position. Do not rely on automatic text shrinking to fit text into wrong-sized boxes, and do not change the original layout to achieve a larger font size.
- Images containing text must not keep the original text and then overlay editable text. After combining the text-free material with text boxes, there should be no double text, residual artifacts, or misalignment.
- If truly illegible text is encountered, complete the other parts first and only ask about the specific segment; do not guess seemingly certain content. If the user requests an early delivery, list unresolved text and do not claim all text has been accurately reconstructed.
5. Assembly, Inspection, and Delivery
Before assembly and during acceptance, read references/layout-and-checks.md to handle coordinate mapping, transparent margins, and native text validation.
Place images and text according to the original image's occlusion order, typically: background, wireframe/illustration, text. Give each object a clear name, such as BG_01, FRAME_01, ICON_01, TEXT_01; avoid locking the entire page into a large non-clickable group.
After exporting the PPTX, render it to an image, and compare the full page and local areas at the same scale as the original image. Correct background residuals, frame position deviations, halos, icon scales, typos, line spacing, and overflow. Check the actual transparent channel of PNGs and the native text in the PPT; merely "appearing transparent" or "having a few text boxes" does not count as completion.
By default, deliver the final .pptx file and one rendered preview. Text should be editable; image layers should be independently movable, replaceable, and scalable. Only deliver additional material packages or internal inspection records when requested by the user. Include a brief note about real limitations affecting usage; do not claim pixel-perfect consistency, full vector editability, or that it has been verified in software that was not used.