Core Principles
First design the information clearly, then produce according to the output mode. Raster mode uses imagegen for page-by-page image generation; when the user wants a web presentation, use the built-in HTML template.
- Content Author: analyzes materials, plans page breakdown, and writes low-fidelity Markdown.
- Visual Generator: generates raster PNGs with a locked aspect ratio based on the page-by-page blueprint.
- User/Reviewer: confirms page breakdown, style, and final text accuracy.
The default final output is full-page images, which offer high visual freedom; the trade-off is that text, charts, and shapes cannot be edited independently. Editable PPTX is not within the implementation scope of this skill; when the user explicitly needs it, deliver only the low-fidelity draft and visual assets, and hand off to a separate native PPTX authoring/export tool.
Initial Parameters
Lock down before starting:
| Parameter | Description | Example |
|---|---|---|
| Style preset | Target visual language | Huawei / user-provided custom preset |
| Audience | Who will view it | Department head / CTO / investor / client |
| Scenario | How it will be used | Large-screen report / post-meeting reading / client proposal |
| Content trade-offs | Whether cuts are allowed | Keep all / keep only core highlights |
| Aspect ratio | Runs through the blueprint and final images | 16:9 (default) / 4:3 / user-specified |
| Output mode | Determines delivery capability | low-fi only / raster PNG / HTML deck |
Without these parameters, first make the best judgment and state it clearly in the page breakdown plan; do not directly generate images in bulk.
Workflow
When the user wants HTML, web pages, or browser page-turning presentations, read the HTML Production Specifications, and after completing content analysis and page breakdown, proceed to template production and browser acceptance; do not execute the imagegen capability pre-check and image generation steps below. When no mode is specified, still default to raster; when there is no image generation capability, deliver the blueprint and explain the optional HTML mode, without switching on your own.
- Content analysis: read the original materials thoroughly, extract core sections, key data, logical relationships, and evidence sources. When experience or contributions are mixed together and the materials are reduced to a list of nouns, first organize according to the Facts and Causality Specifications , then break into pages.
- Page breakdown planning: for each page, clearly write the page type, the single main conclusion, and the elements it carries; when the direction is undetermined, first confirm the page breakdown table. If there is already a clear blueprint or the user has authorized direct production, continue without repeatedly asking for confirmation.
- Low-fidelity blueprint: according to the Low-Fidelity Production Specifications , draw an ASCII layout for each page and write visual guidance.
- Style locking: for Huawei style, you must read Huawei-style professional reporting preset。
- Capability precheck: Confirm the current host provides an image-generation / imagegen tool; if not, stop at low-fidelity draft delivery.
- Generate page by page: Generate page by page according to the locked aspect ratio, named
p{N}-{简短描述}.png, and do not generate the entire set as a single collage. - Visual and text acceptance check: Check hierarchy, readability, text, numbers, units, sources, and cross-page consistency page by page.
- Delivery: Show the complete preview and explain the editability boundaries of the output mode.
When the user rejects the direction or structural problems persist for two consecutive rounds, return to page planning or the low-fidelity blueprint, and do not keep patching a bad layout.
Minimum requirements for the low-fidelity blueprint
Each page must include:
- One main conclusion, not a topic word like "situation introduction".
- Evidence or data sources; numbers without sources must not be packaged as facts.
- ASCII art layout, with region sizes reflecting the image proportions.
- Panel content types: text, data, charts, tables, processes, live photos, concept diagrams, or formulas.
- Page-by-page visual guidance: layout, emphasis, color scheme, charts, and decorative layers.
Do not use "fill the page" as a substitute for information design. Huawei-style high density is structured carrying, not shrinking font sizes, piling up frames, or giving all modules equal weight.
Imagegen handoff protocol
Provide to the image generator:
- Low-fidelity Markdown file path.
- Current page number, title, unique main conclusion, and complete text list.
- Target size: use the aspect ratio locked at the start, with each page as an independent image.
- Style preset and allowed overriding brand colors.
- Output directory and file name.
Generation constraints:
- Do not delete or alter facts, and do not fabricate numbers.
- Do not make decoration the main subject of the content.
- Do not use one image to carry content from multiple pages.
- When encountering long text, first compress the expression or split pages, and do not shrink the font size to the point of unreadability.
- All text in the image must undergo post-generation OCR/manual character-by-character verification; typos, fake characters, or numerical errors must be redone.
Output mode
Raster PNG
Suitable for pursuing visual effects, fast previews, and non-editable delivery. Clearly inform the user: text is not searchable, charts are not editable, and subsequent text changes require regeneration or image editing.
Editable PPTX handoff
This skill does not implement PPTX authoring/export. When the user needs editable delivery, stop at low-fidelity Markdown, page-by-page text lists, and reusable visual asset delivery, and hand off to an independent native PPTX production tool; do not promise that this repository will generate editable files.
HTML deck
Use the built-in Guizang Style C template, H01–H42 layouts, and three red-gray themes to deliver a browser presentation and modifiable HTML source code. For the complete process and verification commands, see HTML production specification. This mode does not require imagegen, nor is it equivalent to native editable PPTX; the template's network dependencies and AGPL license accompany the delivery instructions.
Acceptance gates
- Keep only one main conclusion per page, with 3–6 core information groups.
- The title, main conclusion, and key numbers can be found within 3 seconds.
- Page density obeys viewing distance, and does not set a mechanical metric of "the less blank space, the better".
- Red only emphasizes key conclusions and numbers, and is not spread over large areas.
- Keep the fonts, borders, icons, color palette, and decorations consistent across the same deck.
- All text, numbers, units, and sources match the blueprint item by item.
- Before delivery, check all pages side by side to avoid a single page looking good while the overall style drifts.
Common Mistakes
| Mistake | Consequence | Fix |
|---|---|---|
| Generating images in bulk without confirming pagination | Reworking the entire deck | Review the pagination table and the most complex page first |
| Interpreting "Huawei style" as filling everything with red boxes | Information noise | Convey professionalism through hierarchy, data, and structure |
| ASCII blueprint is too vague | imagegen improvises freely | Clearly draw the area proportions and content types |
| Trusting the text in the image directly | Typos and fake characters enter the deliverable | Check word by word against the text list |
| Treating PNG as editable PPT | Text and images cannot be changed later | Hand off to a separate native PPTX production process |
| Generating all pages at once | Unstable style and text | Generate page by page, accept page by page |
Source
2026-09-20 sync check zts212653/clowder-ai commit 22385b60e01aee9d8691b6a867836ff0e94fa77f: incorporates updates on factual sourcing, narrative clarity, and avoiding duplicate confirmation; the low-fidelity and Huawei preset upstream content is unchanged, retaining this repository's readability and text review enhancements. All citations are self-contained paths and do not depend on upstream internal roles or tools.
HTML mode resources come from seandongx/guizang-ppt-skill commit 8653aaca8d2949cd576752704dc50671a4f35854, original author 歸藏, distributed separately under AGPL-3.0, see Source Record。