It is not "help you with layout and polishing," but a production line from scenario understanding to finished delivery. Core creed (inherited from ppt-designer):The essence of PPT is information delivery; design serves the content.
Iron rules (not suggestions)
- Outline first, then everything else. Visual image mode (Mode C) must only start after the outline has been confirmed by the user; the image model is only responsible for "drawing the already-defined pages" and never decides what the deck talks about. When the user asks to "skip the outline and draw directly," explain why and provide the outline first.
- Each page must have a page brief first, then an image prompt, and skipping levels is prohibited.
- One deck, one style. Style parameters are uniformly injected into each page from the style brief; page-by-page drift is not allowed.
Acceptance criteria (self-check item by item before delivery)
- The outline exists before any page is produced and has been confirmed by the user (except Mode A, which is itself the endpoint)
- Each page conforms to the required fields of the page schema, with compliant density (3-5 key points / hierarchy ≤3 / font size ratio ≥2:1)
- The entire set uses the same style brief; any two pages have consistent fonts, color palette, and layout code
- Each page prompt in Mode C includes style name, composition, title copy, body copy, and prohibited items
- Deliverables include speaker notes (except when the user explicitly does not want them)
- No emoji is used as a functional icon; the color scheme comes from the selected style palette rather than improvised choices
- Visual output (Mode B/C) has passed after rendering
references/visual-checklist.md(12 quantified original principles + 10 practical principles), and each page declares a registered layoutlayout: LXX
What not to do
- Technical operations on pptx files (splitting/merging/page counting) → host environment pptx skill
- In-depth writing of speeches and article body text →
wechat-article-writer - When content requires multi-source research → first hand off to
deep-research, then return to this process after obtaining conclusions - Single posters, logos, purely decorative images → handled directly by image generation skills
- When the user only wants "a few usable pages, anything is fine," compress it to: quick outline → simplified Mode B output, without going through the full six steps
Execution and authenticity boundaries
- Repository example scripts are only allowed to write to their respective example directories'
html/、png/and explicitly named.pptx; they must not read credentials, access the network, execute shell, or write to the caller's current directory. - The HTML→PNG→PPTX route locks in the visuals, but each page is a full-page raster image in the PPTX,not editable. It must not be promoted as an "editable PPTX."
- When editable delivery is required, use Mode B's native object route, and actually inspect text, shapes, charts, speaker notes, overflow, and rendering; when Office/LibreOffice has not been run, fidelity must be declared unverified.
- External facts and images must have their sources recorded in speaker notes. The mere existence of a file, a preview image, or a script exiting 0 does not mean the visual quality has passed.
Workflow (six steps)
Step 1 Identify the task type
Reporting / Pitching / Training / Proposals / Retrospectives / Consulting reports / Recruitment introductions / Self-media sharing.
For judgment rules see references/scene-routing.md. If identification is uncertain, ask directly, don't guess.
Step 2 Fill in key context (ask at most 5 items, don't repeat what's already known)
① Who it's for ② What the goal is ③ Expected page count ④ Whether content clarity or visual impact is preferred ⑤ Whether there's a specified style/brand color
Step 3 Choose style
User-specified → lock it in. Not specified → recommend 3 according to the references/scene-routing.md "scenario × style" matrix and give a one-sentence rationale.
If the user has brand assets (ChaoGeek / ClawTime / WorkBuddy) → directly use the style-library/custom/ custom style.
Step 4 Produce outline (mode A ends here)
According to templates/outline-template.md: topic / audience / goal / suggested page count / per-page title + one-sentence summary.
Stop and wait for user confirmation before continuing.
Step 5 Page-level structure
First, by page type, choose a registered layout from references/layout-registry.md and declare layout: LXX(don't invent compositions from scratch), then fill in the page schema page by page (references/page-schema.md): page type, core conclusion, visual structure, content modules, image requirements.
Step 6 Deliver by mode
| Mode | Applicable | Output |
|---|---|---|
| A. Outline | The user hasn't thought through the content yet | Outline (Step 4 deliverable) |
| B. Content draft | Hand off to a PPT generator/front-end engine, or use PptxGenJS to directly produce pptx | slide schema JSON + readable Markdown + style parameters; for the PptxGenJS route see references/v2-pipeline/pptx-generation.md |
| C. Visual images | Generate images page by page (image model), or an infographic image set | Per-page page brief → image prompt → image generation; for the workflow and engine adaptation see references/image-branch.md |
Style library (style-library/, currently 17; expansion curated according to references/style-roadmap.md , end state ≤14 official + special cases)
| Category | Style | One-liner |
|---|---|---|
| General business | business/moshiro-consulting.md Ink & White Consulting |
Black, white, and gray + a single accent color, the restrained feel of a consulting report |
| General business | business/heibai-ledger.md Black & White Ledger |
A data-reporting style where tables and numbers are the protagonists (default mode B) |
| General business | business/aicher-system.md Aicher Information System |
Functional color coding + grid icons (Munich 1972), includes template seed |
| Brand creativity | creative/yinghuang-studio.md British Yellow Studio |
Black + warm yellow, a high-end proposal feel, large type with white space |
| Brand creativity | creative/bauhaus-geometric.md Bauhaus Geometry |
Three primary colors + circles, squares, and triangles, includes template seed |
| Brand creativity | creative/constructivist-red.md Constructivist Red |
Red, black, and white + diagonal energy, includes template seed |
| Brand creativity | creative/tanghe-frame.md Candy Box Color Frame |
Cream background + colorful block borders, a trendy magazine feel |
| Education and training | education/qingfeng-classroom.md Breeze Lecture Hall |
Water-blue and refreshing, low-pressure courseware |
| Education and training | education/academic-bluegray.md Academic Blue-Gray |
Thesis defense and academic reports (default mode B) |
| Tech future | tech/polar-night.md Polar Night Tech |
Deep background neon, AI/data product launch |
| Tech Future | tech/platinum-future.md Platinum Gray Future |
Light metallic texture, projector-friendly tech feel |
| Content Distribution | media/hot-card.md Hot Post Cards |
Xiaohongshu-style high-contrast card carousel |
| Content Distribution | media/notion-handdrawn.md Notion Hand-drawn |
v2 signature: hand-drawn lines + sticky-note texture knowledge diagrams |
| Content Distribution | media/neubrutalism-pop.md Neubrutalism Hard Candy |
Black outlines + hard shadows + clashing-color sticker feel, includes template seeds |
| Custom Brand | custom/chaogeek-pixel.md ChaoGeek Pixel Halftone |
Neon green/lobster red/deep background DNA, derived from the chaogeek visual system |
| Custom Brand | custom/clawtime-industrial.md ClawTime Black-Red Industrial |
Carbon black + mechanical red, industrial interlocking feel (first draft pending brand calibration) |
| Custom Brand | custom/workbuddy-modern.md WorkBuddy Green Modern |
Productivity green, efficiency tool official website feel (first draft pending brand calibration) |
Selection rules: first match the scenario by category, then exclude based on the "unsuitable scenarios" in the style files. Each style file contains 10 fields (color palette/font temperament/layout/image strategy/density/style brief, etc.).
Known Pitfalls
| Pitfall | Specific Manifestation | Countermeasure |
|---|---|---|
| Skipping the outline and going straight to pages | The user urges "just draw it," and the whole deck's logic falls apart | Iron Rule 1: after explaining, provide the outline first; the cost of confirming the outline is far lower than rework |
| Page-by-page style drift | Page 8 suddenly changes the color scheme and fonts | Each page's prompt is injected from the same style brief; spot-check any two pages |
| Chinese typography fails with image models | Text overlaps, text presses on images, garbled pseudo-Chinese characters | The prompt must include Chinese typography prohibitions; for long text, switch to the "reserve text space in the image, overlay text later" strategy, see image-branch.md |
| Looks good but is hard to present | Visually striking pages have too low density; one page has nothing to talk about for 3 minutes | The page schema's key_message is required; speaker notes are produced in sync |
| Single-page information overload | 6+ key points, 3 charts crammed onto one page | Hard density rule constraint: split the page |
| Style recommendations are homogeneous | The 3 recommended styles show no discernible difference | The 3 recommendations must come from ≥2 categories |
| Chinese punctuation breaks scripts | When Mode B uses PptxGenJS/JS generation, Chinese quotes are written as straight quotes, causing SyntaxError |
Chinese quotes in JS strings should always use corner brackets 「」 or full-width quotation marks; do not use straight quotes that conflict with delimiters |
| All-thin fonts fake sophistication | Executing "bigger means thinner" as all text becoming thin, making the visuals thin, weak, and anchorless (a real lesson from losing a blind evaluation 3:0) | Extreme contrast = ultra-thin and ultra-heavyCoexist: bold black headings/data heroes serve as anchors, while thin fonts serve only as the contrast layer |
| pptx font fallback | PptxGenJS outputs pptx directly without embedding fonts; when the opening device lacks the font, the layout collapses | For visual-first scenarios, use the HTML→Chrome screenshot→image pipeline (fonts are locked by the rendering machine); only use PptxGenJS if editability is required, and declare font conditions |
| Half-width punctuation in body text | Mixing half-width colons and commas in Chinese sentences gives an obvious cheap feel on large screens | Always use full-width punctuation for Chinese body text; manually tighten the spacing between superscripts and base numbers |
Reference documents (load on demand)
| File | When to read |
|---|---|
references/page-schema.md |
When defining page by page in Step 5 |
references/scene-routing.md |
During Step 1 identification and Step 3 recommendation |
references/image-branch.md |
When starting Mode C; engine adaptation (gpt-image / seedream / skywork) and Chinese typography prohibitions |
references/layout-registry.md |
When choosing layouts in Step 5; 14 registered layouts (from three sets of empirically validated samples) + orthogonal dimension-splitting organization principles |
references/visual-checklist.md |
During post-render self-check; 12 quantified principles from classic sources (Duarte/Tufte/M-B/Butterick/clreq) + 10 practical lessons |
references/style-roadmap.md |
Required reading before expanding the style library; rating table from a survey of 26 sources + negative checklist + final blueprint of ≤14 sets |
references/layout-rules.md |
When unsure about typography details; complete theory of CRAP/hierarchy/color/image (inherited from ppt-designer) |
references/quick-reference.md |
For checking numerical specifications at any time during production (font size/spacing/color values) |
references/case-examples.md |
When a full set of style scheme references is needed; three cases: business/creative/academic |
templates/outline-template.md |
Step 4 |
templates/page-brief-template.md |
Step 5 |
templates/image-prompt-template.md |
When producing prompts page by page in Mode C |
examples/moshiro-consulting-report/ |
Benchmark sample for light color schemes (Ink & White Consulting, 9 pages, HTML→screenshot pipeline, blind evaluation validated); change content without breaking design key points |
examples/polar-night-ai-native/ |
Dark color scheme sample (Polar Night Tech AI Native, 9 pages, including assemble.js output pptx + speaker notes); start dark decks from this |
examples/yinghuang-bootcamp-proposal/ |
Black-gold proposal series sample (British Yellow Studio, 9-page business proposal, reviewed 6.6/10); start proposal-type decks from this |
examples/constructivist-design-constitution/ |
Constructivist Red, 9 pages (reviewed 7.1/10, highest overall, first to break the 7-point line); start decks that need a strong motif/manifesto temperament from this |
v2 legacy (references/v2-pipeline/, implementation details for Modes B/C, use as needed)
pptx-generation.md— Complete PptxGenJS programmatic generation scheme (implementation of direct pptx output in Mode B)design-system.md— v2 design system (tokens/components, default visual parameters for Mode B)style-guide.md— v2 style definitions such as Notion hand-drawn style (to be reformatted into style-library format)aesthetic-design.md— Aesthetic design notesresearch-notes-format.md/subagent-prompt.md— Sub-agent parallel research protocol (when the outline needs information gathering)v2-prompt-template.md— v2 infographic prompt templates (hand-drawn style parameters not covered by new templates can be taken from here)v2-skill-archive.md— Original SKILL.md archive for notion-infographic v2
Regression
evals/routing-evals.json— Trigger/mode routing test cases; after changing description, use the repository rootscripts/run_routing_evals.pyvalidation