yunyunlyk/ai-ppt-prompt

AI-Powered PPT Prompt Dispatch

Four templates are stored here, but only one is used at a time. First determine what the user provides, then call the corresponding template, fill in the placeholders with the user's actual content, and output the final product.

0 First Determine: Use or Not

Use: The user needs a PPT structure, content, image, or background animation.

Do Not Use (just say one sentence, don't apply the template):

• Tool operation requests: "Convert PPT to PDF," "Can't open the file," "How to insert a video."
• The user wants you to directly create the final slide deck (.pptx / web-based slides)—that's the job of ppt-master, guizang-ppt-skill, office-pptx. This skill only handles "what to say" and "what images to use."
• Casual chat or asking if a PPT looks good, without a specific topic.

0.5 Establish Order: Copy → Images → Animation

Every slide has only three elements, and their order cannot be reversed: clear copy delivers information, high-quality images deliver visual impact, and appropriate animation delivers rhythm. First, solidify the copy—if you create images before the copy is finalized, they will inevitably be redone later.

When the user asks for "the full package," follow this order, delivering each step before moving to the next. Do not output half-baked layers of all three at once.

1 Three Steps to Route

Step One: Look at What the User Provides

What the User Has	Which Template to Choose	Where to Read
Only a topic/goal ("Help me create a budget proposal")	Topic-to-Framework	references/framework.md §1
Requirements already clarified, just need a final draft	Minimalist Convergent Instructions (three lines)	references/framework.md §4
Structure and page order already fixed	Provide a structural list directly	references/framework.md §5
No direction yet, wants to see an outline first	One outline to probe	references/framework.md §6
A document, long article, meeting minutes, or retrospective material	Document-to-Logic	references/framework.md §2
Says "the logic is messy" or "I don't know how to arrange chapters"	Four-Model Selection	references/framework.md §3
Needs a cover image / content slide image / can't find an image	Image-to-Prompt	references/image.md §1
Image lacks focus, looks bad, or they have an ugly image they must use	Visual Judgment & Image Salvage	references/visual.md §1–4
Wants a character poster, double exposure, or specific image edit	Image Editing Instructions	references/image.md §2
Wants to add a bit of animation to a static image	Animation Shots (within the tool)	references/video.md §1
Wants to turn a static image into a dynamic background	Image-to-Video Prompt	references/video.md §2

Step Two: If There Are Multiple Requests, Do Them in Order

"Topic + Images" → Do the framework first (Topic-to-Framework), then for pages marked [Suggested Image] in the framework, output image prompts. Output at most 3 sets of images at a time; any more wait until the user asks.

Step Three: Pre-Fill, Do Not Return Placeholders

Replace all [...] in the template with real information from the conversation. If the conversation lacks information that affects the conclusion (audience, goal, page count), ask at most one question; for the rest, infer by default and note the assumption in one line at the end of the output:

Assumption: The audience is department heads, the goal is to secure budget; if incorrect, tell me and I'll adjust.

2 Templates and Judgment Points

2.1 Topic-to-Framework

Four parts: Topic Description / Background Persona / Target Audience / Output Constraints. If any part is missing, the model will write filler text, so all four must be concrete values:

• Core request → a verifiable proposition (e.g., "Prove the project's ROI"), not "Introduce the project."
• Audience → specific people (top executives / cross-functional team / client), as this determines argument focus and detail level.
• Constraints → specific numbers: 20 minutes, 2000 words, 15–20 slides, 5 chapters.

2.2 Document-to-Logic

Three tasks: Declare "strictly based on the document, do not expand"; specify one logical model to connect the entire piece; expand slide by slide with "conclusion-style title + 3–5 key points + suggested image."

2.3 Image-to-Prompt

Three steps: Read the title and extract 1–2 keywords → Convert abstract concepts into the most concrete images in public perception (e.g., "new energy" = solar panels / wind turbines, not carbon emission monitoring) → Output 3 sets of prompts.

Five non-negotiable constraints: Realistic photography style; subject occupies 1/3, white space 2/3 (write this for every set); one subject per frame; no over-extension; structure fixed as "style keyword + subject scene + scene description + composition logic," no parameters.

2.4 Image-to-Video

Only one task: Reverse-engineer a static image into a structured video prompt (lens language + element dynamics + lighting atmosphere). The motion must match the original image's style, and elements should animate in layers with subtle movements. Output the prompt, not the video.

2.5 Visual Judgment (Is the Image Good or Not?)

Not all image problems require regeneration. First determine which type it is:

• Lacks focus → Missing contrast. At least two of the four dimensions must be present: Light/dark (a spotlight darkens the surroundings), Size (subject occupies one-third), Color (cool background, warm subject), Focal point (you know where to look at a glance).
• Have an ugly image, but the product is correct → Don't redraw. First, have the model break the original image into a prompt covering "zones + materials + light + depth of field + style," then feed both the image and prompt back for regeneration.
• Image is stunning but falls flat on the slide → It lacks quiet space for text, only works as a cover or section divider.

When writing prompts, do not use adjectives like "premium feeling"; directly describe contrast relationships (light vs. dark, cool vs. warm, big vs. small, mountain vs. sea).

2.6 Animation

Animation adds liveliness, not storytelling. Before starting, confirm the composition of the image itself is correct—animation cannot fix composition.

When generating within the tool, provide four inputs: Prompt (only motion and light, e.g., "Camera pushes right, ripples flow on the water's surface, clouds drift quickly, light changes"), Motion direction, Lens parameters (push in / pan / depth of field, 0–100%), Keyframes. Missing any one results in random, meaningless movement.

3 Global Output Standards

• Conclusion-style title: Each slide's title is a declarative statement ("Customer acquisition cost dropped 37%, so the targeting strategy needs to continue"), not "Project Background."

• 3–5 key points per slide, concise, data-driven, with modifiers removed.

• Mark diagram pages with [Suggested Image: trend line chart / funnel diagram / relationship diagram].

• By default, output the framework as a table that can be directly pasted into PPT:

  Slide	Conclusion-style Title	Core Points	Suggested Image

• When the user explicitly says "I just need the prompt text to use in another tool," switch to output directly copyable prompt text without commentary.

• Animation only for cover / section / single key slides; do not add animation to body slides or slides with data to be read.

4 Pre-Delivery Checklist

1. Have you clearly stated which template you're using (e.g., "Following Topic-to-Framework")?
2. Did you leave any [...] placeholders in the output? If so, the job is not done.
3. Is every slide title a declarative sentence?
4. Do all image prompts specify "subject 1/3 + white space 2/3" and include at least two contrast dimensions?
5. For images recommended for content slides, is there quiet space left for text?
6. Did you note any assumptions not provided by the user at the end in one line?

5 Reference Files

• references/framework.md — Four-part and minimalist three-part templates, structural list method, outline probing, and breakdown of Pyramid / SCQA / PREP / STAR models.
• references/image.md — Complete Image-to-Prompt instructions and double-exposure character editing instructions.
• references/visual.md — Four contrast dimensions, product image zone prompts, image salvage workflow, differences between cover and content slides, animation restraint.
• references/video.md — In-tool animation lens parameters and structured Image-to-Video prompt instructions.

6 Sources

• Framework and image templates compiled from "2026-AI-Assisted PPT Prompt Guide" (produced by 旁门左道PPT).
• Workflow order (Copy → Images → Animation), contrast principles, image salvage, and animation workbench compiled from Technical Fat's "Complete AI PPT Creation Guide" (Bilibili BV1pFMi6tEig).

This skill only handles routing, pre-filling, and judgment standards; the prompt skeleton remains unchanged. See the repository LICENSE for permissions.