视觉画布设计
canvas-design
基于设计原则创作静态视觉作品,用于海报与品牌图形。
presentation-design【文档与办公】Generate a premium 6-slide presentation design board as one single composite image. Use this skill whenever the user asks for PPT design concepts, keynote-style slide mockups, a strategy deck visual, an editorial / sports / news / business analysis deck, a TV-friendly slide layout, or any "6-page slide board" / "design these slides" / "PPT 设计稿" / "幻灯…
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
Create a premium 6-slide presentation design board as one single composite image — a board view that lets the user see the whole deck at a glance, TV-ready and cinematic.
Primary use cases: PPT design concepts, keynote-style slide mockups, strategy deck visuals, event / news / sports / business analysis decks, editorial data stories, executive presentations.
The final board must be generated by an image-generation model (use the built-in image_gen tool by default, or an approved project image-generation workflow). Do not create the board with code, SVG, HTML/CSS, canvas, diagrams, placeholder drawings, or deterministic shape scripts. Those methods may be used only for non-visual file operations or optional post-processing after the model-generated image exists.
Save the model-generated raster output to presentations/{name}.png. If the model output cannot be saved directly, locate the generated raster asset and move/copy it into the required path; do not recreate it programmatically.
One single image containing all 6 slides:
┌──────────────┬──────────────┐
│ Slide 01 │ Slide 02 │
├──────────────┼──────────────┤
│ Slide 03 │ Slide 04 │
├──────────────┼──────────────┤
│ Slide 05 │ Slide 06 │
└──────────────┴──────────────┘
presentations/
{name}.plan.md # the approved deck plan (spec)
{name}.png # the final composite image
Use lowercase, hyphenated, ASCII-safe names. For revisions, suffix with -v2, -final, -alt-a. Do not overwrite existing files unless the user asks for replacement.
Before designing, identify:
If the user does not specify language, use the user's local language.
Extract inputs from the user's brief (see Inputs to Extract).
Ask the user for style mode: light or dark — always ask via AskUserQuestion (do not silently infer). Offer dark cinematic as the default suggestion when the topic is ambiguous.
Load references/style-modes.md and pick the most appropriate sub-mode (e.g. dark sports broadcast for an NBA recap, light minimal keynote for a B2B SaaS pitch). Briefly state the chosen palette and typography direction to the user.
Load references/slide-layouts.md and select 6 layouts that together form a coherent narrative arc. Mix layouts — do not repeat the same one across all 6. Always use hero-title for slide 01 and closer for slide 06 unless the topic demands otherwise. State the chosen layout sequence.
Write the deck plan (mandatory) to presentations/{name}.plan.md. For each of the 6 slides include:
slide_no — 01–06.layout — one of the names from the layout library.title — the headline that will appear on the slide.subtitle — optional one-line framing.content — 3–6 short bullets / key points in slide-native language (no essay paragraphs).visual — description of the hero image, chart, map, portrait, or symbolic visual that occupies the tile, plus how it integrates with the text.accent_use — where the accent color appears on this slide (title, divider, chart highlight, etc.).Show the plan to the user and wait for confirmation or revisions. If the user revises, update presentations/{name}.plan.md and re-confirm before proceeding.
After the user approves the plan, generate the first composite image with image_gen. Build the prompt by composing:
references/style-modes.md.Save the result to presentations/{name}.png.
Self-reflect on the rendered image (mandatory). Inspect the saved PNG and run both passes — Pass 1: Structure & Story and Pass 2: Visual Polish — against the actual rendered result, not just the plan. Check specifically for:
Decide: ship or regenerate. If self-reflection finds material issues, regenerate — either re-render the full board with a tightened prompt, or do a targeted fix (re-prompt focused on the broken slide(s), instructing the model to preserve the rest of the board). Cap automatic regeneration at 2 retries before surfacing remaining issues to the user. Save the final accepted image to presentations/{name}.png (overwrite prior drafts); optionally keep presentations/{name}-v1.png etc. for debugging.
Respond to the user with: the saved image path, the deck plan path, the style mode chosen, the layout sequence, and a short note on anything the self-reflection could not fully resolve.
Each slide should be rich but not cluttered. Preferred density per slide:
Avoid: essay-like writing, too many bullet points, overexplaining, raw article summaries, tiny unreadable text, excessive icons, inconsistent information density.
Use slide-native language: short, sharp, visual, structured, easy to scan.
Always run both passes before declaring the board final.
Check:
Improve: slide sequence, narrative clarity, content hierarchy, factual compression, section naming.
Check:
Improve: contrast, spacing, typography, color balance, hero image integration, chart readability, cinematic depth, consistency.
The final design should feel like:
The final design must not feel like:
Load these on-demand during the workflow — they are not read by default to keep this file lean.
references/slide-layouts.md — named layout library. Load at workflow step 4 when choosing the 6-slide composition.references/style-modes.md — light / dark sub-modes with concrete palettes and typography direction. Load at workflow step 3 after the user picks light or dark.