スキル一覧に戻る
EvalsOne

doc-to-slides-images

by EvalsOne

Shared skills used in ConsoleX AI

0🍴 0📅 2026年1月16日
GitHubで見るManusで実行

SKILL.md


name: doc-to-slides-images description: Generate NotebookLM-like slide images (PNG) from user-provided documents, web articles/URLs, or pasted text. Use when the user asks to create slides, a slide deck as images, or NotebookLM-style slides. Produces 8–12 slides by default (max 20), 16:9 4K. Style is routed by document type with a clean default and user overrides. Render via an image-generation tool call (default nano-banana-pro; user-specified model/tool wins; no local API calls).

Doc → Slides Images (PNG)

Output contract

  • 8–12 slides by default; cap at 20 unless user explicitly requests more.
  • 16:9, 4K (3840×2160).
  • Output: slide_plan.json + slide images (PNG) in order.
  • Slide 1 must be a cover; the last slide must be an ending slide.
  • No page numbers, no footers/headers, no logos.
  • If you include key numeric facts, you may add a short inline source in the same bullet (e.g., (..., Source: ...)).

Workflow

  1. Read references (MANDATORY before any rendering)
  • Read these files carefully before deciding layouts or writing any image_prompt:
    • references/slide_json_schema.json
    • references/slide_blueprints.md
    • references/prompt_templates.md
    • references/style_router.md
  • Use the default image tool/model nano-banana-pro unless the user explicitly specifies another model/tool.
  1. Ingest & normalize input (document / URL / pasted)
  • Extract readable text, remove boilerplate, deduplicate.
  • If long: chunk → summarize per chunk → merge to unified outline.
  • Capture metadata: title, doc type guess, audience, tone.
  • Default language follows user request; otherwise follow source content language.
  • Assume the deck is meant to be readable and shareable without a live presenter (self-explanatory slides).
  1. Decide style + write STYLE INSTRUCTIONS (global)
  • Use default “NotebookLM clean deck” unless router strongly suggests otherwise.
  • If user specifies style, user wins.
  • Read references/style_router.md when choosing style.
  • Create a STYLE INSTRUCTIONS code block that captures the global design decisions (background + hex, palette, typography, icon/illustration style, and layout principles). Treat it as the single source of truth for consistency.
  1. Build slide plan JSON
  • Follow references/slide_json_schema.json.
  • Choose 8–12 slides unless content is very long; never exceed 20 by default.
  • For each slide, choose a layout_blueprint from references/slide_blueprints.md.
  • Build per-slide image_prompt using references/prompt_templates.md.
  • Slide titles must read like human narrative sentences:
    • Avoid Title: Subtitle colon patterns.
    • Avoid clichés and “AI slop” phrasing; use direct, confident, active language.
  • Include enough context on-slide so each slide can be understood standalone (within the legibility constraints).
  • Cover + ending slides must look different:
    • Cover slide: use TITLE blueprint (big title, minimal text).
    • Ending slide: use SUMMARY or CTA blueprint (distinct layout from content slides).
  • Ending slide must NOT be generic (“Thank you”, “Any questions?”). Use a designed closing statement, meaningful quote (if available), or a visual summary that anchors the narrative.
  • In each slide’s notes, include (as text) these 4 sections to help QA and iteration:
    • NARRATIVE GOAL (why this slide exists)
    • KEY CONTENT (exact on-slide text; include short sources for key numeric facts)
    • VISUAL (what to draw)
    • LAYOUT (hierarchy and composition; align to the chosen blueprint)
  1. Render a demo slide first (MANDATORY gate)
  • Explicitly output the full slide_plan.json and the STYLE INSTRUCTIONS block for user review.
  • Pick one representative content slide from slide_plan.json (NOT the cover or ending slide).
  • Render only this one slide first (using the selected image tool/model) and show it to the user for confirmation.
  • If the user requests any style/layout changes, regenerate the demo slide and ask for confirmation again.
  • Treat the latest user-confirmed demo slide as the single source of truth: FINAL_DEMO_SLIDE (older demo renders are obsolete).
  • When a new demo slide version is confirmed, overwrite/replace FINAL_DEMO_SLIDE with that latest PNG output (the exact image you will pass as the reference input).
  • Do not render the remaining slides until the user explicitly confirms the plan + style + FINAL_DEMO_SLIDE.
  1. After user confirms: batch render the remaining slides (MANDATORY consistency)
  • For every remaining slide:
    • Call the image generation tool/model (default: nano-banana-pro; if the user specified another model/tool, use that).
    • Use the slide’s image_prompt.
    • Always use image-to-image edit mode (e.g., mode=edit) and pass FINAL_DEMO_SLIDE as the reference image input to the tool call (if the demo slide was regenerated, use the regenerated one).
    • Explicitly require the generated slide to keep the same base style and background as the demo slide.
    • Require: 16:9, 4K PNG output.
  • If you cannot attach FINAL_DEMO_SLIDE as the reference image input (missing/expired/not available), stop and re-render/re-confirm the demo slide first.
  • You may render the remaining slides in parallel.
  • Do NOT use local scripts to call Gemini/Vertex/HTTP APIs.
  1. QA & regenerate if needed
  • If text illegible, regenerate with stricter prompt:
    • larger font, fewer words, higher contrast, more whitespace, vector look.
  • Keep consistent palette, typography feel, iconography across the deck.
  • If any slide drifts in background/style, regenerate it with the demo slide reference image and stricter “match reference” constraints.

Prompting rules (must follow)

  • Prefer vector/infographic presentation style unless user asks photorealistic.
  • Only include requested on-slide text: title + bullets; no extra labels.
  • Keep bullets short (≤12 words), 3–6 bullets.
  • Enforce safe margins and grid alignment.
  • No footer/header/page number.
  • If sources are needed for key facts, include them inline inside the relevant bullet text (not as a footer or citation marker).
  • If an inline source would make the slide less readable, keep the source in notes only (and keep on-slide text clean).
  • Avoid colon-title patterns and generic endings (“Thank you”, “Any questions?”).
  • After demo approval (including after any re-generation), always include “match the reference image’s base style + background” constraints and provide FINAL_DEMO_SLIDE as the reference image input for every subsequent slide generation call.

Bundled resources

  • Layouts: references/slide_blueprints.md
  • Style routing: references/style_router.md
  • Prompt templates: references/prompt_templates.md
  • JSON schema: references/slide_json_schema.json

スコア

総合スコア

45/100

リポジトリの品質指標に基づく評価

SKILL.md

SKILL.mdファイルが含まれている

+20
LICENSE

ライセンスが設定されている

0/10
説明文

100文字以上の説明がある

0/10
人気

GitHub Stars 100以上

0/15
最近の活動

3ヶ月以内に更新がある

0/10
フォーク

10回以上フォークされている

0/5
Issue管理

オープンIssueが50未満

+5
言語

プログラミング言語が設定されている

0/5
タグ

1つ以上のタグが設定されている

0/5

レビュー

💬

レビュー機能は近日公開予定です