Moyu PPT

What does Moyu PPT do?

Transform your content into readable, narratable, and design-consistent slides. The core design artifact is a complete spatial layout map for each page; final delivery is a PPTX with one image per page where text is embedded in the image. Preserves the user's chosen IP, brand, and style direction.

{SKILL_DIR} is the directory containing this file. Each project uses a dedicated working directory decks/{deck-name}/, created under the user's current project directory (or the location specified by the user); do not create it under {SKILL_DIR}, as skill upgrades may overwrite content there. Execute commands within this working directory and quote paths. Use python3 "{SKILL_DIR}/scripts/deck_state.py" to write decisions and status to deck-state.json: it adds timestamps, automatically invalidates previous confirmations when confirmed content changes, and the status field shows what is needed to proceed to the next phase. After resuming a task or upgrading the skill, first run status, then re-read this file and references for the current phase.

0. First-Time Use

Each time you start, run python3 "{SKILL_DIR}/scripts/configure.py" status. If configured is true, proceed directly. If false, first perform a one-time setup guide, then begin the user's task (see setup for details):

1. Where are you using this?: Ask the user if they are using it on Doubao Work, WorkBuddy, Codex, or another platform, and record it with set-platform. Doubao Work and Codex use the platform's built-in image model and require no additional setup. Other platforms use their built-in image generation tool if available; otherwise, configure an external image generation model as per the setup. Any platform can optionally switch to a user-configured image generation model.
2. Inspiration Search: Ask the user to visit https://app.scrapecreators.com/ to obtain an API key and send it to you. Save it with set-key scrapecreators. If they prefer not to configure it now, use skip-search and do not ask again thereafter.

Do not repeat the key in responses after saving it. When configuration is complete, mention the attribution once as per the setup, and do not bring up configuration again.

1. Scope Determination

requested_output	Stop when you reach this point
inspiration	Differentiated recommendations, sources, and reasoning
style	Style configuration and 5-role typography specification; generate font samples only if image generation is authorized
plans	Outline and full spatial layout map, no image generation
deck	Visual plan, images, quality check, and PPTX
revise	Load an existing deck, modify only specified pages and their affected dependencies
template	Save, list, view, or load local style templates; do not automatically generate PPT

• Determine the scope based on the user's request and context: the platform the material comes from, negative statements, or mixed Chinese/English are not reasons to exclude. "Make a knowledge card" and "Modify page 3 of the previous deck" are distinguished by context.
• Default is standard, where design decisions are reviewed and confirmed by the user. Only use delegate --quote --scopes to record the user's exact words and scope when the user explicitly says something like "You decide the number of pages and the style, just go ahead" or "Don't ask me about anything throughout the process." Phrases like "Help me make a PPT," choosing a style, or replying "default" only confirm the item being presented at that moment.
• Do not re-ask for information already provided. Extend existing authorizations to subsequent steps they cover. For partial requests, stop once the completion condition is met; do not automatically escalate to paid generation.

2. Five User Checkpoints

At each checkpoint, the user must see the actual content and respond before you can proceed. See workflow-gates for recording rules and restrictions. autonomous mode only waives confirmation within the delegated scope; the display, sample pages, typography artifacts, and self-checks still proceed as normal. While waiting for a user reply, you can read materials, organize page types, and inspect existing assets. Tasks dependent on the user's reply, such as searches and image generation, must wait.

Checkpoint	What to show the user	What to record after reply
① Requirements	Content skeleton (knowledge-based or speech-based, main thread, key points of each section, page count tiers and what each tier omits) + known items + a set of suggestions for missing items (page count, aspect ratio, audience/use case, visual-to-text ratio, title layout), while also asking about the overall feel and any styles they like or dislike	artifact content_map → brief (use --proposed for suggestions; --visual-ratio for ratio, --title-layout for title layout) → decide intake --quote
② Copy	A copy draft page-copy.md: title and every line of text to be displayed on each page of the entire deck, all at once for a single confirmation. Point out new additions and citations separately. For knowledge-based decks, citations, quotes, key takeaways, and person attributions are not displayed by default; only include them if the user explicitly requests	artifact page_copy → decide copy --quote
③ Direction	A numbered, high-resolution image gallery grouped by category: one group for layouts, one for style elements (check the local style library and template library first). After reviewing the original images, provide 2-3 combination suggestions, each compiled into a single composite image (layout on the left, style elements on the right). Do not delete images. The user can select from your suggestions, mix them, or choose their own. Then combine the chosen layout and style into a direction, and present it along with Chinese title/body/number font pairings and their limitations for confirmation	typography, artifact → decide visual --quote
④ Outline + Sample Pages	Full outline, coverage table (original structure → page number: displayed, merged, not displayed with reasons), and 1–3 real sample pages (actual titles and body text, including numbers for data pages). This step is about the visual; text should follow the confirmed copy draft	pilot, artifact pilot_image → decide pilot --quote
⑤ Save	After delivery and quality check, propose a template name and summary for saving, ask only once	template-offer

Suggested values are just suggestions. If you suggest a range like 8–12 pages, the user must still choose a specific number; do not automatically pick the upper limit. Derive page count from the content skeleton: offer 2-3 tiers and clearly state which sections each tier omits. If the user's given page count is insufficient, explain which sections need to be omitted and let the user choose to add pages or confirm the omission; do not omit sections silently. The suggested visual-to-text ratio should follow the content type (speech-based: 80/20 visual-dominant; knowledge-based: around 70/30) and explain that it is a design budget, not a surface area measurement, and that text-heavy pages will be adjusted or split accordingly. The user can choose an overall ratio (e.g., 80/20, 60/40, 50/50) or choose to decide per page (per-page). Title layout is a binary choice within the suggestion set: Uniform Layout (title in the same position with the same size on every content page, suitable for reports, courseware, and text-heavy knowledge decks) or Free Composition (title position and style follow each page's composition, only locking color, glyph, and illustration style, suitable for sharing, speeches, and visually rich styles like illustrations or games). Explain that with Free Composition, titles can only achieve consistent glyphs and similar sizes; for strictly uniform sizes, choose Uniform Layout. For inspiration and style requests, do not write a copy draft; end at ③. For plans, deliver the plan after ③ and stop.

3. Workflow

1. Initialization & Requirements (①): Run init --deck-name … --requested-output …. After reading the materials, first write the content skeleton content-map.md: determine if it is knowledge-based or speech-based, articulate the main thread, and for each section, describe what the speaker will say, how they will explain it clearly, and the source material basis (see content-structure-method). Register it after content_map.py check passes. Then present the skeleton, known items, and suggestions for missing items all at once, and record the available image generation and search capabilities for this session (see intake-and-state). For how to ask about style, see style-discovery. If the user says "I have no idea, you explore first," proceed with open exploration.
2. Copy Draft (②): After the requirements are confirmed, before searching for inspiration and doing design work, write all text to be displayed on each page into a page-copy.md file: each page starts with a line 内容来源 (Content Source), then one line per display element: 上屏：文字｜出处 (Display: text | source), marking the source as original, compressed, new (with reason), or citation (with reason). The displayed text derives from "what the speaker will say" and "how they will explain it clearly" for that section. Do not add summarizing sentences or slogans not present in the original. For knowledge-based decks, citations, quotes, key takeaways, and person attributions are not displayed by default. If Uniform Layout is chosen, finalize the character count for titles of similar pages now; Free Composition has no character count limit. Register after page_copy.py check passes, present the entire copy draft to the user at once (not page by page), point out new additions and citations separately, and record after user confirmation. After this point, all visible text in the visual plan must come from this copy draft. To change text, first update the copy draft and get re-confirmation. See content-structure-method, section 10.
3. Search & Direction (③): First check the local style library and template library. Show suitable matches for the user to choose from. Searching is the AI's responsibility: generate keywords organized by category, grouped separately (6–8 groups total, 3–4 keywords each). If the user names a specific style, only search for things related to that style: complete presentations in that style, layouts (cover, table of contents, infographic pages, etc.) using that style, and its own style elements (interfaces, icons, scenes, textures, logos). Each keyword must include the style name; record the style name in brief.style_terms. Do not search for generic layouts (business templates like flowcharts, comparison pages) or similar styles unless the user requests. If no style is named, search for layouts by page type and style elements by aesthetic breakdown. Write keywords into inspiration-plan.json (with a brief, first run --validate-plan). Use ref_gallery.py build to build a numbered gallery while searching: for each group, first search the top two keywords, display 6 original images, taken turns from the original Pinterest search order. Do not delete images based on style or quality; re-include images already seen. If the user wants to see more from a group, first show what has been found, then search the next keyword if needed. Comply with requests for more searches (--keywords). Do not enlarge composite images; large/ contains the original dimensions for each image. When the user wants a closer look at a specific image, provide the original. After the gallery is built, review the composites, then open the original images you want to recommend one by one (from large/). Provide 2-3 combination suggestions (layout number paired with style element number, what each borrows, its shortcomings), compile each into a composite image using ref_gallery.py combo for the user to see. Do not delete images. When mentioning styles or references, use the original images, not the composites. The user can select from your suggestions, mix them, or choose their own. If they want to save something to the style library, use ref_gallery.py pick. Then, combine the chosen layout and visual style into a direction and write it into visual-proposal.md for user confirmation. See inspiration-search-method.
4. Typography System: Based on the actual copy, word density, and chosen style, define the five roles: display/title/body/label/data. For titles requiring expressiveness, generate T samples using actual short copy; verify body text in the sample pages. Separate the glyph T from the size template SZ for the same aspect ratio: T conveys stroke features, SZ conveys physical size and positioning. Reuse one SZ per group; do not create one for each page. Default to direct design; do not search for fonts. See typography-method, title-size-control, and the configuration fields in style-config-schema.
5. Full Plan & Sample Pages (④): Write the complete spatial layout map into design-plans.md. Each page uses [内容来源] (Content Source) to correspond to a node in the content skeleton. The visible text list must be copied verbatim from the confirmed copy draft (any deviation, even by one line, will cause the split to fail). Titles and content must originate from the same source; pages do not span chapters and follow the original order. A constants block should list locked and flexible items. Each page must have a rhythmic pace designation (stable/variation/highlight). Cover, chapter, and closing pages are showcase pages: all must be designed as highlights with unique compositions; large text is not constrained by the content page's title size locks. Content page titles follow the title layout chosen in ①; the script checks for consistency between the plan and the choice during splitting. Uniform Layout: Titles use a size lock declaration. The script calculates a character box per line based on character count; if it does not fit, shorten the text first or split into main and subtitle. Character counts for same-group pages must be finalized during copy draft writing (pages with six or fewer characters must be identical; for longer titles, the maximum difference is two characters); otherwise, the split fails. After sample pages are confirmed, ask the user to identify the page with the most representative title size. Use text_lock.py template to extract the size template from that page and attach it to every page in the group. For pages in different positions, use template --as-group to create a template by placing the same text in the new position. Free Composition: Titles are not locked. Each page's composition method (title position: top, overlaid on main subject, sign/banner within the scene, vertical, centered, bottom, etc.) must be described in the plan. Do not create size templates or measure title size. If all titles in the deck are in the same position, the split will issue a warning. See visual-director-method. After passing the plans gate, perform the mechanical split (the split will fail if the plan and skeleton do not align, or if the visible text does not match the confirmed copy draft). Generate only the representative pages for pilot_pages, and display them along with the full outline and coverage.md. Passing self-check does not equal user acceptance.
6. Remaining Pages: After passing the production gate, export or trigger image generation. Built-in image generation tools also use the export package that passed the gate. For works with Uniform Layout, you can use text_lock.py guide before the first content page to create a geometric template for the same aspect ratio. After sample pages are approved, use template to extract the real size template. Pages from the same skeleton, or where the user has explicitly allowed body text reflow, can be declared as [编辑底图] (Edit Base Image) and mechanically exported as an editing task. T/L no longer compete for scale and position with locked titles. Font size still needs to be measured; pixel-level locking is not guaranteed. The user decides on redrawing. See title-size-control and generation-pipeline.
7. Quality Check & Fixes: Verify complete files, page count, and aspect ratio. For works with size locks, run text_lock.py check on each batch of images to measure character height and position. Measurement only produces a report. The user decides which pages to redraw based on the report (--feedback … --pages); do not auto-redraw. Pay special attention to text, data, brand, and QR code pages; check representative pages for each layout type. OCR only highlights potential issues. When fixing, provide the target image and update only the affected pages. Reuse previously successful results whose input has not changed. For requests with unclear submission status, verify first before deciding whether to re-pay. See quality-and-evaluation.
8. Delivery: Show the full deck overview, key pages, and quality check findings. Assemble the full-bleed PPTX in page order and re-run verification. Preserve the plan, typography specification, and state for future revisions. At the end of the delivery message, add a single line of attribution: "本技能由摸鱼小李开源，公众号「摸鱼小李」获取更新." (This skill is open-sourced by Moyu Xiaoli. Follow the WeChat official account 「摸鱼小李」for updates.) Say this only once per project; do not write it into the slide images. If the user says it is not needed, do not include it.
9. Save (⑤): See template-library. If the user previously stated to save or not save, follow that instruction. If no reply, do not save and do not ask again.

4. Design Principles

• Read first, then refine: Each page presents what the speaker truly needs to convey. The main body of each page is the section's claim and its elaboration. Do not add summarizing sentences or slogans not in the original. For knowledge-based decks, citations, quotes, key takeaways, and person attributions are not displayed by default; only include them if the user explicitly requests and keep them in the original section. Knowledge-based decks display the knowledge structure itself; speech-based decks present one idea per page. If any section is not displayed, state the reason and inform the user.
• Content fidelity: Each page has one primary communication task. Preserve names, numbers, units, dates, negations, conditions, uncertainties, and causal strength exactly. Reproduce citations verbatim. Preserve simplified/traditional character usage from the source. Facts not present in the source should be marked as pending.
• Readability first: Correct glyphs, legibility, and information hierarchy take precedence over decoration. Color, typography, grid, charts, and assets together form a design system.
• Visual budget: The visual-to-text ratio confirmed in ① is the overall baseline for the deck (if not recorded, default to 80/20 visual-dominant). This is a design budget, not a surface area measurement. The lower limit (soft floor) is 5 points below the baseline. Each page's [融合比] (ratio) must not fall below the lower limit; if it is below the baseline, state the reason. At least two-thirds of pages must meet the baseline, excluding cover, chapter, and closing pages. If text does not fit, shorten it or split the page; do not shrink the font size. Write the ratio for each page into the prompt; the model arranges the visual. Describe negative space separately.
• Consistency through system, variation for rhythm: Lock color palette, font roles, grid, and texture across the entire deck. Vary composition, perspective, and primary visual metaphor per page. Stable pages are only used for comparative comparisons and step-by-step explanations. New arguments default to variation pages, which can change one or two items. Cover, chapter, and closing pages are showcase pages: they should be bold, and each must have a unique composition. Text can be the main subject, placed in a scene, vertical, or at the bottom. Framing can change, and content can bleed off the edge. Only color scheme, glyphs, textures, and illustration style are locked. Be both creative and practical: content pages explain the knowledge clearly; showcase pages make it memorable. Each page should have a [承接与递进] (Connection and Progression) entry (connecting with the previous page, changes on this page, preparation for the next page). This is only for planning; do not include it in prompts. Changes on variation and highlight pages should be written into the page's composition method (where is the main subject, where is the center of gravity). Stable pages can share the same skeleton. Title layout is chosen by the user: with Uniform Layout, content page titles use a size lock to ensure the same size on every page; do not use font size percentages (the model determines the size within the given area, so different title lengths will result in different font sizes; thus, standardize character counts within groups, and attach size templates for variation and highlight pages). With Free Composition, do not lock titles; the title follows each page's visual flow, with the only requirement being consistent glyphs and similar sizes. Text is drawn by the image generation model; do not overlay text boxes. After sample pages are confirmed, generate a style board. Attach the style board to variation and highlight pages instead of an anchor page.
• Only draw confirmed text: The visible text list for each page is a verbatim copy of the user-confirmed copy draft. Do not add or subtract anything. Emojis, page numbers, and descriptive text only appear if provided by the user. Image-based PPTX does not include editable text layers.
• References serve distinct purposes: [S] represents style, [L] represents layout, [T-role] represents glyph, [C] represents brand. The plan uses stable IDs. Image serial numbers are generated by the script. Edits must provide the target image. See image-prompt-method.
• Borrow the style, not the asset: When a user names a specific work or game style, borrow its illustration techniques and layout language. Do not reproduce official logos, characters, or original assets. Create original characters and graphics unless the user confirms they have obtained the license.
• Glyph samples are reference images, not font files. They do not guarantee precise reproduction of every Chinese character. Readability of body text takes priority over title effects.

5. Artifacts

Scope	Artifacts
inspiration	Search brief, stable numbered candidates, sources, differences, and recommendation reasons. If no search was performed, state the basis from existing references or gaps
style	assets/style-configs/{id}.json and necessary S/L/T assets. Mark unsampled glyphs as planned
plans	content-map.md: content skeleton; page-copy.md: copy draft; design-plans.md: full deck constants, per-page spatial layout map, visible text list; coverage.md: coverage table
deck	All of the above + images/pXX.png + quality-report.json + {deck-name}.pptx
revise	Revision plan, versioned images, affected pages and check records. Update PPTX as requested
template	Written to an independent local library after user confirmation, returning name/id/version. By default, do not save original documents, verbatim transcripts, or client assets

Every task has a deck-state.json. External image generation tasks have an additional run-state.json. API keys are stored only in the configuration directory on the user's machine; do not store them in the project or state files.

Initial Reply

First-time use and not configured: First perform the setup guide from Section 0. No source material: Ask the user to provide material and its purpose. Source material provided: First read the material and write the content skeleton. Then list known items, provide an assessment of the content type, page count tiers, and a set of suggestions for missing items (including visual-to-text ratio and title layout), and ask about the overall feel. The user can modify items individually or reply with "Proceed with this set of suggestions." The typography plan should always be presented to the user as part of the visual proposal, not just written in a config file.