Create image-based PowerPoint decks by (1) turning raw article content or notes into a detailed per-slide message plan when needed, (2) turning that message plan into a slide display plan and then a visual-production plan, (3) generating one 16:9 slide image per slide with all displayed text baked into the image (English by default; multilingual slide text supported), and (4) assembling an images-only .pptx that simply concatenates those images full-screen. Use when the user wants polished, consistent visuals with extensible style packs (cinematic dark, cinematic light, cinematic editorial, illustrative cinematic, animated feature, editorial, warm pastoral, tech, youth social, academic, corporate, whiteboard sketch), prefers not to hand-layout PPT objects, or wants a repeatable prompt workflow to iterate over time.
Create image-based PowerPoint decks by (1) turning raw article content or notes into a detailed per-slide message plan when needed, (2) turning that message plan into a slide display plan and then a visual-production plan, (3) generating one 16:9 slide image per slide with all displayed text baked into the image (English by default; multilingual slide text supported), and (4) assembling an images-only .pptx that simply concatenates those images full-screen. Use when the user wants polished, consistent visuals with extensible style packs (cinematic dark, cinematic light, cinematic editorial, illustrative cinematic, animated feature, editorial, warm pastoral, tech, youth social, academic, corporate, whiteboard sketch), prefers not to hand-layout PPT objects, or wants a repeatable prompt workflow to iterate over time.
Infographic PowerPoint Deck
Default input model (raw article first)
This skill should work directly from the user's raw article content by default.
If the user provides a full article, sermon, report, or long notes:
Read the source content and infer the audience, burden, density, and argument/story structure.
Build slides_message_plan.md from the article structure.
Turn slides_message_plan.md into slides_display_plan.md.
Turn slides_display_plan.md into slides_visual_plan.md.
Write one image prompt per slide and generate the deck.
If the user already provides slides_message_plan.md or an equivalent content-first slide table, you can skip the article-intake step and use that directly.
If the user provides one or more reference slides/screenshots, use references/reference_slide_intake.md to extract the reusable visual grammar before writing prompts.
Artifact model
Use three distinct planning artifacts:
slides_message_plan.md
content / presenter planning only
what each slide is meant to teach, explain, or say in full
the detailed speaking content or narrative burden behind the slide
source/article anchors, scripture/evidence anchors, and must-preserve claims
explicit on-slide text after text-budget and archetype decisions
can be much shorter than the message plan
slides_visual_plan.md
downstream visual-production planning
deck archetype
style pack choice
layout routing
scene choice
text budget and other production constraints
slides_message_plan.md is the source of truth for what the presenter means to communicate.
slides_display_plan.md is the source of truth for what text the audience actually sees.
slides_visual_plan.md is the source of truth for design and prompt production.
raw article content / notes and use references/article_intake_workflow.md, or
a prebuilt slides_message_plan.md, or
an underspecified content slide table that you will normalize first
optional reference slide images / screenshots and use references/reference_slide_intake.md
If starting from a raw article, create slides_message_plan.md first.
Infer one deck archetype from the approved message plan using references/deck_archetype_routing.md.
If the user explicitly requests a style, use that as an override.
Pick one style pack folder.
If user does not specify style and archetype does not strongly imply a different choice, default to editorial-light.
Style pack chooses the deck's visual language (palette, lighting, typography attitude, scene bias). It does not force one fixed layout for every slide.
Create slides_display_plan.md from slides_message_plan.md.
Decide what the audience should actually see on each slide after considering archetype, density, and likely layout family.
Do not force the full speaking content onto the slide.
Create slides_visual_plan.md from slides_display_plan.md.
Add style, layout, scene, and text-budget decisions here, not in the message plan.
Gather the style-pack material:
You may run scripts/compose_style_pack_blocks.py --pack-id <id> as a helper, or manually read the style-pack files.
Treat the script output as helper context only, not as the final prompt body.
The concrete prompt should still be authored by the model in plain instruction prose.
For each slide in slides_visual_plan.md, fill references/prompt_template.md:
Use the style-pack material as helper context and rewrite it into plain instruction prose.
Write the prompt instructions in English by default.
Paste exact required on-slide text from slides_display_plan.md in the user-specified language(s).
Add slide-specific scene layer details and icons using the visual plan.
Keep internal routing IDs such as L1 or L10 in planning metadata only. Translate the routed layout into plain composition language inside the concrete image prompt.
If using a reference image or edit-style tool call, keep that mechanic in the tool call. The concrete prompt should still read as result-oriented art direction such as Using the provided image as the compositional base..., not tool language such as edit this image.
Generate each slide as one 16:9 image.
Allowed slide-creation modes in this skill are image generation and image editing/reference only.
In the slide prompt text, explicitly include a ratio lock sentence (e.g., Hard canvas constraint: 16:9 widescreen. Do not generate a square image.).
In the tool call, explicitly set ratio config when supported (e.g., generation_config with aspect_ratio: "16:9").
If consistency needs reinforcement, you may use an approved prior slide image or background exploration image as an input/edit reference, but the image tool must still render the final full slide image.
QA readability + text accuracy; regenerate only failed slides.
If any slide is not 16:9 or has text defects, regenerate the image directly.
Do not crop, pad, resize, or post-process generated images unless the user explicitly asks.
Do not render text afterward with Python/PIL/PPT overlays as part of the normal skill workflow.
Assemble images-only PPTX with scripts/build_images_only_pptx.py.
Outputs to produce in the user’s workspace:
slides_message_plan.md (detailed per-slide speaking / teaching plan; source of truth for what each slide is meant to communicate)
slides_display_plan.md (audience-facing on-slide text plan; source of truth for displayed text and labels)
slides_visual_plan.md (visual-production plan; source of truth for archetype/style/layout/scene decisions)
prompts.md (the exact prompts used for each slide)
slides/slide01.png ... (final 16:9 images)
deck_images_only.pptx (concatenated images)
Optional structured slide table input
For higher-quality multi-step workflows, you may start from either a raw article or an already-structured message slide table:
Raw article content is a valid first-class input to this skill. Use references/article_intake_workflow.md.
If the upstream artifact is already slides_message_plan.md, preserve it as the message source of truth.
If the upstream artifact already decides deck length, section coverage, or which beats must stay on separate slides, treat that as the controlling message plan unless the user explicitly asks for further compression or expansion.
If the upstream content table is simple or underspecified, normalize it into slides_message_plan.md using references/slide_table_normalization.md before display planning.
Prompt stack (minimal vs full)
Minimal stack (recommended for most decks):
references/article_intake_workflow.md (when starting from raw article content)
references/deck_archetype_routing.md (infer the overall deck form first)
references/style-pack-catalog.md (choose pack ID)
references/style-pack-system.md (load order + composition rules)
references/layout_routing_policy.md (auto-route layouts by default)
references/prompt_example_library.md (use examples and anti-examples to avoid panel drift)
references/high_fidelity_prompt_playbook.md (when the user wants a more precise prompt)
references/reference_slide_intake.md (when user wants the deck to learn from a reference slide)
This avoids cross-file mismatch and makes new style creation deterministic.
Note: despite the Bible-themed examples, this workflow works for any topic. Swap scenes/props in references/scene_library.md to match your domain (product, research, training, etc.).
Decision model (keep style, layout, and scene separate)
Message plan = content-level communication model: what the presenter means to say on each slide and how many slides are needed to say it well.
Display plan = audience-facing text model: what short text, labels, verse fragments, bullets, captions, and module headers should actually appear on the slide.
Deck archetype = deck-level viewing model: what kind of visual presentation best fits the approved message plan.
Layout = slide-level composition grammar: split-panel, framework, comparison, warning, didactic teaching board, or full-bleed overlay. It should normally be auto-routed during visual planning. The best result may feel highly organized without looking like a rigid template.
Scene ID / visual scene = slide-level depiction choice chosen during visual planning.
Text budget = production constraint describing how much copy the chosen layout needs to carry.
Use the deck archetype to drive the rest of the decisions:
message plan -> slide roles, source anchors, and coverage depth
message plan + archetype + text budget -> display plan
display plan -> on-slide text and label density
deck archetype -> style pack
deck archetype -> visual-plan layout bias
slide roles + text budget -> layout routing
slide role + source meaning -> scene choice
optional reference slide -> prompt wording, layout bias, and possible style-pack refinement
Slide count is not a deck-skill default. It should be determined by how many slides are actually needed to deliver the content well.
Use this decision rule:
identify the distinct beats that need separate treatment
decide how much compression each beat can tolerate without hurting clarity, persuasion, or usability
set slide count from that coverage need
Upstream slide plans, agent instructions, runtime expectations, or user requests can constrain the deck, but they are not the core planning logic.
If an explicit brevity request would force harmful compression, preserve the content logic first and either:
keep the richer deck, or
consciously produce a compressed version while making that tradeoff explicit in slides_message_plan.md
Do not collapse a rich article into a short deck merely because the skill can make a shorter deck.
Infographic does not mean one universal text density.
Some decks should feel like sparse keynote slides.
Other decks should feel like self-contained didactic infographics or teaching posters with more visible words, labels, formulas, diagram captions, and explanatory sections.
Choose the density and layout family that best serves the content instead of forcing every deck toward the same low-text aesthetic.
Use one style pack across the deck by default, and let the skill choose layout per slide unless the user explicitly overrides it:
Short emotional or opener slides often route to L4, L5, or L6.
In cinematic, editorial, warm, airy, animated, and youth decks, short/medium context or application slides should try L4 or L6 before L1 if the scene can provide a calm text-safe zone.
Teaching, framework, comparison, and didactic poster slides often route to L1, L2, L3, L7, L8, L9, L10, or L11 when structure or density requires those layouts.
The same style pack can legitimately use both split-panel and full-bleed layouts inside one deck.
If you need article-first intake guidance, read references/article_intake_workflow.md.
If you need archetype routing guidance, read references/deck_archetype_routing.md.
If you need routing guidance, read references/layout_routing_policy.md.
Slide prompt recipe (copy/paste template)
Read references/prompt_template.md and fill it per slide. Keep it extremely explicit:
Start from the inferred article structure, message plan, display plan, and deck archetype, not from arbitrary slide decoration.
Put all required on-slide text from slides_display_plan.md into one exact-text block in the concrete prompt.
If the slide is a dense didactic infographic, include every visible section header, label, callout, formula label, diagram caption, and module heading in that exact-text block rather than expecting the model to invent them correctly.
For visuals, describe scene layers (far/mid/foreground), plus 3–8 concrete objects.
For high-fidelity results, do not stop at naming the style pack or the layout family. Also specify geometry, palette/material treatment, typography attitude, divider/module structure, and lighting behavior in plain prose.
Specify what must be subtle/low-contrast so text stays readable.
Keep prompt instructions in English by default even when the required on-slide text is Chinese, German, or bilingual.
Do not clutter concrete prompts with redundant language metadata in the default English case. Only state prompt/output language explicitly when the slide uses non-English or multilingual rendered text, or when the language setup might otherwise be ambiguous.
Use both presentation families intentionally: split-panel / board-style infographic layouts for dense teaching content, didactic poster layouts for self-contained educational slides, and full-bleed cinematic overlay layouts for title cards, quote slides, and scene-led emotional beats.
Treat split-panel wording as explicit layout instructions, not as generic presentation language. If the user prefers text directly on the image, route away from L1 before writing the prompt.
Treat L1-L11 as internal routing handles, not model-facing prompt text. The image model should receive composition instructions such as use a mirrored two-zone teaching board with a central divider, not use layout L9.
Treat style-pack files the same way: they are internal guidance material, not literal prompt fragments. The final prompt should read like direct art-direction and display instructions, not like serialized metadata.
Treat input-image and edit mechanics the same way: they are tool-side mechanics, not prompt-side prose. The final prompt should describe the desired finished slide, not the fact that an edit tool is being used.
If the final prompt still sounds generic after removing the metadata, enrich it rather than shortening it. High-fidelity prompts normally need explicit visual specificity, not just raw content and one style label.
For full-bleed layouts, default to direct text on the image itself with composition-based readability support. Do not ask for visible rounded rectangles, frosted cards, or caption boxes unless the user explicitly wants that treatment.
Do not ask the user for per-slide layouts by default. Route layouts internally from slide role, text budget, display density, and overall deck style. Only use explicit Layout hint when the user or slides_visual_plan.md provides one.
Do not ask the user for style-pack IDs by default when the article content already makes the fit obvious. Infer archetype first, then select style.
If you need article-first intake guidance, read references/article_intake_workflow.md.
If you need deck archetype selection guidance, read references/deck_archetype_routing.md.
If you need inspiration for more cinematic scene material and prop ideas, read references/scene_library.md.
If you need stable scene IDs + tags for selection/filtering, read references/scene-catalog.md.
If you need copy-ready visual scene blocks, read references/scene-preset-library.md.
If you need fast, reliable infographic compositions or full-bleed title-card layouts, read references/layout_library.md.
If you need automatic routing rules for when to use each layout, read references/layout_routing_policy.md.
If you need concrete good/bad prompt examples before writing prompts, read references/prompt_example_library.md.
If you need a higher-precision checklist for turning style/layout/scene choices into a concrete prompt, read references/high_fidelity_prompt_playbook.md.
If the user gives a screenshot or example slide and wants the skill to learn from it, read references/reference_slide_intake.md.
If you need shot-language / time-of-day / weather / lighting presets, read references/shot_mood_library.md.
If you need style modularity, read references/style-pack-system.md and use scripts/compose_style_pack_blocks.py.
If you want fewer typos and more readability, paste these into every slide prompt: references/typography_spacing_lock.md, references/text_fidelity_block.md, references/negative_prompt_block.md.
If you want storyboard-like viewing rhythm, read references/storyboard_library.md.
If a slide uses language-specific rendering constraints or long non-Latin quote blocks, read the relevant reference note before writing the concrete prompt (for example, references/chinese_quote_compression.md).
Quality bar (and how to iterate)
For each slide:
Message fit first: the deck should reflect the structure and intent of the article and the richer message plan, not just produce isolated pretty slides.
Display discipline second: the slide should show only what the audience needs to see, not the entire speaking content.
Readability first: text never overlaps busy imagery; use either a dedicated text panel or a deliberately clean overlay-safe region, depending on the selected layout.
Auto-layout first: unless the user explicitly overrides, let the skill choose the layout from slide role, text budget, and style rather than forcing one composition family.
No hallucinated text: forbid extra words/logos/watermarks.
Text fidelity in the requested language/script: if any character, accent, word, or punctuation is wrong, regenerate that slide with:
shorter quote blocks,
All required on-slide text must be exact. Do not rewrite. Do not add or remove punctuation or spaces.,
simpler backgrounds.
Engagement: add background scenes + textures + icon clusters, but keep them low-contrast.
Didactic density is allowed when the content needs it: a self-contained teaching slide may legitimately carry more visible words than a keynote slide, as long as hierarchy, spacing, and diagram structure stay clear.
Common failure fixes:
Deck feels visually polished but structurally wrong: revisit references/article_intake_workflow.md and the message/display boundary; fix the slide roles or display compression before rewriting prompts.
Tool call fails/intermittent errors: retry after a short wait; keep prompt stable; reduce scene complexity only if repeats.
Text too small: reduce bullet count; split into two slides; demand large type and comfortable line spacing.
Slide feels too empty for an infographic teaching use case: reroute to L9, L10, or L11 and explicitly include the needed labels, captions, formulas, or comparison headings instead of pretending the content is sparse.
Too busy: force a low-contrast background, keep the scene mostly on the right, and keep the left panel clean.
Deck too dark: switch to editorial-light or airy-relaxed, disable vignette, force daylight scene IDs, and add brightness override.
Full-bleed overlay feels boxed in: remove visible cards, rounded rectangles, frosted panels, and boxed banners; ask for direct text on the image with only subtle local contrast support.
Panel drift appears on a cinematic/editorial/warm slide: compare against references/prompt_example_library.md, reroute L1 to L4 or L6, and remove any text panel wording before retrying.
Auto-routed layout feels wrong: reroute the slide using references/layout_routing_policy.md; do not keep forcing a weak layout just because an earlier guess picked it.
Language/script fidelity drifts: shorten copy first, simplify the background, use the relevant language-specific rendering reference, and only then adjust style/layout.
Ratio mismatch: regenerate that slide with stricter 16:9 constraint in prompt; do not crop or otherwise post-process.
Ensure both prompt text and tool-call config specify 16:9 in the same retry.
Operating mode (be patient; avoid “probe images”)
Image generation can be slow. Prefer generating the real slides directly.
Avoid generating test/probe images (e.g. blank backgrounds) unless the user explicitly asks for diagnostics.
If an image call fails intermittently, retry the same slide prompt after a brief pause; do not change multiple variables at once.
Default delivery mode is raw outputs: keep generated images as-is and assemble the PPT directly from them.
If you create a background exploration image to stabilize later generations, treat it only as an image-tool reference input, not as a canvas for later text overlay or any non-image-tool composition step.
Tooling constraints (important)
Some image tools only allow writing output images to specific directories (often ~/Downloads or a workspace path). When blocked:
Generate images into an allowed directory (e.g. ~/Downloads/<deck_id>/).
Copy images into the project/workspace output folder.
Bundled scripts
scripts/compose_style_pack_blocks.py
Compose inherited style-pack blocks into one paste-ready bundle.
Use before prompt writing to prevent style mismatch.
Read references/prompt_template.md when you need a high-detail prompt skeleton that produces the “rich background infographic” style reliably.
Read references/article_intake_workflow.md when the user gives full article content and expects the skill to infer the deck strategy.
Read references/deck_archetype_routing.md when deciding what overall presentation form best fits the article.
Read references/style-pack-system.md for composition rules and how to add new style packs.
Read references/style-pack-catalog.md for intent-to-style routing and pack selection.
Read references/style-packs/ for self-contained style definitions.
Read references/slide_table_normalization.md to normalize underspecified upstream content slide tables into slides_message_plan.md before display and visual planning.
Read references/scene-catalog.md for reusable scene IDs and tags.
Read references/scene-preset-library.md for ready-to-paste scene visual blocks.
Read references/scene-entry-template.md when adding new scene IDs to the catalog.
Read references/scene_library.md to quickly add scene-based / cinematic elements (locations + props) without making the slide messy.
Read references/layout_library.md to choose a layout that matches your text density (so slides stay readable).
Read references/layout_routing_policy.md to auto-route layouts when the user does not explicitly request them.
Read references/shot_mood_library.md for cinematic shot + lighting + time-of-day presets.
Read references/motif_pack.md when you want optional recurring motif guidance for stronger deck cohesion.
Read references/deck_consistency_block.md when you want optional cross-slide consistency locks for light direction, rhythm, and texture behavior.
Read references/typography_spacing_lock.md to prevent tiny text and keep spacing consistent.
Read references/text_fidelity_block.md to reduce text errors and forbid any extra words in any language.
Read references/negative_prompt_block.md to avoid common generator artifacts (watermarks/unspecified text/UI).
Read references/storyboard_library.md to give the whole deck a narrative arc.
Read references/chinese_quote_compression.md only when the required on-slide text is Chinese and long enough to need splitting without paraphrasing.
