slider

# slider > Research slides are evidence displays with narrative pacing. > The LM has creative freedom to compose each slide to best communicate its content. Turn an arXiv paper into a ready-to-present conference talk, output as a single self-contained HTML file. ## Priority order 1. **Technical faithfulness.** Never invent claims, numbers, or evidence. 2. **Evidence presentation.** Main deck must show the paper's key visuals, not just text summaries. 3. **UI polish.** Restrained, projector-safe, research-grade design. ## Core constraints - Output is a **single self-contained HTML file**, minified and ready to open directly in any browser. No build step required by the user. - The deck is a **rigorous and comprehensive talk**. Each paper must generate **at least 15 slides** targeting a ~15-minute presentation. - Every slide has **one dominant message**. - Main deck serves **audience understanding**, not paper section order. - Paper figures, tables, and diagrams are **first-class content**, not optional decoration. ## Renderer contract The deck uses a **fixed 1200×675 coordinate system** (stage model). This contract applies to all rendering modes: - **Stage-consistent fullscreen.** Both windowed and fullscreen modes render the same fixed canvas. Fullscreen only changes the `transform: scale()` factor — it never reflows content, never switches typography units, and never changes alignment logic. The slide looks identical at any scale. - **No viewport-unit typography.** All font sizes, padding, and gaps inside slides use `rem`, `px`, `em`, or `clamp()`. Never `vh`/`vw` for slide content. Viewport units are only permitted for viewport chrome (body background, controls positioning). - **No fullscreen-specific layout overrides.** If a slide overflows at 1200×675, fix the content or choose a different layout — do not add fullscreen-only CSS band-aids. - **Composition stability.** A presenter must be able to rehearse in windowed mode and present in fullscreen with zero visual drift. Text wrapping, vertical alignment, whitespace distribution, and element proportions must be identical. ## References ### Evidence and visual policy - paper visual inclusion policy: [references/paper-visual-inclusion-policy.md](references/paper-visual-inclusion-policy.md) - research visual priority: [references/research-visual-priority.md](references/research-visual-priority.md) - faithful redraw policy: [references/faithful-redraw-policy.md](references/faithful-redraw-policy.md) - paper figure embedding pipeline: [references/paper-figure-embedding.md](references/paper-figure-embedding.md) - paper type policy: [references/paper-type-policy.md](references/paper-type-policy.md) ### Design system - design control plane: [assets/design/DESIGN.md](assets/design/DESIGN.md) - deck design specification: [references/deck-design-control-plane.md](references/deck-design-control-plane.md) - external design principles: [references/external-design-principles.md](references/external-design-principles.md) - research slide patterns: [references/research-slide-patterns.md](references/research-slide-patterns.md) - page-role layout families: [references/page-role-layout-families.md](references/page-role-layout-families.md) - anti-patterns: [references/research-slide-anti-patterns.md](references/research-slide-anti-patterns.md) - fullscreen contract: [references/fullscreen-contract.md](references/fullscreen-contract.md) - reading comfort rubric: [references/reading-comfort-rubric.md](references/reading-comfort-rubric.md) ### Workflow and rendering - talk arc and compression: [references/presentation-logic.md](references/presentation-logic.md) - output mode selection: [references/output-contracts.md](references/output-contracts.md) - paper ingestion: [references/paper-ingestion-workflow.md](references/paper-ingestion-workflow.md) - figure and table triage: [references/figure-table-extraction.md](references/figure-table-extraction.md) - visual selection scoring: [references/visual-selection-scoring.md](references/visual-selection-scoring.md) - slide-data planning: [references/slide-data-generation.md](references/slide-data-generation.md) - pdf visual export: [references/pdf-visual-cropping.md](references/pdf-visual-cropping.md) - artifact rendering: [references/artifact-rendering.md](references/artifact-rendering.md) - slide quality rubric: [references/slide-quality-rubric.md](references/slide-quality-rubric.md) - bug and conciseness audit: [references/bug-and-conciseness-audit.md](references/bug-and-conciseness-audit.md) - verification workflow: [references/verification-workflow.md](references/verification-workflow.md) ### UI and interaction - UI pattern mapping: [references/ui-patterns.md](references/ui-patterns.md) - presenter UX: [references/presenter-ux.md](references/presenter-ux.md) - progressive builds: [references/progressive-builds.md](references/progressive-builds.md) - dual-screen speaker view: [references/dual-screen-speaker-view.md](references/dual-screen-speaker-view.md) - live tweaks protocol: [references/live-tweaks-protocol.md](references/live-tweaks-protocol.md) - HTML engine rules: [references/html-engine-template.md](references/html-engine-template.md) - React rules: [references/react-implementation-rules.md](references/react-implementation-rules.md) - multi-file React project: [references/multi-file-react-project.md](references/multi-file-react-project.md) ### Visual processing - auto-detection strengthening: [references/auto-detection-strengthening.md](references/auto-detection-strengthening.md) - multi-panel figure detection: [references/multi-panel-figure-detection.md](references/multi-panel-figure-detection.md) - multi-panel geometry: [references/multi-panel-geometry-detection.md](references/multi-panel-geometry-detection.md) - panel boundary detection: [references/panel-boundary-detection.md](references/panel-boundary-detection.md) - panel separator detection: [references/panel-separator-detection.md](references/panel-separator-detection.md) - separator geometry: [references/separator-geometry-refinement.md](references/separator-geometry-refinement.md) - table-aware crop: [references/table-aware-crop-refinement.md](references/table-aware-crop-refinement.md) - table row and column inference: [references/table-row-column-inference.md](references/table-row-column-inference.md) - semantic table parsing: [references/semantic-table-parsing.md](references/semantic-table-parsing.md) - semantic table role labeling: [references/semantic-table-role-labeling.md](references/semantic-table-role-labeling.md) - semantic role confidence: [references/semantic-role-confidence-calibration.md](references/semantic-role-confidence-calibration.md) - confidence calibration: [references/confidence-calibration.md](references/confidence-calibration.md) - visual binding confidence: [references/visual-binding-confidence.md](references/visual-binding-confidence.md) - metrics visual rendering: [references/metrics-visual-rendering.md](references/metrics-visual-rendering.md) ### Styling - theme presets: [references/theme-presets.md](references/theme-presets.md) - venue style variants: [references/venue-style-variants.md](references/venue-style-variants.md) - venue visual thresholds: [references/venue-specific-visual-thresholds.md](references/venue-specific-visual-thresholds.md) - venue appendix policy: [references/venue-specific-appendix-policy.md](references/venue-specific-appendix-policy.md) - redundancy suppression: [references/redundancy-suppression.md](references/redundancy-suppression.md) - score-aware slide insertion: [references/score-aware-slide-insertion-refinement.md](references/score-aware-slide-insertion-refinement.md) ### Platform - platform adaptation guide: [references/platform-adaptation-guide.md](references/platform-adaptation-guide.md) ### Meta - architecture: [ARCHITECTURE.md](ARCHITECTURE.md) - test matrix: [TEST_MATRIX.md](TEST_MATRIX.md) - changelog: [CHANGELOG.md](CHANGELOG.md) ## Bundled scripts - evidence extractor: [scripts/extract_paper_evidence.py](scripts/extract_paper_evidence.py) - visual candidate locator: [scripts/find_visual_evidence.py](scripts/find_visual_evidence.py) - visual scorer: [scripts/score_visual_candidates.py](scripts/score_visual_candidates.py) - slide-data generator: [scripts/generate_slide_data.py](scripts/generate_slide_data.py) - pdf visual exporter: [scripts/export_pdf_visuals.py](scripts/export_pdf_visuals.py) - visual asset binder: [scripts/bind_visual_assets.py](scripts/bind_visual_assets.py) - visual asset report: [scripts/visual_asset_report.py](scripts/visual_asset_report.py) - slideshow renderer: [scripts/render_slideshow_artifact.py](scripts/render_slideshow_artifact.py) - browser slide audit: [scripts/browser_slide_audit.py](scripts/browser_slide_audit.py) - research slide audit: [scripts/audit_research_slides.py](scripts/audit_research_slides.py) - design comfort audit: [scripts/audit_design_comfort.py](scripts/audit_design_comfort.py) - deck design compiler: [scripts/compile_deck_design.py](scripts/compile_deck_design.py) - health audit: [scripts/check_skill_health.py](scripts/check_skill_health.py) - platform export: [scripts/export_platform_configs.py](scripts/export_platform_configs.py) - offline KaTeX bundler: [scripts/offline_katex.py](scripts/offline_katex.py) - template retrofit: [scripts/retrofit_to_template.py](scripts/retrofit_to_template.py) ## Starter assets - HTML starter: [assets/html-slideshow-starter/paper-presentation.html](assets/html-slideshow-starter/paper-presentation.html) - React starter: [assets/react-slideshow-starter/App.tsx](assets/react-slideshow-starter/App.tsx) - React project starter: [assets/react-project-starter/](assets/react-project-starter/) - Design system: [assets/design/DESIGN.md](assets/design/DESIGN.md) ## Source handling Accept: arXiv URL, paper title, PDF, extracted text, DOI. Prefer full paper over abstract-only input. If only a title is given, resolve it first. If figures matter, visually inspect the PDF pages. --- ## Required workflow ### Step 1. Ingest and normalize Follow [references/paper-ingestion-workflow.md](references/paper-ingestion-workflow.md). Extract at minimum: - title, authors, venue, year - one-sentence TL;DR - problem, motivation, prior gaps - core idea, main contributions - method overview, architecture, objective/loss - datasets, metrics, baselines - main results, ablations, qualitative findings - limitations - visual candidates (all figures and tables in the paper) ### Step 2. Classify paper type Classify the paper using [references/paper-type-policy.md](references/paper-type-policy.md): - architecture-heavy - optimization-heavy - benchmark-heavy - theory-heavy - qualitative-heavy A paper may have multiple types. Record the classification. ### Step 2.5. Compile deck design Before rendering, compile a deck-level design using [references/deck-design-control-plane.md](references/deck-design-control-plane.md). 1. Run `scripts/compile_deck_design.py` with paper type, venue, and any user style preferences. 2. The script outputs `deck_design.json` — the intermediate representation consumed by the renderer and auditor. 3. The design locks: mood family, semantic palette, type/spacing scales, page-role→layout-family bindings, motion policy, content-handling rules, anti-pattern enforcement, and fullscreen contract. 4. The renderer must respect all bindings in `deck_design.json`. Ad-hoc visual decisions during rendering are prohibited. 5. If the user requests a style change, re-run the compiler — do not patch individual slides. ### Step 3. Identify mandatory main-deck visuals Apply [references/paper-visual-inclusion-policy.md](references/paper-visual-inclusion-policy.md) and [references/research-visual-priority.md](references/research-visual-priority.md). 1. List all visual candidates. 2. Classify each as Priority A, B, or C. 3. Determine which Priority A visuals are mandatory for the main deck based on paper type. 4. For each mandatory visual: can it be cropped cleanly? Needs faithful redraw? Needs reconstruction? 5. Apply suppression rules from [references/research-visual-priority.md](references/research-visual-priority.md). ### Step 4. Export and inspect key visuals When a PDF is available: 1. Run `scripts/find_visual_evidence.py <source>` to locate candidates. 2. Run `scripts/score_visual_candidates.py` to rank them. 3. Run `scripts/export_pdf_visuals.py <paper.pdf>` to crop candidates. 4. Visually inspect the crops. Decide: use as-is, faithful redraw, or reconstruct. 5. Apply [references/faithful-redraw-policy.md](references/faithful-redraw-policy.md) for any redraws. **Paper figure auto-embedding (MANDATORY for self-contained output):** When paper figures are identified and cropped: 1. Convert the cropped PNG/JPG to base64. 2. Embed directly as `<img src="data:image/png;base64,...">` in the HTML. 3. This ensures the output is fully self-contained — no external file dependencies. 4. Set the `visual` field on the slide data with `src` pointing to the base64 data URI. 5. If a figure cannot be extracted (e.g., no PDF available), use a faithful redraw or clean placeholder instead. 6. **Never leave a `mustIncludeVisual=true` slide without a visual.** Extract, redraw, or reconstruct. ### Step 5. Create slide plan with content composition Build the slide plan using [references/slide-data-generation.md](references/slide-data-generation.md). Before writing the final `slides` array, lock four short planning artifacts. These can stay inline in the conversation unless the user explicitly wants files: - `deckBrief` — audience, talk length, thesis, and delivery tone. - `deckScript` — page roles and one-sentence thesis per slide. - `visualPlan` — page-aware visual choices. For each image-bearing slide, state which figure/table/sample is needed and why it earns space. - `designLock` — one primary visual direction, up to two supporting directions, and a motion policy. Do not let the deck drift into unrelated slide themes. **This must be consistent with `deck_design.json`.** The goal is to lock the deck as a presentation before rendering HTML. Do not skip directly from paper extraction to decorative code. **You have creative freedom to compose each slide's layout.** Choose between: **Option A — Classic layout** (for standard patterns): - Set the `layout` field to a pattern name (cover, method-overview, split, etc.) **Option B — Content blocks** (preferred for custom compositions): - Set `contentBlocks[]` with ordered content primitives - Each block has a `type`, `content`, optional `placement`, `style`, and `emphasis` - Optionally set `gridTemplate` for CSS Grid control and `atmosphere` for per-slide mood For each slide, specify: - `id`, `title` (argumentative), `purpose` - `pageRole` — what job this slide plays in the talk arc (hook, bottleneck, method overview, headline result, limitation, close) - `keyMessage`, `claim`, `speakerNote` - `mood` — chosen from the mood library only after the `designLock` is set. Default to one primary deck mood family and use stronger contrast moods paper-slideonly where the story needs them. - `evidenceType`, `evidenceSource`