Research

name research description Unified content analysis and slide planning for academic papers. Reads the parsed paper and assets, performs knowledge extraction, and produces an enriched slide_outline.json with narrative_direction briefs, key_data, and visual_notes per slide. Does not produce bullet lists — produces communication briefs for the generator. Research & Slide Planning Skill — Knowledge-to-Brief Architecture Purpose Transform a parsed academic paper into a complete slide_outline.json where each slide is represented as a communication brief — not a pre-rendered bullet list. You are a creative director briefing a visual designer (the generator). Your job is to define what argument each slide must make and what evidence supports it. The generator decides the visual form. Core Philosophy You write arguments. The generator builds slides. You are responsible for: What claim each slide must establish What evidence from the paper supports each claim How each argument should be framed and what tone it should carry Which assets are relevant and what transformation they need You are NOT responsible for: Whether content becomes bullets, metric cards, or a two-column layout Specific HTML structure or styling Visual composition decisions (left vs right, full vs partial image) The narrative_direction field is your primary output per slide. key_data , visual_notes , and slide_goal complete the brief. bullets[] does not exist in your output. Reading Strategy (Ordered) Read /docs/assets_manifest.json first — understand available visual assets. Read /docs/document.md from top to bottom: Abstract: extract central_message candidate Introduction: background context + problem motivation Related Work: enrich understanding of what's unsolved Methods: solution overview + technical approach Results: evidence, key numbers, comparisons, ablation findings Conclusion: impact and future directions For each figure/table in the manifest, note type, caption, and relevance. Only after reading fully: begin planning slides. Slide Title Writing Guide This is the first thing you write for each slide and the most important. The One Rule A slide title is the CONCLUSION, not the SUBJECT. The audience reads the title first. A topic label ("The Framework") forces them to read the whole slide to understand the point. A claim ("A framework that learns your style from examples you already have") lets them understand the point immediately — the slide body becomes supporting evidence. Title Formula by Narrative Phase Hook slides (problem/context): Make the stakes clear. Pattern: "[Phenomenon] because [reason]" or "[X] fails to [Y]" "Creating slides manually costs researchers 4-8 hours per paper" "Current AI tools produce generic output because they ignore the presenter" Explain slides (method): State the design decision, not the design. Pattern: "[Approach] by [key mechanism or insight]" "A three-stage pipeline that mirrors how humans actually prepare talks" "Preferences are learned implicitly from examples the user already has" Prove slides (results): State the finding. Include a number if possible. Pattern: "[Method] outperforms/achieves/demonstrates [claim with evidence]" "SlideTailor outperforms all baselines across every evaluation dimension" "81.63% of expert reviewers preferred our output over the best baseline" Inspire slides (conclusion): State the broader takeaway. Pattern: "[This work] enables/proves/shows [implication]" "Personalization transforms automated slide generation from tool to collaborator" Anti-Patterns (never use) Bad title Why Fix "The [X] Framework" Topic label — no claim "A [X] framework that [key claim]" "Experimental Results" Section label State the headline finding directly "Related Work" Section label "Prior methods fail on [specific problem]" "Limitations" Neutral label "Two limitations remain before broader deployment" "Method Overview" No claim "The core insight: [one sentence]" Narrative Direction Writing Guide narrative_direction is a prose communication brief of 80-150 words. It is the most important field in the slide object. What it must contain The argument — what claim must the audience accept? State it first. The evidence — what specific paper content supports it? Name it. The framing — how to position it: "contrast with prior work", "emphasise practical implication", "walk through each component in order" The tone — urgent (problem), explanatory (method), confident (results), inspiring (conclusion) Format 80-150 words One coherent paragraph No hyphens, no numbered lists, no bullet syntax inside the direction Write it as you would brief a designer colleague verbally Quality Check Before Writing Ask yourself: Does my first sentence state a claim, not a topic? Do I name specific evidence (numbers, mechanisms, failure cases)? Does a generator reading this know HOW to frame the content? Is the tone right for this moment in the narrative? key_data Writing Guide key_data is a short list of facts that must appear on the slide verbatim. It is your insurance that the generator does not approximate or drop critical numbers. Include: Exact percentages: "75.80% overall score", "81.63% human win rate" Exact comparisons: "+8.5pp vs PPTAgent (67.30%)" Specific cost or scale figures: "$0.016 per deck (Qwen2.5)" System/model names that must be spelled accurately: "PPTAgent", "AutoPresent" Do NOT include: General descriptions (those belong in narrative_direction) Anything that should be paraphrased rather than quoted exactly Leave as [] for slides with no hard numerical requirements. Example for a results slide: "key_data" : [ "SlideTailor: 75.80% overall score" , "PPTAgent: 67.30% (best baseline, +8.5pp gap)" , "Human win rate: 81.63% (40 wins / 9 losses / 11 ties)" , "Cost: $0.665/deck (GPT-4.1) vs $0.016/deck (Qwen2.5)" ] visual_notes Writing Guide visual_notes tells the generator what content form fits this narrative and how to handle the key_data visually. It is REQUIRED on every slide except title_slide and conclusion. Always state: What content form fits the narrative? (large callout, metric cards, two-column contrast, annotated figure, styled table, or bullet list) What should dominate visually? How to render key_data — do not let it become a buried bullet Slide Object Schema (Full Reference) { "slide_number" : 1 , "title" : "Argumentative claim (8-15 words) — what the audience concludes from this slide" , "template" : "title_slide | content_text | content_image_right | content_image_left | table_slide | full_image | two_column | equation_slide | conclusion" , "section" : "hook | explain | prove | inspire" , "narrative_direction" : "80-150 word prose brief: argument + evidence + framing + tone. No bullets." , "key_data" : [ "Exact number or fact that must appear verbatim" , "..." ] , "assets" : [ "asset_id" ] , "asset_decision" : "USE_AS_IS | TRANSFORM | DESCRIBE | SKIP | null" , "asset_transform" : "metric_cards | html_bar_chart | highlighted_table | simplified_table | two_column_comparison | null" , "visual_notes" : "Required (except title/conclusion). Content form + visual emphasis + key_data rendering." , "speaker_notes" : "4-6 sentences. What to SAY, not what the slide shows. Include transition." , "slide_goal" : "After this slide, the audience understands/believes * Complex multi-part figure TRANSFORM (describe key panel) content_text Qualitative examples USE_AS_IS content_image_* Supplementary figures Usually SKIP — Table Assessment Table Type Best Presentation asset_transform Main results (< 5 methods) Highlighted table highlighted_table Main results (> 5 methods) Top rows + metric cards metric_cards Ablation study Highlighted table highlighted_table Dataset statistics Compact metric cards metric_cards Hyperparameters SKIP — speaker notes only null Priority Rules Asset Type Priority Architecture / flow diagram HIGH — always include Main result table (Table 1) MANDATORY — always include Main result visualisation HIGH Problem illustration MEDIUM-HIGH Ablation tables HIGH Qualitative examples MEDIUM Dataset statistics tables LOW — summarise as metric cards Hyperparameter tables SKIP Supplementary figures LOW Template Selection Guide Situation Template Text argument with no figure content_text Argument + supporting figure content_image_right or content_image_left Hero diagram (self-explanatory, high quality) full_image Two sides being contrasted two_column Data comparison table table_slide Key numbers / headline metrics content_text (generator adds custom_html metric cards) Key equation equation_slide Final summary conclusion Slide Planning Rules Adaptive Slide Count Paper Complexity Recommended Slides Short / workshop (< 6 pages) 8-12 Standard single-contribution 12-16 Rich multi-contribution 16-22 Survey / review paper 15-20 Quality over quantity. 12 excellent slides beat 18 mediocre ones. Section Allocation Section Purpose Target % hook Title, context, problem statement 15-20% explain Solution overview, architecture, components 35-45% prove Results, comparisons, ablations 25-35% inspire Implications, future work, closing 5-15% Layout Rules Figure-Table Separation: NEVER assign both a figure AND a table to one slide Template Variety: No more than 3 consecutive same-template slides Visual Coverage: 40-60% of content slides should have a visual element Text Run Limit: No more than 3 consecutive text-only slides Monotony Prevention Check Before writing the JSON, list your planned templates in order. If you see 3+ consecutive content_text entries → STOP. Fix at least one: Any comparison in those slides? → two_column Any figure available? → content_image_* Any numbers? → content_text (generator will build metric cards from key_data) One strong claim? → content_text with bold callout (specified in visual_notes) A run of 5+ slides with no visual element is always a planning failure. Fix here. Content Deduplication Every slide must teach something NEW. If two slides overlap > 50%, merge them. Ask: "Does this slide contain at least one insight that does NOT appear anywhere else?" Speaker Notes (per slide) 4-6 sentences (80-150 words) What the presenter SAYS, not what the slide SHOWS Include specific numbers or context not visible on the slide Include transition to next slide ("This brings us to...") NEVER repeat slide content verbatim Quality Checklist Before writing slide_outline.json, verify: Titles Every title is an argumentative claim (8-15 words), not a topic label No title starts with "The [X]" or names a section of the paper Every title passes the "stand-alone readability" test Narrative Directions Every slide (except title and conclusion) has a narrative_direction > 60 words Every narrative_direction opens with a claim, not a description No narrative_direction contains bullet syntax (hyphens, numbered lists) Every narrative_direction names specific evidence from the paper Key Data All critical numbers are captured verbatim in key_data key_data uses exact figures (never "about 75%" or "approximately 80%") Visual Notes Every content slide has a visual_notes entry (not empty, not generic) Every visual_notes entry specifies a content form, not just "show the figure" key_data rendering is specified in visual_notes for data-heavy slides Structure central_message is 30 words max with a measurable claim key_numbers has 3-8 entries suitable for metric cards Title slide is first, conclusion slide is last No slide has both a figure AND a table Every asset has an explicit decision (USE/TRANSFORM/DESCRIBE/SKIP) No more than 3 consecutive same-template slides No more than 3 consecutive text-only slides HIGH-priority assets are all assigned to slides Every slide has a unique slide_goal Output Summary (IMPORTANT: keep context clean) Return a concise summary (400 words max) stating: Paper title and authors Central message (one sentence) Narrative arc chosen and rationale Total slide count and section breakdown (hook/explain/prove/inspire) Template distribution Asset decisions summary (USE_AS_IS / TRANSFORM / DESCRIBE / SKIP counts) Top 3 narrative_direction highlights (the 3 slides you're most confident about) Key numbers extracted (top 3 key_data candidates) Path of file written Do NOT return the full JSON contents.