Research Presentations

Research Presentations

Use this skill for research presentation planning and quality control. File creation, object editing, export, and rendering should be handled by the official Presentation/Slides or LaTeX capability when available.

Boundary

• Use for group meeting, academic talk, seminar, conference, journal club, defense, methods/model/result update, or Asteria/TRACE-to-deck work.
• Do not use for minor text, color, alignment, or object edits to an existing PPTX/Google Slides deck.
• Do not use generated whole-slide images as a substitute for editable PPTX unless the user explicitly asks for image/PDF slides.
• Use presentation themes/templates for deck-wide color. Scientific palettes in palette/ may inform embedded figures, but raw palette ids should not become the slide theme.
• Choose the presentation format from the user's deliverable: PPT, PowerPoint, .pptx, editable, Slides, Google Slides, or "I need to edit it later" means editable Presentation/Slides; Beamer, LaTeX slides, .tex, academic PDF, or a venue/project-locked TeX template means Beamer/LaTeX; outline/storyline-only requests can stop at the deck plan.
• In the presentation-desktop profile, an unspecified group meeting, research update, paper talk, journal club, seminar, defense, methods/results update, or "research slides" request defaults to the exact CUHK Beamer route: source-editable .tex from ../../shared/templates/cuhk/beamer/source/ plus rendered PDF.
• Use ../../shared/deck-plan.schema.json as the default intermediate representation.
• Use metadata.mode: research-group-meeting for PhD group meetings, supervisor discussions, and research progress updates where the primary job is to update beliefs, expose failures, define next experiments, or ask for a supervisor decision.
• research-group-meeting mode must create a Research State and Evidence Board before slide planning. The state fields are internal planning data and must not be pasted mechanically onto slides.
• Chinese slide text uses writing-fidelity plus chinese-prose for final wording; English scientific slide text can use scientific-prose. These are handoffs to installed writing skills, not duplicate writing rules inside this skill.

Narrative

Research decks should answer:

why this was done -> what changed -> mechanism/method -> evidence -> limits -> discussion needed -> next step

Group-meeting decks should additionally answer:

previous question -> prior belief -> new evidence -> belief update -> failures -> largest uncertainty -> next discriminating experiment -> decision needed

Do not expose internal labels such as belief_update, evidence_quality, route promotion, handoff, or deck implication as slide body text. Translate them into scientific questions, evidence, limits, failure diagnosis, and advisor decisions.

Workflow

1. Read source material and identify source anchors: Markdown sections, PDF pages, figures, tables, code outputs, prior slides, or review comments.
2. For research-group-meeting, build the Research State and Evidence Board first. Inventory available figures, medical images, qualitative examples, quantitative plots, model diagrams, equations, experiment logs, failed experiments, literature figures to redraw, and missing evidence.
3. Choose page archetypes from the scientific job: RESULT_FIGURE, FAILURE_CASE, MEDICAL_IMAGE_COMPARISON, STATISTICAL_MODEL, METHOD_DIAGRAM, EXPERIMENT_DESIGN, NEGATIVE_RESULT, RESEARCH_UPDATE, NEXT_EXPERIMENT, or SUPERVISOR_DECISION.
4. Produce deck-plan.yaml before creating non-trivial slides. In research-group-meeting mode, every slide must include page_function, required_evidence, source_evidence_ids, scientific_objects, evidence_status, layout_rationale, allowed_fallback, forbidden_fallback, and qa_criteria. Run planning validation while evidence is still being gathered, then final validation before generation.
5. Keep one main research action per slide. Slide titles may be claim titles for mature results, but exploratory updates, negative results, and uncertainty pages may use question or in-progress titles.
6. Decide the format from the user's requested deliverable:
   • PPT, PowerPoint, .pptx, editable, Slides, or later manual edits -> editable Presentation/Slides route.
   • Beamer, LaTeX slides, .tex, academic PDF, or a locked TeX venue/project template -> Beamer/LaTeX route.
   • Group meeting, research update, paper talk, journal club, seminar, defense, methods/results update, or research slides in a desktop presentation context with no format specified -> exact CUHK Beamer route, using source-editable .tex plus rendered PDF.
   • Outline, storyline, or page-by-page plan only -> stop at deck-plan.yaml. The normal one-call exact-CUHK production entry is ../../shared/scripts/generate_research_presentation_production_entry.py --input-bundle <bundle.json> --out-dir <output-dir>. It reads a file/path-oriented research bundle, creates the source-fidelity map and deck plan, performs normal gold selection, consumes Stage 3 layouts, renders .tex -> PDF -> PNG, and writes task-local visual-review inputs. Do not call benchmark generators as the production entrypoint. The official existing-deck revision completion gate is ../../shared/scripts/validate_existing_deck_revision_entry.py --revision-packet <packet.json> --out <summary.json>. Use it for targeted revisions of a reviewer-seen deck after the revised source has been rerendered. It consumes the reviewer-seen baseline, accepted-element ledger, targeted feedback, rerender evidence, high-resolution problem pages, first-use dependency checks, rendered scientific-object QA, English final pass, and independent visual review. It may close only as PASS_REVIEWED; otherwise it returns REVISE or BLOCKED.
7. If a slide claim has no real evidence, convert it to missing evidence, next experiment, speaker notes, backup, or delete it. Do not fill it with rounded cards, icons, slogans, empty tables, or generic arrows. If the frozen plan or user request requires real plot/image/table/data evidence, a fabricated proxy, decorative graphic, or conceptual illustration cannot satisfy that requirement. Conceptual grounding is allowed only when clearly labeled as conceptual and kept separate from evidence claims.
8. Preserve equations in LaTeX inside the deck plan. Before compiling .tex or producing Beamer/PDF, invoke the locally installed render-chinese-math-pdf skill. Use that skill to probe for the LaTeX compiler, TeX packages, font availability, writable TeX caches, local render resources, and PDF QA tools instead of assuming xelatex/lualatex paths. On the YuukiAS workstation, preserve /home/yuukias/render_resources/chinese_math_pdf as the current TRACE-compatible render resource location when the probe reports it. If the local skill is not installed or any required resource is missing, block and report the exact missing dependency for user installation rather than substituting a retired server-only path.
9. Use the CUHK default template when no stronger project, course, company, or conference template is specified. For the default academic/research route and for exact CUHK Beamer reproduction, use ../../shared/templates/cuhk/beamer/source/ as the canonical source; the title slide layout is locked and only content fields such as title, subtitle, author, institute, and date may change. ../../shared/templates/cuhk/design-tokens.json, ../../shared/templates/cuhk/beamer/main.tex, and ../../shared/templates/cuhk/pptx/ are derived/non-exact scaffolds and must not be treated as the exact production source.
10. After file creation, render the deck to PDF/images and run scientific visual QA. The normal one-call route also writes a deck contact sheet, sequence summary, and bounded quality-loop state so an independent reviewer can judge deck-level rhythm instead of aggregating page PASS results. A deck is not complete merely because a file exists or mechanical checks pass. For editable PPTX regression, the PDF/images must come from the PPTX through a real presentation engine such as soffice, libreoffice, or an explicitly configured renderer. For English research slides, run a scientific-prose final pass only after scientific structure, formulas, claims, and citations are frozen; then rerender before final QA.
11. If deck-level visual review finds a clear blocker, consume only the structured reviewer evidence and apply at most one source-faithful repair cycle. Allowed repairs are bounded to source-dependency reorder, transition cue adjustment, overdense page split, redundant page merge/removal, compatible gold-layout swap, primary-object rescale, or existing annotation/legend repair. Unknown findings, unsafe mappings, or blockers after the single repair cycle must fail closed with no winner.
12. Use the research presentation reference library by page function, scientific domain, statistical subdomain, and evidence type. Keep downloaded source assets in .cache/research-presentation-reference-library/; commit only metadata indexes and page-level lessons.
13. For real-world deck revisions or user/advisor feedback loops, read references/real-world-presentation-guardrails.md after this file. It is the runtime entry for mature cross-project revision rules; source-only maintenance TODOs and project-specific histories stay outside ordinary installed plugin payloads. Natural requests such as "继续按这些批注返修我现有的组会PPT", "这个版本其他页别动，只修我指出的页面", or "保留上轮接受的布局，继续改剩下的问题" are existing-deck revision requests, not authorization to regenerate a new deck from scratch.

Revision Scope

When revising an existing rendered deck from user, advisor, reviewer, or visual-review feedback, treat the cited problem as a regression constraint for the next version. Keep an accepted_element_ledger in the working notes or deck plan when feedback is targeted: record which slides/components were accepted, which can only be locally adjusted, and which may be restructured. A local correction does not authorize a global redesign, removal of accepted structures, or restoration of a previously rejected layout. Compare the revised render against the version the reviewer actually saw; unrelated large visual changes must be justified or reverted.

Evidence And Concept Grounding

First-use notation and abstract concepts should be grounded in the audience's current context. Define what the symbol or concept is, where it comes from, and what role it plays in the model or experiment. When active data or audited source material contains a real example, prefer a short real example over placeholders such as group 1 or toy evidence. If only a conceptual example is available, label it as conceptual and do not use it to satisfy a real-evidence requirement.

Diagram Gate

Use diagrams only when the relationship, computation, mechanism, or experimental path is itself scientific content. Do not draw a diagram that merely places prose into boxes. Nodes must be real scientific objects or operations, and connectors must be structural connectors that encode data flow, dependency, transformation, control flow, time order, or experimental comparison. Arrows should be generated as connectors with semantic anchors and consistent direction, not typed arrow characters or decorative lines. Containment should use enclosure, braces, group labels, or shared background rather than misleading arrows.

Research Group Meeting QA

Rendered-slide QA must answer these questions per page:

• Does the page have a real scientific object?
• What evidence supports the page?
• Does the visual encode data, mechanism, experimental unit, case, formula, comparison, or uncertainty?
• Can the main question or result be understood in 5-10 seconds?
• Are figures, axes, legends, case labels, and formulas readable?
• Does the page look like consulting, a report page, or a card dashboard?
• If decorative frames are removed, is scientific content still present?
• Is the page worth 30-90 seconds in a group meeting?

If the answer is clearly no for scientific content, the page must not pass.

Deck-level QA must also inspect a real contact sheet or equivalent full-sequence object for adjacent density jumps, repeated layout face, overfull/empty pages, redundant filler, result-to-failure-to-next-experiment rhythm, workstream transition balance, and title/object/formula/image alternation. The top-level review package PASS is not enough without item-level judgement and observations for the deck-level object.

The generator must not assign final PASS. It may write PPTX, raw assets, evidence manifests, contact sheets, quality-loop state, and render status. An independent scientific visual reviewer must read the plan/evidence/rendered PNGs/reference IDs/contact sheet before assigning PASS, REVISE, or BLOCKED. If real PPTX rendering is unavailable, status is BLOCKED_REAL_PPTX_RENDER; do not reconstruct a parallel PDF and claim it represents the PPTX.

Anti-Patterns

Treat these as QA failures when they replace evidence: title plus slogan plus giant empty table; rounded-card dashboard; consulting language; generic arrows without scientific objects; fake visualization; paragraph pasted onto slide; unreadably shrunk paper figure; every slide using the same layout; decorative icons replacing evidence; vague next steps; internal planning language leaking onto slides; evidence-free roadmap.

If an urgent real group meeting must be delivered before the editable PPTX path passes regression, use mature Beamer or the user's older template first. Plugin experimentation must not block the real meeting.

References

• references/real-world-presentation-guardrails.md
• ../../shared/deck-plan.schema.json
• ../../shared/template-routing.md
• ../../shared/ppt-skill-routing.md
• ../../shared/source-fidelity.md
• ../../shared/visual-qa.md
• ../../shared/references/RESEARCH_GROUP_MEETING_MODE.md
• ../../shared/references/RESEARCH_SLIDE_ARCHETYPES.md
• ../../shared/references/RESEARCH_PRESENTATION_ANTIPATTERNS.md
• ../../shared/references/research_slide_reference_index.csv
• ../../shared/references/reference_sources_manifest.json
• ../../shared/references/reference_source_search_matrix.csv
• ../../../../../docs/workflows/RESEARCH_PRESENTATION_REFERENCE_LIBRARY.md
• ../../shared/templates/cuhk/