Research Presentation

Research Presentation

Turn a paper or research corpus into a presentation that an audience can follow, verify, and discuss. The core promise is paper to slides without losing the evidence: preserve source meaning, choose a cumulative narrative, make figures legible, keep claims bounded, and deliver an editable deck with notes and a rendered QA trail.

This Skill owns the workflow from source inspection through final slide QA. It does not replace the author's scientific judgment, invent results, or silently strengthen a claim.

Route the request

Choose the narrowest route that satisfies the request:

Before building, infer one sentence:

By the end, [audience] should [understand/decide/discuss] because [central evidence-backed takeaway].

Ask at most one blocking question when a missing choice would change the deck materially (for example, defense versus journal club). Otherwise make a conservative assumption and state it in the run log.

Non-negotiable contracts

1. Evidence first. Read the source before planning the slide sequence. Never use a figure, number, table cell, citation, mechanism, or limitation that is not traceable to the supplied source or an explicitly cited external source.
2. Bounded claims. Preserve population, setting, time frame, comparison, uncertainty, and claim strength. Distinguish author conclusions from presenter interpretation and label the latter.
3. One narrative job per slide. Use takeaway titles that answer the question raised by the previous slide; do not build an inventory of paper sections.
4. One visual anchor per content slide. Use a readable original figure/table, a meaningful metric, or a simple source-faithful diagram. Avoid decorative diagrams and generic stock imagery.
5. Readable before complete. Start body text at 12 pt (11 pt floor), titles at 28–32 pt, cover title at 40–56 pt, and keep sub-11-pt characters below 10% of the deck. Add a slide or cut copy instead of shrinking type.
6. Source anchors everywhere. Every content slide carries a visible anchor such as Fig. 2, Table 1, §3.2, or p. 7; speaker notes include a [Sources] block for claims and assets.
7. No silent raster-only handoff. Prefer native/editable text, tables, shapes, and charts. Use source figures as images when needed, but preserve their aspect ratio and provide the original asset or source map.
8. Render before delivery. Render every final slide, inspect the full-size pages, run overflow/coverage/font checks, and fix unintended overlap, clipping, wrapping, stretched figures, and empty placeholders.

Paper-to-Slides workflow

1. Inventory and extract

Record the input files, target audience, language, venue, duration, expected slide count, output format, and available tools. Extract the entire paper text in one pass before section planning. Then work through the paper in order and write an evidence ledger containing:

• research question, gap, hypotheses, and contributions as the authors state them;
• study design, data, sample/population, baselines, parameters, and analysis steps;
• every meaningful quantity with unit, comparison, direction, and source pointer;
• every figure and table: what it shows, how to read it, and which panels matter;
• limitations, threats to validity, undefined terms, and out-of-scope claims;
• claims that are author judgments versus presenter synthesis.

For a PDF, use the host PDF tools or PyMuPDF/PDF extraction available in the environment. Use a lossless figure/table extraction tool when available; otherwise record the fallback and crop from rendered pages without changing the data. Read references/paper-extraction.md for the full ledger contract.

2. Choose the talk contract

Use the venue to set the arc:

• Journal club / lab meeting: question → gap → model/method → strongest evidence → limitations → discussion questions.
• Conference talk: problem → contribution → method → result → implication, with fewer side details.
• Thesis defense: problem → research program → methods → contribution-by-contribution evidence → limitations → future work.
• Seminar / teaching: motivating example → concept → worked evidence → application → recap.

Default to 12 content-bearing slides unless the user specifies another length. A 12-minute talk usually needs 10–14 slides; add pages when the evidence cannot remain readable.