Paper To Talk

paper-to-talk

Overview

Build a presentation deck from source papers with three non-negotiable disciplines:

1. Source fidelity — every equation, figure, and attribution traces to a specific page/equation/figure in the source PDF. Nothing recalled from memory.
2. Clean slide design — the default slide has open space, one visual or one idea, sparse text, and no more information than can be spoken comfortably.
3. Argument structure — every assertion holds a (claim, evidence pointer, caveat) triple. Strong evidence → main slide. Weak or parametrization-sensitive → grain-of-salt slide.

Each slide has one thesis sentence. If you can't write that sentence, the slide isn't ready.

Design defaults

Optimize for a 15-minute spoken path plus discussion. Prefer fewer, cleaner slides over comprehensive coverage.

• Target 10-12 spoken frames plus title and backup for a 15-minute talk; use 12-14 only for a full 20 minutes.
• Use one primary figure, equation, or table per slide. If a slide has two visuals, remove most text.
• Keep most slides to 3 bullets or fewer; TL;DR and take-away slides may use 4 bullets.
• Keep bullets to one line when possible. If a bullet wraps twice, compress or split the idea.
• Put detailed derivations, reference tables, robustness catalogs, and source pointers in backup unless they are essential to the story.
• Leave a visible blank region on every slide. Do not fill all quadrants just because space exists.
• Prefer \small or normal size for spoken content. Use \scriptsize only for citations, compact tables, or backup.
• Use source lines as muted footers, not paragraphs.

When a slide feels dense, fix it in this order:

1. Delete secondary claims.
2. Move caveats to the grain-of-salt slide or backup.
3. Replace prose with a labeled figure or 2-3 word callout.
4. Split only if both halves are essential to the spoken path.

Workflow

1. Source extraction

For each paper:

• Fetch the arXiv abstract/PDF or use the user-provided PDF.
• Read the saved PDF page ranges. Hit in order: title page → abstract → intro → methods/theory → results → appendices.
• Capture verbatim: first 3 authors + et al, equation numbers, figure numbers, exact symbols, footnote caveats.
• Distinguish speaker (presenter at the talk) from paper author (first author of the PDF). They are not the same.

Never write equations from memory. For every equation on a slide, hold a pointer (paper, eq. N, page).

2. Recorded talk / transcript extraction

Use video and transcript material as interpretation context, not as slide filler.

For a recorded talk, conference page, or captioned video:

• Fetch the webpage and identify direct transcript/caption files first (.srt, .vtt, .txt, embedded JSON). Use the video only when captions are unavailable or visual content matters.
• If the user asks whether you can watch a video, answer honestly: you can analyze captions/transcripts and accessible video metadata; direct visual/audio inspection depends on available tools.
• Preserve timestamps in working notes, but do not paste long transcript excerpts into slides.
• Segment the talk into: setup, method, headline result, interpretation, caveats, and Q&A.
• Extract the speaker's emphasis, not every statement:
• what they repeated or framed as the main point,
• what they contrasted against prior work,
• what they treated as interpretation rather than measurement,
• what they were cautious about,
• what they said future work should do.
• Extract Q&A as discussion prompts:
• convert audience questions into short journal-club questions,
• include the speaker's answer only as a compact context clause when needed,
• group related questions instead of listing the full exchange,
• keep the discussion slide to 4-5 questions maximum.

Good slide patterns:

• "What the speaker emphasized" — 3 bullets plus a small diagram linking the paper result to the speaker's interpretation.
• "Discussion prompts" — 4-5 concise questions derived from Q&A, with one muted source footer.
• "Talk vs paper framing" — two columns with 2-3 rows: paper result, speaker emphasis, implication.

Avoid:

• transcript tables,
• timestamp-by-timestamp summaries,
• literal Q&A dumps,
• long quotes unless the user explicitly asks for them and copyright limits allow it.

3. Equation derivation discipline

When the equation you need is not in the text, or its derivation is missing, or the user asks for one, derive it. While deriving, sanity-check via at least two of:

• Dimensional analysis — both sides have the same dimensions.
• Unit consistency — SI / cgs / natural units balance.
• Scaling — the answer scales correctly in known parameters; no hidden factors of $k$, $a$, $H$, $c$, $\hbar$ etc.
• Limit convergence — reduces to a known formula in a known limit (Newtonian, weak-field, low-z, linear, deep-MOND, classical, non-relativistic, etc.).

If a derivation cannot pass at least two of these checks, do not put it on the slide. Flag the gap to the user.

4. Slide-flow checklist

Target ~10-12 spoken frames for 15 min; ~12-14 for 20 min:

1. Title — paper(s), authors, arXiv ID, your name, venue.
2. TL;DR — 3-4 bullets, one stating the headline number with its sigma / CI.
3. Method / estimator zoo — situate the paper among siblings. One row per family. Mark the row this paper occupies. If two papers share the session, mark both rows.
4. Context primer — 1 frame of background the audience needs but the paper assumes.
5. Concept / data setup — 1-2 frames.
6. Method — 1-2 frames covering the new derivation / parameterization / estimator.
7. Results — 1-2 frames with the paper's headline figures.
8. Author/talk emphasis — optional, only when a talk/transcript adds interpretation beyond the paper.
9. Discussion prompts — optional, especially for journal club; use Q&A material here rather than literal transcript.
10. Grain-of-salt slide — caveats, parametrization sensitivities, untested systematics, scope limits. Required, not optional.
11. Take-aways — 3-4 bullets mirroring TL;DR with what was actually shown.
12. Backup — key equations, derivations, references, robustness details.

Delete or move to backup before exceeding the target. A 15-minute journal-club deck should not try to teach every appendix.

5. Argument structure rule

For every assertion: hold a triple

• claim — what the slide says.
• evidence pointer — paper section / equation / figure number, or external citation.
• caveat — qualifier (parametrization sensitivity, missing systematic, scope limitation, weak prior).

Strong evidence + small caveat → main slide. Weak evidence or large caveat → grain-of-salt slide. No bare assertions.

6. Layout discipline (Beamer 16:9, 10pt)

Usable area: ~14.6 × 8.2 cm.

Density budget:

• No more than 35-45 words of body text on a visual-heavy slide.
• No more than 65 words on a text-only slide.
• No more than 5 displayed equations in the entire spoken path for a 15-minute talk; put the rest in backup.
• No more than one dense table in the spoken path. The estimator zoo may be that table.
• If a slide needs \tiny, it is backup or it needs deletion.

Bullets:

• Set \itemsep to at least 0.45em on main slides.
• Prefer 3 bullets with keywords in \acc{} or \wrn{} over 5 complete sentences.
• Use paragraph text only for a single thesis sentence or a source footer.

Tables:

• Sum of p{...cm} widths plus (n-1) × 2 × \tabcolsep must be .tex 2>&1 | grep -E "^!|Error|Output written|Overfull"

Iterate until the grep output contains only `Output written` (or vanishing residuals like `Overfull \vbox (0.5pt too high)`). **Ignoring `Overfull` warnings is a quality violation.**

**Visual verification loop (mandatory for substantial decks):**
- Render the PDF to PNGs or inspect it with a PDF viewer.
- Create a contact sheet for fast density review.
- Check that each slide has visible white space, readable labels, and no stale pages from earlier renders.
- If a slide looks like notes, convert it into a backup slide or cut it.

### 7. Notation hygiene

- Pin notation once per slide when symbols are ambiguous across papers ($T$ vs $\delta T$ vs $\bar T$, $g$ vs $\mathbf g$ vs $g_N$, $f$ as growth-rate vs filling-factor, etc.). Add a `\mut{...}` declaration.
- Use canonical macros (`\acc`, `\wrn`, `\mut`, `\Lcdm`, `\rd`) consistently across the deck. See `template-beamer-16x9.tex`.
- For multi-paper talks, harmonize notation across papers in the zoo slide; flag any cross-paper symbol clashes.

## Failure-mode catalog

| Failure | Counter |
|---------|---------|
| Misattributing the paper to the speaker presenting it | Check the PDF title page; first author ≠ speaker. |
| Pasting transcript or Q&A into slides | Extract emphasis and discussion prompts; keep transcript in working notes only. |
| Treating a talk as identical to the paper | Add a "speaker emphasis" slide only for differences in framing, caution, or interpretation. |
| Conflating differently-named-but-same-physics methods | Build the zoo slide early; explicitly distinguish overloaded terms in a caveat row. |
| Notation drift across papers | Pin convention once per slide via `\mut{...}` declaration. |
| Wrong scaling / dimensional error in a derivation | Re-derive from governing equation; check at least 2 of dim/unit/scaling/limit. |
| Inferring appendix content from the section name | Read the appendix. Always. |
| Ignoring `Overfull` warnings in the log | Compile-and-grep is mandatory before "done". |
| Equations recalled from memory | Trace every equation to a source pointer or derive with sanity checks. |
| Bare assertions without evidence | Every claim has a `(claim, evidence, caveat)` triple. |
| Main slides look like paper notes | Enforce the density budget; cut, move to backup, or replace text with a figure. |
| Tables / images that overflow the frame | Compile-and-grep. Adjust column widths, font size, or split content. |
| Cropped or off-center full-page images | Use `keepaspectratio` for letterbox or drop it for full-bleed; never let Beamer infer placement. |

## Bundled assets

- `template-beamer-16x9.tex` — preamble with color palette and macros, sample frames (title, TL;DR, 2-col, table, full-bleed image).
- `compile-check.sh` — wrapper running `pdflatex -interaction=nonstopmode -halt-on-error` and grepping the log for issues.

## Red flags — stop and verify

- "I'll write this equation from memory" → No. Derive or extract from PDF.
- "The compile is mostly clean" → grep the log; count overfull boxes; decide explicitly.
- "Speaker = author, probably" → Check the PDF title page.
- "The appendix probably does X" → Read the appendix.
- "Close enough on the centering" → Render the PDF and look; don't infer.
- "This estimator is the same as that one, I think" → Build the zoo slide; force yourself to articulate the distinction.
- "I'll skip the grain-of-salt slide, the result is solid" → No. Every paper has caveats. The grain-of-salt slide is non-negotiable.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [MinhMPA](https://github.com/MinhMPA)
- **Source:** [MinhMPA/paper-to-talk](https://github.com/MinhMPA/paper-to-talk)
- **License:** MIT

Install and usage instructions live in the source repository linked above.