present-paper

name	present-paper
description	Academic presentation preparation — paper-driven (journal club, grand rounds, seminar) and lecture/teaching decks (course material, workshop slides, conference talks). Analyzes source material, finds supporting references, drafts audience-adapted speaker scripts, generates or augments PPTX with speaker notes, and prepares Q&A.
triggers	present paper, paper presentation, journal club, seminar presentation, grand rounds, academic presentation, presentation prep, lecture, lecture material, teaching slides, course slides, 강의자료, 발표자료, 슬라이드, pptx
tools	Read, Write, Edit, Bash, Grep, Glob
model	inherit

Present-Paper Skill

Purpose

Prepare a polished academic presentation from a research paper. The skill walks through a 5-phase pipeline: paper analysis, supporting research, script writing, slide note injection, and Q&A preparation.

Use it when:

• preparing a journal club or seminar presentation
• presenting a paper for a graduate course
• preparing grand rounds or conference talks based on a published paper
• building speaker notes for an existing slide deck

---

Communication Rules

• Communicate with the user in their preferred language.
• Use English for medical, statistical, and methodological terminology.
• Add pronunciation guides for drug names and technical abbreviations in the user's language.
• Be direct about paper limitations, but frame them constructively.

---

Phase 0: Init & Outline

Step 0a — Load design references (read before drafting outline)

Three of these are read now, in full — they change what you produce. The rest are read when the answer to Q0 tells you which one you need, because a talk has one venue and one style, and reading the others costs roughly seven thousand tokens to learn nothing you will use.

Read now (always):

A. references/ai_slide_tells.md — the marks a generated deck leaves. Read all of it, first. The complaint about AI decks is not that they are ugly — templates solved ugly. It is that they stop communicating, because they were built to make the maker comfortable rather than to serve the audience. This file is why the deck does not need catching later; scripts/check_slide_tells.py catches it after (Step 3.6). It overrules older guidance where they conflict — in particular the eyebrow-on-every-slide and brand-footer rules this project used to mandate, which are the single most-cited visual tell.

B. references/presentation_archetypes.md — the skeleton, chosen by where the speaker is standing: conference oral, journal-club critique, case-anchored grand rounds, didactic lecture, defence, keynote (Duarte's sparkline, the Jobs STAR moment, Takahashi/Lessig), lay talk, decision brief (Minto's pyramid, action titles, Kawasaki's 10/20/30). A deck has two independent choices and conflating them is why talks fail: the archetype is what the talk has to do; the visual style is what it looks like. A conference oral in a keynote's skeleton dies (no data on the slides); a keynote in a conference oral's skeleton dies harder. The skin is a preference; the skeleton is not. Its mechanical half is scripts/check_deck_budget.py.

C. references/presentation_design_guidelines.md — the enforceable rules (assertion headlines, 24-pt floor, negative space, ≤3 colours, colourblind-safe palettes, redraw-don't-screenshot, animation discipline) plus the G1–G10 self-check the Phase 3.5 critic scores against.

Read on demand — after Q0/Q2 tell you which one:

File	Read it when	Cost if read blindly
references/medical_presentation_templates.md	the venue is one of the five medical ones — then read that section only	~3,700 tokens, of which you use a fifth
references/slide_visual_styles/CATALOG.md → one style file	Q2 has chosen a style	~2,300 tokens per style
references/slide_design_principles.md	you are stuck on why a slide is not landing — Reynolds / Duarte / Knaflic / Tufte, the theory under the rules in C	~2,600 tokens of theory you mostly already applied

These mirror the entry-point pattern used in make-figures/references/design_principles.md (Step 1 "Specify"). Both skills share the same Reynolds / Knaflic / Tufte foundations — slide-level (this skill) and figure-level (make-figures) are companions, not duplicates.

Required Inputs

Before starting, collect these from the user:

Input	Why
Paper	PDF path, DOI, or PMID
Presentation time	Determines depth and slide count
Target audience	Specialty mix, knowledge level — controls terminology depth
Context	Course name, conference, journal club format, prior session topics
Template / visual style	Institutional template (.pptx/.potx) to fill, or a visual style to generate in. Default: ask (Step 0b)
Extension section	Optional topic to include (e.g., AI directions, clinical implications). Default: none

Step 0b — Template & visual style selection

After collecting the inputs above and before drafting the outline, settle how the deck will look. Ask the user two questions (use AskUserQuestion; skip a question if the user already answered it in their request):

Q0 — "Where are you standing, and for how long?" (venue + minutes)

This decides the archetype — the skeleton — before any question about looks. Map the answer with the selector table in references/presentation_archetypes.md, and carry archetype + minutes forward: Step 3.6 checks the built deck against them. A 40-word slide is an ordinary academic slide and a catastrophic keynote slide; there is no universal answer to "how much text is too much", only an answer for this room.

If the user gives only a topic and no venue, ask. Do not guess: a deck built for no particular room comes out generic in exactly the way every reviewer can see.

Q1 — "Do you have an institutional or branded template to use?"

• Yes → the user supplies a .pptx/.potx. Switch to Mode C (Phase 3, "Fill an institutional template"): run scripts/inspect_pptx_template.py <file> to list its layouts/placeholders/theme, then fill by placeholder index, preserving the master and logo. See references/slide_visual_styles/institutional_brand.md. Do not also ask Q2 — the template's theme is the style.
• No / none → ask Q2.

Q2 — "Which visual style should I generate in?" Offer the CATALOG.md menu with a one-line preview each (make the recommended option first and label it):

Option	One-line preview
Nature / Lancet (recommended for medical academic talks)	White, navy + coral accent, hairline dividers, Inter/Pretendard — restrained editorial-academic
Clinical Blue	White/light-blue, navy-teal, calm and trustworthy, colorblind-safe — grand rounds / CME
Editorial Mono	High-contrast black-on-white, oversized type, one accent — single big-message keynote
Dark Modern	Deep-slate background, off-white text, electric accent — AI / method / tech talks
Other	Describe a palette/feel, or name a journal/brand to emulate

Record the choice; pass the matching style spec to Phase 3. If the user has no preference and the talk is a medical academic talk, default to Nature / Lancet (~/.claude/rules/academic-lecture-style.md). Style choice does not change the outline, script, or Q&A — only Phase 3 rendering.

Q3 — conference decks only: is slide 1 a submission requirement? Many societies require the title slide to carry the title, authors, affiliations and country exactly as entered in the abstract submission. Those fields are not yours to improve. A crowded title slide is a real temptation to shorten an affiliation to its institution, and doing so breaks the requirement while satisfying the density check — which is the one place in this skill where a gate and a rule point in opposite directions. The requirement wins. Read the fields off the submission portal, copy them, and record SLIDE_TOO_DENSE on slide 1 as consciously overruled with that reason.

Paper Analysis

Read the paper and produce a structured analysis:

## Paper Analysis

### Citation
[Full citation with DOI]

### Background
- What gap does this paper address?
- What was known vs. unknown before this study?

### Study Design
- Type: [RCT / cohort / case series / meta-analysis / etc.]
- Subjects: [n, inclusion/exclusion]
- Methods: [key methodological choices]
- Primary outcome: [what was measured]

### Key Results
1. [Finding 1 with effect size and CI/p-value]
2. [Finding 2]
3. [Finding 3]

### Patient/Case Summary Table
[If applicable — structured table of individual cases or subgroups]

### Limitations
1. [Limitation 1]
2. [Limitation 2]

### Significance
- Why does this matter?
- What changes because of this paper?

Slide Outline

Create a slide-by-slide outline with time allocation:

## Slide Outline ([N] slides, [M] minutes)

| # | Title | Time | Key Content |
|---|-------|------|-------------|
| 1 | Title slide | 0:30 | Paper citation, presenter |
| 2 | Context / Prior sessions | 1:00 | How this connects to prior knowledge |
| 3 | Background | 1:30 | The gap this paper fills |
| ... | ... | ... | ... |
| N | Take-home messages | 0:30 | 3-5 key points |

Gate: User approves outline before proceeding.

---

Phase 1: Supporting Research

Search Strategy

Find references that strengthen the presentation:

1. Follow-up studies — Has the main finding been replicated or extended?
2. Clinical trial data — Large-scale data that contextualizes the findings
3. Review articles — Authoritative summaries that frame the topic
4. Contradicting evidence — Important for balanced Q&A preparation

Efficiency rule: Limit supporting references to 5-8 total. Only search categories that the approved outline (Phase 0) actually requires. Skip categories not needed for the presentation type (e.g., skip clinical trials for a methods-focused paper).

Selection Criteria

Do NOT summarize every paper found. Extract only:

• Specific data points needed for slides (incidence rates, OR/HR, AUC values)
• Findings that directly support or challenge the main paper
• Context that helps the audience understand significance

Output

## Verified References

### Main Paper
1. [Citation] — PMID: XXXXX, DOI: XX.XXXX/XXXXX

### Supporting References
2. [Citation] — PMID: XXXXX
   → Used for: [specific data point or context]
3. [Citation] — PMID: XXXXX
   → Used for: [specific data point or context]

### Key Data for Slides
- [Statistic 1]: [value] — Source: [Ref #]
- [Statistic 2]: [value] — Source: [Ref #]

Every reference must have a verified DOI or PMID. Mark unverified references with [UNVERIFIED].

---

Phase 2: Script & Content

Speaker Script

Draft a complete speaker script with these requirements:

1. Language: User's preferred language for narration; English for technical terms
2. Audience adaptation: Adjust explanation depth based on Phase 0 audience profile
   • For mixed audiences: add one-line plain-language explanations for specialty-specific terms
   • Example: "FLAIR sequence — an MRI technique that suppresses fluid signal to highlight edema"
3. Pronunciation guide: Include native-language pronunciation for drug names, abbreviations
   • Example: "lecanemab (leh-KAN-eh-mab)" or local equivalent
4. Timing markers: Note approximate time per slide
5. Transition phrases: Connect each slide to the narrative arc

Structure

## Speaker Script

### Slide 1: Title (0:30)
"[Opening — introduce yourself and the paper]"

### Slide 2: Context (1:00)
"[Connect to prior knowledge or clinical relevance]"

...

### Slide N: Take-home Messages (0:30)
"[Summarize 3-5 key points. Thank audience. Invite questions.]"

Extension Section (Optional)

Only include if user requested in Phase 0. Examples:

• AI/computational research directions stemming from the paper
• Clinical practice implications
• Policy or guideline implications
• Connections to the user's own research

Gate: User reviews script before proceeding.

---

Phase 3: Slides & Notes

Three Modes

Mode A = generate a new deck in a chosen visual style. Mode B = add notes to an existing deck. Mode C = fill the user's institutional/branded template (chosen at Step 0b). Pick the mode from the Step 0b answer.

Mode A: Generate new slide deck

Generate a fully-editable PPTX from structured inline data using python-pptx. Two canonical template libraries:

• ${CLAUDE_SKILL_DIR}/references/generate_pptx_templates.py — generic T_lead / T_text / T_table / T_image_right / etc. templates with smoke-tested main(). Use for journal club, grand rounds, conference talk, and short paper talks.
• ${CLAUDE_SKILL_DIR}/templates/build_pptx_nature_lancet.py — Nature/Lancet visual style (white + navy + coral, Inter/Pretendard, 47-slide academic lecture proven). Use for academic lecture multi-paper survey (template #5). Functions: new_presentation, add_title_slide, add_toc_slide, add_section_divider, add_transition_slide, add_content_slide, add_glossary_slide, add_closing_slide, plus fix_app_xml() helper. Style spec: references/slide_visual_styles/nature_lancet.md.

For lecture decks pulling figures from PDFs (rather than from /make-figures output), use ${CLAUDE_SKILL_DIR}/scripts/extract_pdf_figures.py — pdftoppm + PIL crop with normalized (0–1) box coordinates. Supports both single-crop CLI and YAML batch config.

After raw extraction, run ${CLAUDE_SKILL_DIR}/scripts/trim_caption.py to auto-remove journal headers / figure captions / surrounding whitespace so that only the figure body remains — the Adobe-Acrobat-crop equivalent in automation. The script uses horizontal-projection segmentation plus text-band detection (height + density + gap + line-pattern signature) and preserves multi-panel figures intact:

python3 "${CLAUDE_SKILL_DIR}/scripts/trim_caption.py" \
  --in-dir  figures/extracted \
  --out-dir figures/cropped

Handles four common journal layouts: top running-head bar, bottom multi-line caption (sparse text), bottom caption fused with figure body (no clear gap, detected via narrow dark/light alternation), and multi-row tables with footnotes (footnote cut, table rows preserved). No tesseract / OCR dependency — Pillow + numpy only. Verified on 12-figure academic deck (80–95% height retention; captions, journal banners, and CellPress-style headers all removed). When the deck slot expects only the figure body (default for build_pptx_nature_lancet.py), point FIG_DIR at the cropped output dir.

Word-boundary aware markdown parser (mandatory for HLA-rich decks)

When the build script parses inline **bold** / *italic* markers in slide body or speaker notes, the italic rule must use word-boundary lookahead / lookbehind so asterisk-bearing scientific tokens (HLA alleles like DRB1*07:01, HLA-A*02:01, SNP IDs, footnote markers) are not eaten as italic delimiters:

import re
pattern = re.compile(
    r"(\*\*(?:(?!\*\*).)+?\*\*"                           # bold; inner single * allowed
    r"|(?<![A-Za-z0-9])\*[^*\n]+?\*(?![A-Za-z0-9]))"      # italic (word-boundary)
)

Two regex tricks together:

1. Italic with boundary: (?<![A-Za-z0-9]) and (?![A-Za-z0-9]) reject * adjacent to alphanumerics, so DRB1*07:01 is left intact.
2. Bold tolerates inner single *: (?:(?!\*\*).)+? allows **DRB1*04:02** (HLA allele inside bold) to match as a single bold span.

Without these, a naive \*[^*]+\* italic pattern silently corrupts every HLA allele in the deck. Add the regex to add_styled() (or equivalent) in every Nature/Lancet-style build script.

Pronunciation auto-augment for non-native presenters

For decks where the presenter is uncomfortable with English pronunciation of acronyms, author names, drug names, or gene symbols, append a per-slide [ Pronunciation ] section to the speaker notes (audience sees nothing — only Presenter View). Use ${CLAUDE_SKILL_DIR}/scripts/inject_pronunciation_notes.py:

python3 "${CLAUDE_SKILL_DIR}/scripts/inject_pronunciation_notes.py" \
  input.pptx output.pptx \
  --dict pron_dict.yaml \
  --header "[ 발음 ]"            # or any header you like

The script:

• Loads a YAML/JSON PRON_DICT (term → [reading, full_name]) supplied by the caller. The dict is domain-specific — assemble it for your audience (Korean readings, French readings, Spanish readings, etc.).
• Uses word-boundary regex (?<![A-Za-z0-9_]) … (?![A-Za-z0-9_]) so short acronyms (e.g. AE, OR) only match when standalone, never inside other words.
• Recognizes allele-style tokens via a separate regex (\b(?:HLA-)?[A-Z]{1,5}[0-9]?\*[0-9]{2}:[0-9]{2}\b by default) and synthesizes their reading from the base allele entry in the dict.
• Skips slides that already contain the header (idempotent — safe to re-run).

Realistic yield on a 47-slide academic deck: ~38 slides receive a section, ~300 total term entries, 5–10 per annotated slide. Transition and divider slides have empty notes and are auto-skipped.

Speaker notes statistics density

When the slide body already shows exact OR / 95% CI / p-value, the notes should NOT repeat the same numbers — the presenter ends up reading statistics aloud and the audience cannot keep up. Notes should be a narrative (key anchors + one-line "see the slide body for the exact numbers" reminder), not a numeric listing.

Quick measurement to spot dense slides during QC:

import re
text = slide.notes_slide.notes_text_frame.text.split(pron_header)[0]
n_char = len(text)
n_stat = len(re.findall(r"\b(?:OR|p|CI)\s*[=<>]?\s*\d|\d+\.\d+|\d+%|×10", text))
needs_compression = n_char > 1000 and n_stat >= 5

Rule of thumb: 700–1,000 chars + 0–2 stat tokens is fine (30–60-second narrative). >1,000 chars + ≥5 stat tokens → compress to narrative tone and point at the slide body. Exact numbers belong in the slide body and footnotes (SSOT), not the notes.

Numbers in the notes are numbers you will say out loud

Everything the rest of this toolkit enforces about figures in the slide body applies to the notes, because the notes are what the speaker reads at the microphone. Two failures, both observed:

• A quoted benchmark typed from memory. "95% of 152 models were rated high risk" — the source said 87%, of 148 of 171; a second note said "98% of 62 oncology models" where the source said 84% and 62 was the number of papers, not models. The papers were cited correctly in the reference list. Only the digits were remembered. Every external number in the notes is checked against the source, exactly like a number in the body.
• An injected value hand-typed during a rewrite. When a build script draws numbers from an artifact — f"{N['primary']}" — compressing the prose is where they get flattened into literals. Twelve of them went that way in one 14-minute-to-9-minute pass. The slide body was gated and the notes were not, so the deck would have said one number while the speaker said another. When you edit generated text, the count of injection expressions must not fall. Before and after:

import re
len(re.findall(r"\{N\[", src))   # must not decrease across a rewrite

Pull the hand-typed numbers out for checking with the same regex, inverted — anything numeric that is not inside an injection expression is a literal somebody typed.

Sharing-ready notes-stripped variant

After the presentation, when the deck is shared with the audience (e.g. a professor asking for the slides), the speaker notes typically contain presenter-only material — second-language narrative, pronunciation hints, self-referential reminders ("Prof. ○○ will likely ask about …"). Stripping notes is mandatory before circulation. Use ${CLAUDE_SKILL_DIR}/scripts/strip_notes_for_sharing.py:

python3 "${CLAUDE_SKILL_DIR}/scripts/strip_notes_for_sharing.py" \
  presenter_v9.pptx share/<topic>_<initials>.pptx

The script:

• Clears every slide's notes_text_frame (idempotent, slide body and figures untouched).
• Re-writes docProps/app.xml with the correct Slides= and Notes= counts so PowerPoint Mac does not show its repair dialog (see also the app.xml canonical fix in pptx-mac-compatibility.md §5).
• Verifies that zero notes characters remain.

Recommended 3-file sharing package (filename pattern <topic>_<initials>):

• <topic>_<initials>.pptx — notes-stripped variant for slide reuse
• <topic>_<initials>.pdf — same deck, PDF for environment-agnostic preview (LibreOffice --convert-to pdf automatically drops the cleared notes pages)
• <topic>_<initials>_references.zip — optional bundle of the reference PDFs; if it exceeds the email attachment limit, send a Google Drive link.

In the cover email, mention the PPTX is included specifically so the recipient can reuse individual slides if useful.

The companion documents leave with the deck — check them too

_qa_prep.md, _quick_review.md and any handout are drafted from the same working material as the notes, and they travel further. Three things to check before any of them goes out:

• Retired numbers. A gate on the deck is not a gate on its siblings. In one talk the slides said 26.3% and the last-minute review sheet still said 54.1% — the deck was right and the document the speaker would answer questions from was wrong, which is the worse way round. Sweep the sibling .md files for the superseded values whenever the deck's numbers move. Where an old figure is deliberately quoted, fence it (<!-- superseded-quotation -->) rather than exempting the file.
• Operator material in an audience document. A brief written to prepare you for a meeting carries the register it was written in — "the point they will push back on", "the fallback position", a minute-by-minute plan. Handed to the person it describes, it reads as a strategy for managing them. When one document has to serve both purposes, it is two documents: an internal one with the contingencies, and a neutral one with the findings and what you are asking them to confirm. This is the same rule as stripping the speaker notes, applied to the files beside them.
• Live-only devices, and what points at them. A redistributed deck is read at a desk, not attended: a break slide, a timer, "as we just saw in the exercise" mean nothing there, and another speaker's break is their own decision. Remove them — and then grep the whole deck, bodies and notes, for the sentences that referred to them, because those are left dangling by the deletion. While you are there, re-verify anything the material asserts about an upstream project (install commands, counts, versions) against that project's own source rather than against the copy you wrote months ago: an exact number with a date stays accurate longer than a vague one, which drifts silently and cannot be checked.

Architecture

inline structured data (lists/dicts in build_*_slides())
    ↓ template functions (T_lead / T_text / T_table / ...)
editable PPTX with native text frames (selectable, restyleable in PowerPoint)

Three rules that keep slides stable:

1. No markdown parsing. Every slide is a function call with explicit inline data.
2. No cur_top cumulative position tracking. Use the fixed coordinate zones below — cur_top accumulates rounding errors and breaks layout after ~10 slides.
3. No Marp. Marp renders to images; the deck becomes uneditable and reviewers cannot copy text or restyle.

Slide-type templates

Template	Use for	Required fields
T_lead	Title slide, section divider	title, subtitle?, extra?
T_text	Bullet body (most common)	title, body_lines[], subtitle?
T_table	Cohort tables, comparisons	title, headers[], rows[][], body_before?
T_image_right	Body + figure on right	title, body_lines[], img_path, img_pct? (PNG ≥300dpi or vector PDF — see Figure source formats below)
T_quote_slide	Verbatim citations, witness quotes	title, quotes[], body_after?, img_path?
T_two_col	Compare/contrast	title, left_lines[], right_lines[]
T_two_col_with_box	Compare + emphasis	as above + metaphor_col, metaphor_lines[]
T_highlight_slide	Single key result	title, highlight_lines[], body_before?
T_metaphor_body	Body + analogy footer	title, body_lines[], metaphor_lines[]
T_table_two_col	Take-aways + numeric table	title, left_lines[], headers[], rows[][]

Figure source formats (when consuming /make-figures output)

When the deck pulls figures from analysis/figures/ produced by /make-figures:

• Preferred for slides: PNG at ≥300 dpi. python-pptx add_picture() handles this directly. Set img_pct (template T_image_right) so the figure occupies ≥40 % of slide width on a 13.33 × 7.5-in widescreen layout.
• Vector source available: prefer PDF only if the slide will be projected at >1080p or printed as a handout — convert PDF → PNG at the target DPI (pdftoppm -r 300 input.pdf out_prefix) before insertion, because python-pptx PDF embedding is unreliable across PowerPoint versions.
• Forbidden: TIFF (Mac PowerPoint silently drops it — see Mac compatibility checklist below); JPEG for line art (compression artifacts on diagonal lines); raw SVG (PowerPoint Mac handles it inconsistently).
• Caption / legend: re-draft for spoken-narration context, not the journal legend verbatim. The journal legend assumes a reader; the slide caption assumes a listener with 5–10 seconds of attention.
• Put the claim in the slide, not in the raster. A figure carries data; a conclusion drawn into the PNG ("Value = triage & safety net", a headline percentage) stops tracking the talk the moment the bullet above it is edited. Nothing catches that: text search sees the slide and not the image, so the only thing that finds it is a person looking at the render. Numbers and conclusions live in the slide's own text where they can be read, grepped, and corrected.

Diagrams and plots are drawn as CODE, then inserted (not out of autoshapes)

Hard rule. This is the highest-yield rule in the skill, and it is the one thing practitioners report actually working when they hand slide-making to an agent:

"에이전틱하게 PPT 도구를 사용하거나 / 웹페이지 형식으로 구성하는 경우는 거의 100% 실패함. 그나마 성공률을 높일 방법은 다이어그램 / 플롯을 모두 잘 알려진 도구(matplotlib 등)를 활용해 '코드'로 그리도록 시킨 다음, 그 결과를 그대로 삽입하도록 지시하는 방법인 듯."

Content	Draw it with	Never
Any chart	matplotlib / R (/make-figures)	Hand-placed shapes pretending to be a chart
Flow, mechanism, pipeline, hierarchy	matplotlib, or Graphviz DOT when the graph is the point	python-pptx autoshapes
Study flow (STROBE/PRISMA)	/make-figures flow builders	Boxes drawn one at a time

Then insert the rendered PNG (≥300 dpi) with add_picture().

Check the rendered PNG before you insert it. Drawing in code buys a second coordinate system, and a stroke laid on the figure's boundary is half-cut by the render. On the slide that does not read as a crop; it reads as a box with a side missing:

python3 scripts/check_diagram_edges.py diagrams/ --json qc/diagram_edges.json

DIAGRAM_EDGE_CLIP reports ink within a few pixels of the image border, and leaves full-bleed images alone (a photograph has ink on all four edges by construction). Run it straight after savefig, where the fix is an inner margin plus bbox_inches="tight" and a pad — once the PNG exists the stroke is already half gone and cropping cannot bring it back. People find these one at a time: the first time this guard ran it immediately found a second clipped diagram beside the one a reader had noticed.

Why the ban. Building a diagram out of autoshapes produces both AI tells at once: a row of identical rounded rectangles (SHAPE_MONOTONY) joined by arrows nobody labelled (ARROW_NO_SEMANTICS). Graphviz makes the second one structurally hard to get wrong — a DOT edge must be written A -> B [label="seeds along"], so the language itself demands the arrow declare what it claims:

digraph mechanism {
  rankdir=LR; node [shape=box, fontname="Inter"];
  catheter -> tract   [label="seeds along"];
  tract    -> nodule  [label="grows into"];   // an arrow that says what it means
}

An arrow is a claim — causes, becomes, flows into, is compared with, predicts. Six claims, one glyph. Drawn unlabelled, every person in the room supplies a different verb, and one wrong arrow can derail an entire discussion. See references/ai_slide_tells.md §4–5.

The one exception: a single, deliberate, labelled shape used as an accent (a callout box, a highlight frame). One shape is a choice; eight identical ones are a generator.

Helpers (used by templates — usually you do not call directly)

Helper	Role
_text	Single text box with **bold** inline markup
_multiline	Multi-line block with bullet (- , ✓ ) and ### subhead support
_title_block	Title + teal underline + optional subtitle
_table	Styled table (teal header row, alternating rows)
_quote	Blockquote — teal left bar + light-blue background
_highlight	Yellow rounded box + orange 2pt border
_metaphor	Same shape as quote, lighter font
_image	PIL aspect-preserving image insert (handles iPhone EXIF if you transpose first)
_slidenum	Bottom-right page number

Design tokens (defaults — change to fit institution/journal)

NAVY    = #1B2A4A   # title text, section divider background
TEAL    = #0072B2   # subtitle, underline, table header bg, quote bar
ORANGE  = #D55E00   # highlight box border
GRAY    = #333333   # body text
FONT    = 'Arial'   # present on both platforms; see the font-portability check below

The font is a delivery decision, not a taste decision

A typeface that is not installed on the machine the deck opens on is substituted silently: the words stay, the metrics change, line breaks move, and a box that fitted stops fitting. It is invisible on the authoring machine by construction — you have the font — and it surfaces on the projector.

python3 scripts/check_font_portability.py output/presentation.pptx --json qc/font_portability.json

FONT_NOT_PORTABLE names any typeface bundled with one operating system and absent on the other, with a count per font so a 1,000-run body face reads differently from a stray monospace in three code lines. It is a blocklist, not an allowlist: a hospital's licensed brand face is not this check's business. It exempts fonts the deck embeds, and it treats a theme-level default as inert until the deck actually contains text of the script that slot serves.

Two ways to be safe, and both have a cost worth knowing:

• Embed the fonts (PowerPoint: Save > Embed fonts in the file). Licence permitting.
• Carry a PDF. The portable fallback — but PDF drops embedded video, so a deck with a clip must ship its MP4s separately or the fallback is not one.

Fixed coordinate zones (16:9 = 13.333" × 7.5")

ML / MR = 0.8"     MT = 0.5"     CW = SW − ML − MR = 11.733"

TITLE_Y = 0.5"    TITLE_H = 0.8"
SUB_Y   = 1.3"    SUB_H   = 0.5"
BODY_Y  ≈ 1.9"    BODY_H  ≈ 5.1"

Build script responsibilities

A from-scratch generation script must:

• Reference every input by a path relative to the presentation directory, and keep every input that produced an artifact inside it. A build script written during a session tends to point at wherever the work happened to be — a session scratch directory, a temp path. That directory is gone next week, and with it the ability to rebuild: the figure PNGs survive, the scripts that drew them do not, and a label baked into a figure can no longer be corrected to match a body line that has since changed. Save the figure-generation scripts next to the deck, not next to the session.
• Assign all four placeholder coordinates together (see Step 3.7) — a partial assignment writes a shape with no area and hides it from every check.
• Convert TIFF images to PNG before add_picture (Mac PowerPoint silently drops TIFF).
• Apply EXIF transpose to iPhone photos before insertion.
• After inserting/removing slides, sync docProps/app.xml (<Slides>, <Notes>, HeadingPairs, TitlesOfParts) to the actual count, or PowerPoint Mac will raise a recovery dialog on open.
• If you copy <a:srcRect> from another deck, copy the values verbatim — they are 1/1000-percent (cap 100000), never EMU. A unit conversion bug here crops 99% of the image off-slide.
• Print slide count, notes count, file size, and editability check at the end.

Forbidden in Mode A

• ❌ Marp CLI for PPTX (always image-rendered, uneditable).
• ❌ Markdown auto-parsing into slides (layout drifts on every regeneration).
• ❌ cur_top cumulative top tracking (accumulates rounding error).
• ❌ Direct iPhone photo insert without EXIF transpose (rotated 90° in PowerPoint).
• ❌ Using python-pptx from-scratch rebuild to edit an existing deck — see Patch over Rebuild below.

Mac PowerPoint compatibility checklist

PowerPoint Mac is stricter than Windows / Keynote / LibreOffice on OOXML defects. Verify before delivering any deck destined for a Mac viewer:

Defect	Detect	Fix
TIFF images	find ppt/media -iname '*.tif*'	sips -s format png in.tif --out out.png + replace .tif→.png in _rels/*.rels
<a:sp3d> in rPr	grep -l '<a:sp3d>' ppt/slides/*.xml	Regex-strip the <a:sp3d>...</a:sp3d> block (renders as red outline only on Mac)
app.xml count mismatch	<Slides> value + HeadingPairs count + TitlesOfParts size vs actual slide files	Sync all four fields to real count
srcRect corruption	Any value > 100000 (1/1000-percent cap)	Compare with original deck; restore verbatim

Validation must run on PDF export AND Mac PowerPoint — neither alone catches all four. PDF misses sp3d outlines and srcRect corruption.

Patch over Rebuild — editing an existing PPTX

When the user supplies an existing deck and asks for surgical edits (textbox width, image crop, font swap, sp3d removal), prefer regex/sed patching of the unzipped XML over regenerating with python-pptx. From-scratch rebuild loses:

• <a:srcRect> image crops
• <a:sp3d> / <a:scene3d> (when intentional)
• Slide master / layout / theme details
• app.xml and core.xml metadata

unzip -q original.pptx -d /tmp/work
python3 -c "
import re; from pathlib import Path
p = Path('/tmp/work/ppt/slides/slide23.xml')
s = p.read_text()
s = s.replace('cx=\"9504720\"', 'cx=\"11200000\"')
p.write_text(s)
"
cd /tmp/work && zip -rq ../patched.pptx . -x '*.DS_Store'

python-pptx is reserved for (a) brand-new decks built via the templates above, or (b) appending speaker notes via slide.notes_slide.notes_text_frame.text. The skill's scripts/inject_speaker_notes.py is the canonical example of (b). It parses inline **bold** / *italic* into run-level styling by default (python-pptx stores text verbatim, so the markers would otherwise show literally in Presenter View — the failure mode pptx-speaker-notes.md warns against); pass --no-markdown for legacy plain text. A reproducible check lives at tests/test_speaker_notes_markdown.py.

Standard structure (10–15 min paper talk)

1. Title slide (T_lead) — paper citation + presenter
2. Background (T_text × 1–2)
3. Study design / Methods (T_text or T_two_col)
4. Key results with figures (T_image_right / T_table × 2–3)
5. Discussion (T_text)
6. Limitations (T_two_col_with_box works well)
7. Take-home (T_text or T_highlight_slide)

Output

Save to output/presentation.pptx. Speaker notes go into the notes pane only — never modify slide design when adding notes.

Step 3.5 — Slide critic (run before delivering deck)

After exporting the PPTX, run the slide critic rubric at references/critic_rubrics/slide.md. Score each slide and the deck-level Mac compatibility checks (Section F) as PASS / PARTIAL / FAIL. Produce concrete edits for every FAIL or PARTIAL item before treating the deck as ready.

Mandatory deck-level checks (cross-link with ~/.claude/rules/pptx-mac-compatibility.md):

# F.22 No TIFF
find ppt/media -iname '*.tif*' || true   # must be empty

# F.23 No 3-D bevel
grep -l '<a:sp3d>' ppt/slides/*.xml      # must be empty

# F.24 app.xml count sync
grep -c '<Slides>\|<Notes>' docProps/app.xml
ls ppt/slides/slide*.xml | wc -l         # must match

# F.25 srcRect bounds (any value > 100000 = bug)
grep -oE '"[0-9]{6,}"' ppt/slides/*.xml | head

Record critic_pass: yes | partial | no and refine_rounds: N in _quick_review.md.

Step 3.6 — AI-tell audit (deterministic; run on the built deck, not the build script)

python3 scripts/check_slide_tells.py  output/presentation.pptx --json qc/slide_tells.json
python3 scripts/check_deck_budget.py  output/presentation.pptx --json qc/deck_budget.json \
        --archetype <from Q0> --minutes <from Q0>

Write the --json under the project's own qc/ directory, not only to the terminal. A verdict that exists solely in scrollback cannot be counted later: how often a check fires, and on what, is the only evidence that it is worth keeping, and a check nobody can measure is one nobody can retire either. qc/ is where the rest of this toolkit leaves its envelopes, and each one names the detector that wrote it.

check_deck_budget.py is the mechanical half of the archetype: slides against the clock (DECK_OVER_BUDGET), words per slide against what this room can absorb while also listening (SLIDE_TOO_DENSE), and the type floor for the back row (TYPE_TOO_SMALL). It takes an archetype rather than a universal threshold because a single global number would have to be wrong for most venues. --list prints the budgets.

It also reports ZERO_AREA_TEXT, which is not about the room at all — it is about whether the other three verdicts mean anything. See "Placeholders inherit geometry" below.

Six verdicts, each one a mark reviewers say they can spot instantly. Every one must be cleared or consciously overruled, with the reason written down:

Verdict	What it found	The fix
CHROME_ON_EVERY_SLIDE	Eyebrow labels / brand footers on ≥60% of slides	Keep the page number and the dividers. Delete the rest.
SCAFFOLD_PHRASE	A slide (or note) narrating its own construction — "요약하자면", "The key takeaway is…"	Delete the sentence; say the thing it was pointing at.
TOPIC_TITLE	A content slide titled "Results" instead of stating the result	Assertion headline: "Adjunctive ablation halved local recurrence (12% vs 26%)."
SHAPE_MONOTONY	The same box, eight times, at the same size	Parallel ideas → one table. Non-parallel ideas → different shapes.
DEAD_SPACE_BAND	A mostly-empty slide with a hole through the middle	Say more, or say one thing large.
ARROW_NO_SEMANTICS	≥2 arrows, none labelled	Label every arrow, or add a legend. An arrow is a claim.

The detector is stdlib-only and reads any .pptx, so it also works on a deck a colleague sends you, or one you did not build here.

It is not a style opinion, and it does not detect "was AI used". Used as a booster, AI leaves none of these marks. Used as a button, it leaves all of them.

Step 3.7 — Placeholders inherit geometry: assign all four coordinates, or none

A layout placeholder gets its position and its size from the layout. Assign one of them —

title.width = Inches(11.5)          # ← and nothing else

— and python-pptx materialises an <a:xfrm> for the whole shape, writes the value you gave it, and records the rest as zero or absent rather than as inherited. The box renders with no height, or no width. The text is in the file and is not on the slide.

shape.left, shape.top, shape.width, shape.height = (       # all four, together
    Inches(0.8), Inches(0.6), Inches(11.5), Inches(1.0))

The reason this has its own step, rather than a line in a checklist, is what it does to the checks above it. A shape in that state has no <a:off>, the deck reader drops it, and its words and its type size leave the deck before check_deck_budget counts anything. A title slide carrying 69 words at 11.5 pt passed the budget check for exactly this reason; correcting the coordinates surfaced three findings that had been there all along. The green was made out of the defect. ZERO_AREA_TEXT now reports it, and reports it first, because every verdict after it was computed without the text that is missing.

If you see ZERO_AREA_TEXT, fix the geometry and run the check again. Treat the first run's silence on everything else as unread, not as passed.

Step 3.8 — Does it fit? Measure the render; never estimate it

python-pptx will write more text than a box can show and say nothing about it. PowerPoint reveals it on the screen, which is where the audience is.

The tempting check is arithmetic — font size × line spacing × lines — and it fails in both directions. A line-height constant of 1.42 under-estimated CJK line pitch and let a body block cross into the footer; measuring the render gave 1.60; raising the constant to 1.62 then refused about 290 passages that rendered perfectly well. Separately, sizing a block at font × 1.06 while forgetting the roughly 1.2 leading PowerPoint adds on top made a 21-line list compute to 4.1 in when it needed 5.1. Those two numbers are why there is no estimator in this skill: they are what the estimate is wrong by, in each direction.

The render already knows. You are exporting a PDF anyway — it is half the Mac-compatibility check and the portable fallback for a venue without your fonts:

soffice --headless --convert-to pdf output/presentation.pptx
python3 scripts/check_text_overflow.py output/presentation.pptx --pdf output/presentation.pdf \
        --json qc/text_overflow.json

OFF_SLIDE is a line ending in the reserved band at the foot of the slide; CARD is a line whose bottom passes the bottom of the filled block it sits in. Both report the measured distance, because "0.03 in below the block" and "0.6 in below the block" call for different repairs.

Without a render it exits 2 — could not measure — rather than reporting a pass. A check that answers when it did not look is worse than no check, because the answer gets quoted.

Mode B: Add notes to existing slides (more common)

Before touching the deck, run the two Step 3.6 checks on the deck you were given and report what they say. A request phrased as "just fix the style" is not a diagnosis, and taking it as one is how an afternoon goes into underlines while the actual defect — 120 words a slide, the keyword buried in the fourth line, slides ordered the way the thought arrived rather than the way it lands — survives untouched. SLIDE_TOO_DENSE, TYPE_TOO_SMALL, SCAFFOLD_PHRASE and TOPIC_TITLE answer "is this a content problem or a design problem?" in about ten seconds. Put the answer in front of the user and let them choose what you work on. Often the design was fine.

Then:

• Read existing PPTX to understand slide structure and count
• Map speaker script sections to corresponding slides
• Generate inject_notes.py script tailored to the specific presentation

Note Injection Script

Generate a tailored inject_notes.py following the pattern in ${CLAUDE_SKILL_DIR}/scripts/inject_speaker_notes.py. The generated script should contain only the notes dictionary customized for this presentation and the main injection loop from the template.

Critical Rule

Speaker notes are injected without modifying slide design, layout, text, or images. The script only touches the notes pane. Verify by comparing slide content before and after.

Mode C: Fill an institutional / branded template

When the user supplied a .pptx/.potx at Step 0b (university, hospital, society template with a fixed logo and theme), fill it — do not redesign it. This is patch-over-rebuild (~/.claude/rules/pptx-mac-compatibility.md §2): a from-scratch Presentation() would drop the institution's master, theme, and logo.

1. Inspect: python3 ${CLAUDE_SKILL_DIR}/scripts/inspect_pptx_template.py <template> → lists every layout (index, name) with its placeholders (idx, type, size) plus theme fonts/colors. Read it before writing content. The template has already drawn things: a society layout often carries a full-bleed background image whose colour band is the header, and a master that already has a title frame. Deleting the placeholders and drawing your own boxes puts your title half on the band and half off it. Find the existing design region — the placeholder rectangles from the inspector, and if the header is painted into a background image, where its colour changes — and put text inside it rather than over it.
2. Map each outline slide to one of the template's existing layouts (Title / Title+Content / Section Header / Closing). Do not invent layouts.
3. Fill by placeholder_format.idx (from the inspector) so the institution's fonts, sizes, and logo are inherited — never add free text boxes for title/body. Code pattern and the no-usable-body-layout fallback are in references/slide_visual_styles/institutional_brand.md.
4. Notes: inject with scripts/inject_speaker_notes.py as usual (notes are template- independent).
5. Verify: open in Mac PowerPoint (no repair dialog, logo on every slide, fonts intact); confirm the logo media is still embedded; sync docProps/app.xml after adding/deleting slides (pptx-mac-compatibility.md §5–5.1).

The content rules (presentation_design_guidelines.md) still apply inside the brand — one idea per slide, redrawn tables, ≤3 colors within the institution's palette.

---

Phase 4: Q&A Preparation

Question Generation

Generate questions from multiple perspectives:

1. Methodology critics: "Why this design? Why not...?"
2. Domain experts: Deep technical questions about the specific field
3. Generalists: "What does this mean for clinical practice?"
4. Students/trainees: Clarification questions about unfamiliar concepts

Answer Structure

Every answer should follow the pattern:

Acknowledge → Evidence → Conclude

"That's an important limitation. [Acknowledge the concern honestly.]
However, [cite specific supporting evidence — author, year, finding].
So while [restate limitation], [conclude with the paper's contribution despite it]."

Quick Review Sheet

A single-page reference for last-minute review:

## Quick Review

### Must-Know Numbers
| Metric | Value | Source |
|--------|-------|--------|
| [Key stat 1] | [value] | [Ref] |
| [Key stat 2] | [value] | [Ref] |

### Common Pitfalls
- Don't confuse [X] with [Y]
- [Classification A] and [Classification B] are independent frameworks
- Slide says [rounded value], precise value is [exact value]

### Key Takeaways (memorize these)
1. [Point 1]
2. [Point 2]
3. [Point 3]

---

Output File Structure

All outputs go in the user's presentation directory:

{presentation_dir}/
├── _analysis.md              # Phase 0: Paper analysis + outline
├── _references.md            # Phase 1: Verified references + key data
├── _script.md                # Phase 2: Speaker script
├── _qa_prep.md               # Phase 4: Expected Q&A
├── _quick_review.md          # Phase 4: Pre-presentation review sheet + critic_pass record
├── _slide_critic.md          # Phase 3.5: Slide rubric scores per slide
├── inject_notes.py           # Phase 3: Tailored note injection script
├── figures/                  # Extracted paper figures (if needed)
└── reference/                # Supporting paper PDFs (if downloaded)

Cross-skill / Cross-rule integration

This skill composes with adjacent skills and global rules:

When	Use	Why
Need a figure on a slide (ROC, forest, KM, flow)	/make-figures first, then embed	Both skills share Reynolds/Knaflic/Tufte foundations; figure-level + slide-level companions
Manuscript reporting checklist parallel	/check-reporting for the same paper	Paper presentations often shadow manuscript revision; reporting-guideline gaps surface in Q&A
Visual abstract / Central Illustration	/make-figures visual-abstract templates	Then verify against ~/.claude/rules/journal-ai-image-policies.md (JACC prohibits, Radiology allows with disclosure)
PPTX edits to existing institutional template	~/.claude/rules/pptx-mac-compatibility.md	Patch over rebuild; preserve master/layout/srcRect
Manuscript companion deck	~/.claude/rules/manuscript-style-classical.md	Heading style, AI-Disclosure policy, em-dash discipline carry over to slides for senior MA reviewer audiences
References on slides	/verify-refs (audit-only) before delivery	Same anti-hallucination gate as manuscript references

---

Constraints

• Never fabricate references. Every citation must be verified against PubMed, DOI, or the PDF itself.
• Never modify slide design when injecting notes. Notes and slides are separate concerns.
• Always ask audience first. Do not start drafting until the target audience is defined.
• Extension sections are opt-in. Do not add AI/clinical/policy sections unless explicitly requested.
• Respect presentation time. Script length must match allocated time (roughly 130-150 words per minute for academic presentations).

Anti-Hallucination

• Never fabricate references. All citations must be verified via /search-lit with confirmed DOI or PMID. Mark unverified references as [UNVERIFIED - NEEDS MANUAL CHECK].
• Never invent clinical definitions, diagnostic criteria, or guideline recommendations. If uncertain, flag with [VERIFY] and ask the user.
• Never fabricate numerical results — compliance percentages, scores, effect sizes, or sample sizes must come from actual data or analysis output.
• If a reporting guideline item, journal policy, or clinical standard is uncertain, state the uncertainty rather than guessing.

Global-rule references

Some passages in this skill cite a path of the form ~/.claude/rules/<name>.md. Those are the maintainer's personal global rules, kept outside this repository. They are not shipped with this skill and will not exist on your machine; they appear only as provenance for where a convention came from. If one of them looks like it is standing in for an instruction you actually need, that is a bug — please open an issue, because the instruction belongs here.