premium-deck-builder

name	premium-deck-builder
description	Construct PowerPoint, Keynote, Reveal.js, or Marp decks that are both legible from the back of the room AND feel premium / crafted / expensive. Use for pitch decks, interview decks, board decks, conference talks, slidedocs, investor memos, lecture slides. NOT for logo / wordmark design (→ brand-guidelines), copywriting (→ copywriting), motion graphics for video (→ motion-design-web), live-stage stagecraft (out of scope), or live-presentation coaching (→ tech-presentation-interview).
license	MIT
allowed-tools	Read,Write,Edit,Glob,Grep,Bash
metadata	{"category":"Design & Communication","tags":["presentations","powerpoint","keynote","reveal-js","typography","design-systems"],"runtime":{"canonical-skill-path":"premium-deck-builder/SKILL.md"},"provenance":{"kind":"first-party","owners":["erichowens"]}}
io-contract	{"kind":"deliverable","produces":[{"kind":"code","description":"Presentation deck source (PowerPoint/Keynote/Reveal.js/Marp) legible from the back of the room"},{"kind":"design-doc","description":"Narrative arc and visual-system spec for the deck"}]}

Premium Deck Builder

Build slide decks that read from the back of the room and feel like the maker chose every margin on purpose. Encodes typography, composition, and aesthetic discipline drawn from Duarte, Reynolds, Tufte, Alley, Knaflic, and the visual conventions of high-end consulting and tech-pitch decks.

When to load

• User says: "deck", "slides", "presentation", "pitch", "PowerPoint", "Keynote", "Reveal", "Marp", "slidedoc", "investor deck", "interview deck", "board deck".
• File evidence: .pptx, .key, .md with marp: frontmatter, index.html importing reveal.js, python-pptx in requirements.txt.
• Defect signals: "the deck feels cheap", "looks corporate", "title-and-bullets default", "can't read it on the projector", "feels busy", "client said unprofessional".
• Adjacent skill needs depth on slide design specifically (typography-expert is web-typography; this skill is its slide cousin).

Do not load for: building the website about the deck, choosing the brand, writing the underlying narrative, video production with kinetic typography, or coaching the human delivering it.

Core decision: pick the deck MODE before any pixel

flowchart TD
  A[Deck request] --> B{Where will it live?}
  B -->|Projected live, 6-20 ft audience| C[LIVE MODE<br/>Reynolds rules]
  B -->|Screen-shared on Zoom| D[SCREEN-SHARE MODE<br/>Reynolds + bumped sizes]
  B -->|Sent as PDF / read async| E[SLIDEDOC MODE<br/>Duarte rules]
  B -->|Pitched to investors / interview / sales| F[PITCH MODE<br/>Sequoia / Airbnb pattern]
  B -->|Internal report to executives| G[CONSULTING MODE<br/>McKinsey / Minto pattern]
  C --> H{Mode locked}
  D --> H
  E --> H
  F --> H
  G --> H
  H --> I[Apply mode-specific tokens<br/>see references/composition.md]

The cardinal sin is mixing modes in one deck: slidedoc density projected live, or one-idea-per-slide pitched as a takeaway memo. Decide first, design second.

Operating rules

1. Mode first, content second, slides third. Don't open a template and start typing.
2. One assertion per slide in live mode. Replace "topic-phrase title + bullets" with Alley's sentence-headline-as-title + visual evidence. (Alley/Garner/Zappe, IEEE Trans. Prof. Comm. 2009 — assertion-evidence beat topic-bullets on comprehension AND retention.)
3. Three colors maximum: one neutral, one ink, one accent. Chart series get muted tints of the accent — never new hues.
4. Margins ≥ 8–10% of slide width. White space is taste. Filling the slide is the loudest tell of cheapness.
5. Sentence case for titles longer than 3 words. ALL CAPS is 13–20% slower to read (Tinker 1963; Arditi & Cho 2007). Reserve all-caps for short eyebrows (≤2 words, weight ≥600, letter-spacing ≥0.1em).
6. Tabular figures + slashed zero on every numeric surface. Prevents column jitter and 0/O confusion at glance distance.
7. No "Thank you / Questions?" closer. Highest-recall slot — use it for the single takeaway or the CTA.
8. Live mode body ≥ 24pt. Screen-share ≥ 28pt (the viewer sees you at 25–50% scale). Slidedoc body 10–12pt is the only context where sub-18pt is defensible — and tag the file so nobody projects it.
9. Don't ship with embedded foundry fonts unless you control the playback machine. Either embed-and-test, or export to PDF, or use a system-safe stack (Calibri / Aptos / SF Pro / Helvetica Neue).
10. Verify by rendering. Always soffice --headless --convert-to pdf + pdftoppm (see scripts/render_deck.sh) before declaring done. Reading the script is not the same as seeing the slide.

The five modes (one-line summaries; see references/composition.md for details)

Mode	Pattern	Body floor	Words/slide	Backgrounds	When
LIVE	Reynolds: one idea, image-led, speaker carries message	24–32pt	15–20	warm off-white or near-black	conference, internal talk
SCREEN-SHARE	LIVE + bumped sizes for downsampled viewing	28–36pt	15–20	warm off-white	Zoom / Teams / Meet
SLIDEDOC	Duarte: dense hybrid of slide + document, async read	10–12pt	75–150	white or off-white	board memo, leave-behind
PITCH	Airbnb / Sequoia: 10–12 slides, sparkline narrative	20–28pt	20–40	off-white or near-black	investor, interview, sales
CONSULTING	McKinsey: action-title + executive box + supporting exhibits, Minto's pyramid	14–18pt	60–100	white	client deliverable, internal exec

Failure modes (high-frequency)

FM-1. Web-grid translated to opaque PPTX → reads as "black on B&W texture"

Symptom: maker copied a portfolio's pale-blue alpha grid (rgba(143,166,199,0.28)) into the slide background and the rendered deck looks like text fighting a wireframe. Root cause: PowerPoint shapes don't honor CSS alpha. A 28%-alpha blue line on warm paper composites to a soft pale blue on the web; the same hex baked into a RGBColor() PPTX line is a saturated solid stroke. Fix:

1. Pre-composite the alpha against the actual paper color. rgba(143,166,199,0.28) over #FAF7F0 ≈ #EDEEEC, not #C7D3E5.
2. Sparser default density: every 0.5″ fine + every 2″ bold for content slides (vs. every 0.25″ + every 1.25″ on the web).
3. Thinner strokes: 0.125–0.25pt, not 0.5pt. Slide-resolution lines are bigger than CSS lines at the same px.
4. Reserve the visible grid for COVER + SECTION-DIVIDER slides where the grid IS the visual. Content slides get a near-invisible grid or none.

See the worked example for the specific RGBColor / line-width values that ship.

FM-2. Body sits behind the grid (or any background texture) and reads as illegible

Symptom: text correctly sized at 24pt+ still reads as "busy" or "hard to skim." Root cause: grid / noise / illustration runs through the glyphs without a contrast break. Fix: either (a) reduce the background to near-paper as in FM-1, (b) add a paper-colored "card" backing behind the text block (a rectangle in the slide background color, slightly larger than the bounding box), or (c) remove the background texture from content slides entirely.

FM-3. The font you designed in renders as Aptos on the client's machine

Symptom: deck looks beautiful on your Mac, looks like a default PowerPoint on the audience's PC. Root cause: PowerPoint font embedding has been unreliable across PPT/Keynote/Google Slides round-trips for fifteen years. Foundry fonts are silently substituted. Fix: either (a) export to PDF for delivery and keep the .pptx for the maker only, (b) embed AND test on the actual playback OS, or (c) use a system-safe stack (Aptos / Calibri / SF Pro / Helvetica Neue / Georgia / Menlo).

FM-4. Render verification skipped → "the script said it worked"

Symptom: agent declares the deck done after prs.save(...) returns OK. Root cause: opening the .pptx in Python and reading the shape count is not the same as seeing the slide. Render-and-look is mandatory. Fix: ALWAYS run scripts/render_deck.sh <deck.pptx> (which wraps soffice --headless --convert-to pdf + pdftoppm) and actually look at the PNG output before reporting done.

FM-5. Fabricated content dressed as the candidate's voice

Symptom: deck contains specific technical claims, quoted aphorisms, or characterizations of past work that the agent has no source for. Reads as "AI garbage" to anyone who knows the actual material. Root cause: agent over-reached past the verifiable source (resume / constants file / portfolio copy) and generated plausible-sounding detail. Worst form: text inside actual quotation marks — implying an attributable quote when nobody said it. Fix:

1. Identify the single source-of-truth file (e.g. a project's constants.ts , the candidate's resume, the brief). Treat anything not in there as fabricated.
2. Strip every quoted aphorism. If it's not attributed to a real person with a date, it doesn't belong in quote marks.
3. Replace fabricated specifics with bracketed prompts: [ERICH: what was the actual cold-start handling approach?]. The candidate fills these in.
4. Reduce each case-study slide to the verifiable facts + explicit [FILL] slots.
5. Margin notes / handwritten annotations should NOT use quotation marks — they're editorial marginalia, not quotes. Reserve " " for actual attributed quotes from real sources.

This is the worst failure mode in the skill because it directly damages the candidate or speaker in the room. A reviewer who knows the topic detects fabrication immediately and writes off the entire deck. The agent's job is to build the chrome and the structure — the substance belongs to the human.

FM-6. Mode mid-deck swap (live + slidedoc in one file)

Symptom: first half of the deck is one-idea-per-slide; second half drowns in dense bullets because "the audience will want the details in writing." Root cause: trying to make one deck do two jobs. Fix: build TWO decks. A live-mode deck for the room; a slidedoc for the leave-behind. Or commit to one mode and stay there.

Anti-patterns (named so a reviewer can grep)

AP-1. Title-then-bullets default

Novice: opens template, types "Q3 Results", adds 6 bullets. Expert: chooses slide TYPE first (assertion-evidence, comparison, data-led, image-led, quote, transitional) and writes the action title — "Q3 revenue grew 18%, driven by enterprise renewals." Timeline: bullet templates were industry default 1990–2006; Tufte's Cognitive Style of PowerPoint pamphlet (2006, Columbia disaster) and Alley's assertion-evidence research (2009) ended their defensibility. Still the default in 2026 because templates haven't caught up.

AP-2. Twelve-point body for projection

Novice: copies print typography to projector (Word habit). Expert: 24pt floor for projection (Kawasaki 10/20/30 rule, Reynolds), 28pt for screen-share. The only sub-18pt context is a slidedoc designed to be READ at 18 inches. Timeline: print-on-screen was always wrong; became visibly wrong with high-DPI projectors in 2015+ when fine details started rendering but the audience still couldn't read them.

AP-3. The "Thank you / Questions?" closer

Novice: ends with social-courtesy slide because the template suggests it. Expert: ends on the single takeaway or call to action. The final slide gets the highest visual recall — wasting it is malpractice. Timeline: convention since corporate-overhead days (1980s); challenged by Reynolds' Presentation Zen (2008) and now actively wrong in any deck that wants to be remembered.

AP-4. Rainbow chart palette

Novice: every chart series gets its own hue from the 12-color default. Expert: one accent color + muted tints (Cole Knaflic, Storytelling with Data, 2015). If five series can't read with tint variation alone, the chart has too many series. Timeline: rainbow palettes became actively bad practice when high-DPI displays exposed how garish they look at scale (~2015 onward).

AP-5. Embedded designer font with no fallback

Novice: picks Gotham/Proxima/Founders Grotesk in Keynote and emails the .pptx to the client. Expert: either embeds AND tests on the client's actual Windows PowerPoint, OR exports to PDF, OR uses a system-safe stack (Aptos / Calibri / SF Pro / Helvetica Neue / Georgia). Timeline: PowerPoint font embedding has been unreliable through every PPT/Keynote/Google Slides round-trip since 2010; still trips people up because foundry fonts look so good on the maker's machine.

AP-6. Default PowerPoint/Keynote template

Novice: opens template, accepts Aptos title + bullets + corporate beige. Expert: starts from a blank slide with an explicit grid + token system (paper color, ink color, accent color, type scale, margin scale all defined before any content). Timeline: Aptos replaced Calibri as PowerPoint default in 2024 — same template-default failure with a new face.

AP-7. Fabricated content / AI-generated aphorisms

Novice: writes plausible-sounding bullets, drops in quoted "leadership wisdom" with no attribution because "it sounds good." Expert: only writes what's verifiable from the source-of-truth file. Marks every unknown with [FILL: …]. Refuses to put text inside quotation marks unless it's an attributed quote from a real source. Timeline: the failure mode is as old as deck-making but became named and easy to detect with the rise of LLM-generated content. Reviewers in 2024+ are specifically pattern-matching for "AI cliché": punchy aphorisms, the rule-of-three closer, the "X is really about Y" reveal. Don't ship these.

AP-8. Mode-mixing within a deck

Novice: makes one deck that's "both for the meeting AND to send around after." Expert: builds TWO decks — a live-mode deck for the room and a slidedoc for the leave-behind. Or one deck that knows which it is. Timeline: Duarte named this in Slidedocs (2014) — the mode collision predates that but had no name.

Slide types (use deliberately; never default)

• Title — deck title + presenter + date + audience. Image-led OK.
• Section divider — large numeral or word. Palate-cleanser between acts.
• Content — assertion-evidence by default. Bullets only when content is genuinely enumerable and parallel.
• Comparison (split) — 2 or 3 columns, equal weight, parallel structure top-to-bottom.
• Data / chart — one chart, one headline conclusion, redundant encodings stripped, source line.
• Quote / pullquote — large type, attribution, single idea.
• Image-led — full-bleed, headline overlaid on a calm region with sufficient contrast.
• Transitional — bridges between sections; restates where we are in the argument.
• Agenda / tracker — show once up front; optionally a persistent footer marker.
• Closer / Q&A — single takeaway or CTA. Never "Thank you!" alone.

Quality gates (run before declaring done)

• Mode declared in a comment at the top of the deck source.
• Render verified via scripts/render_deck.sh <deck.pptx> — actually look at the PNGs.
• Body size ≥ 24pt for any slide intended for projection; ≥ 28pt for screen-share. Run scripts/check_deck_quality.py to count violations.
• ≤ 3 colors across the deck (one neutral + one ink + one accent). Chart tints OK.
• ≤ 2 typefaces (display + body). Optional third: mono for metadata.
• No "Thank you" / "Questions?" final slide.
• Every chart has the conclusion in the title. "Revenue 2023" → "Revenue grew 18% in Q4, driven by enterprise renewals."
• Action titles, not topic phrases. "Q3 Results" → "Q3 results beat plan by 22% on enterprise."
• Tabular figures + slashed zero on every numeric column.
• All-caps only on eyebrows ≤ 2 words, weight ≥ 600, tracking ≥ 0.1em.
• Margins ≥ 8% of slide width on left/right, ≥ 6% top/bottom.
• System-font-safe unless playback machine is controlled. If embedding foundry fonts, tested on the actual client OS.
• PDF export attempted as fallback delivery — and tested.

References (load on demand)

• references/typography.md — sizes by context, serif vs sans at distance, contrast, case, font pairing, numeric figures. Load when picking type or auditing legibility.
• references/composition.md — grid systems, focal modes, assertion-evidence framework, slide-type catalog with worked layout sketches. Load when laying out a new deck or auditing slide-type discipline.
• references/premium-aesthetic.md — what makes a deck feel premium vs cheap, concrete examples, color discipline, paper-stock metaphor. Load when the brief says "polished" / "investor-ready" / "premium" / "expensive."
• references/tooling.md — copy-pasteable code for python-pptx, Reveal.js, Marp, and the soffice/pdftoppm verification pipeline. Load when picking a build tool.
• references/bibliography.md — annotated reading list (canonical books, exemplar decks, design systems). Load when researching or citing.

Worked example (load examples/notebook-deck.md)

A 15-slide ET-Book-style interview deck — a notebook/graph-paper aesthetic deck for a Staff+EM applicant interviewing at OpenAI/Anthropic. Two case studies (Comment Ranking 2015, Spaces VR Avatars 2016–2018), full token system mirroring a portfolio website, rendered via python-pptx, verified via soffice + pdftoppm. Walks through every decision named in this SKILL. The build script itself lives in the author's private portfolio repo, not this bundle — see the worked example for what's actually reusable.

Tooling quick-reference (full code in references/tooling.md)

Build:

• python-pptx (Python) — best for templated programmatic decks. Used in the worked example.
• Reveal.js (web) — best for talks where you want web embeds, live demos, or audience-controllable.
• Marp (Markdown → PPTX/PDF/HTML) — best for engineers who want Git-versioned decks.
• Keynote / PowerPoint (native) — best for one-off polish on a deck someone else will edit.

Verify (always run these):

# .pptx → PDF → per-slide PNGs (requires LibreOffice + Poppler)
brew install --cask libreoffice  # macOS
brew install poppler

RENDER_DIR=~/coding/tmp/deck-render   # disposable, but not swept like /tmp
mkdir -p "$RENDER_DIR"
soffice --headless --convert-to pdf deck.pptx --outdir "$RENDER_DIR"
pdftoppm -png -r 144 "$RENDER_DIR/deck.pdf" "$RENDER_DIR/slide"
# Open the PNGs and ACTUALLY LOOK at them. Render is the source of truth.

Run scripts/check_deck_quality.py deck.pptx for an automated audit (body size, color count, font count, title-case violations).

Positive trigger test cases

These should activate the skill:

1. "Help me build a pitch deck for our seed round"
2. "Make a Keynote template for our customer onboarding"
3. "The exec deck looks cheap — make it feel premium"
4. "Convert this brief into an interview deck"
5. "Reveal.js talk for a conference next month"

Negative trigger test cases

These should NOT activate the skill (delegate as noted):

1. "Coach me on how to present this deck" → tech-presentation-interview
2. "Design our company logo" → brand-guidelines
3. "Write the script for the keynote video" → copywriting
4. "Animate the title text" → motion-design-web
5. "Build the conference website" → web-design-expert / frontend-design