SuperAPI AI Directory

slide-design-principles

Design & creation

View repository

Updated 10-16 06:21

About this skill

Slide Design Principles

This is the shared foundation for every deck builder in this category. It defines two things:

  1. The slide-spec — a tool-agnostic, per-slide intermediate the builders produce before generating any .tex or .pptx. The spec is where structure is decided and corrected; the artifact is a mechanical render of an approved spec.
  2. The design rules that apply regardless of format.

Use this skill directly for tool-agnostic slide feedback. Otherwise it is loaded as a dependency by:

  • beamer-academic-deck — seminar, conference, lecture (LaTeX Beamer)
  • beamer-proposal-report-deck — proposal, progress/final report (LaTeX Beamer)
  • business-report-deck — industrial / executive report (PowerPoint via python-pptx, or Canva)

Planning a talk's narrative and pacing (what the one takeaway is, what to cut) is a separate, upstream job owned by research-talk-planner. This skill and the builders assume that planning is done — or do a lightweight version of it inline when no plan exists.


The slide-spec

Every builder produces this before any code. One block per slide. Keep it in markdown so the user can edit it directly.

### Slide N — <one-line working title that is a claim, not a label>
- **Role**: title | agenda | problem | gap | method | result | ablation | timeline | budget | risk | summary | backup
- **Message**: the single sentence this slide must land. If you can't write one, the slide is doing two jobs — split it.
- **Body**: the bullets / equation / statement. Max ~5 lines, ~7 words each. Use "—" for sub-points sparingly.
- **Visual**: figure | diagram | table | none. Describe what it shows and where it comes from (paper Fig. 3, new TikZ, screenshot, generated chart).
- **Build**: (optional) staged reveals — "1) baseline, 2) overlay ours". Only when sequence aids understanding.
- **Notes**: speaker notes / talking points the slide body should NOT contain.

Rules for the spec:

  • One message per slide. The Message field is a forcing function. "And also…" means a second slide.
  • The title is the message compressed, phrased as a claim. Results is a label; Adaptive walks win on all 5 benchmarks is a claim.
  • Body is evidence for the title, not a transcript. Anything you'd say but not show goes in Notes.
  • Mark every visual's provenance so the artifact step knows whether to reuse a file, draw TikZ, or generate a chart.

Confirm the spec with the user before generating the artifact. Correcting structure in markdown is cheap; correcting it in .tex/.pptx is not.


Design rules (format-independent)

1. One signal per slide

A slide answers one question. If a listener can't say what the slide was about in five words, it carried too much. Density is the most common failure; when in doubt, split.

2. Claim-titles

The title is the most-read line on every slide. Spend it on the takeaway, not a category name.

Weak (label) Strong (claim)
Experimental Results Our method cuts error 23% with half the compute
System Overview Three stages, one bottleneck: retrieval
Background Prior work assumes static graphs — ours don't

3. Visual hierarchy

Largest element = most important element. The eye should land on the message, then the evidence, then the detail. If everything is the same size, nothing is emphasized. Use position (top-left first in LTR), size, weight, and color — in that order of strength.

4. Density limits

  • Prose: never a paragraph on a slide. Bullets are compressed speech, not sentences.
  • ~6 bullets max, ~7 words each (the "6×7" heuristic — a ceiling, not a target).
  • One table per slide; ≤ 6 rows / 5 columns visible. Bigger tables are appendix/backup.
  • One primary figure per slide. Two only if they're a direct comparison.

5. Figure adaptation

Paper figures are built for a reader who can zoom and reread; slides are seen once from across a room. Adapt:

  • Strip multi-panel figures to the one panel that makes the point.
  • Increase font sizes; a label readable in a PDF is invisible projected.
  • Remove gridlines, legends, and annotations the audience won't parse live.
  • Add one highlight (arrow, circle, color) that points at the takeaway.

6. Color and contrast

  • Pick one accent color and use it only for emphasis; everything else neutral.
  • Ensure text/background contrast survives a washed-out projector (dark on light is safest).
  • Never encode meaning by color alone — pair with shape, label, or position (colorblind-safe).
  • Consistency across the deck: the same concept gets the same color throughout.

7. Consistency and chrome

  • Fixed layout grid: titles, page numbers, and logos in the same place every slide.
  • Footers/logos are chrome — make them quiet. The content is the message.
  • One font family, ≤ 2 weights. Math in the body font's companion math font.

8. Progressive disclosure

Reveal complex content in steps (builds) so attention follows the explanation. Build equations piece by piece; overlay results onto baselines. Builds should reveal logic, never decorate.


When to push back

Tell the user to fix structure (not styling) when you see:

  • A slide whose Message you cannot write in one sentence.
  • More than ~7 slides per 10 minutes of talk (too fast to follow).
  • A wall-of-text slide that is really speaker notes pasted onto the canvas.
  • A figure copied from the paper at paper resolution and density.
  • Titles that are all labels — a sign the deck has no argument, just sections.

Output

When invoked directly: return the corrected slide-spec block(s) and a short list of the rules each change applies. When invoked as a dependency: hand the agreed slide-spec to the calling builder, which renders it to the target format.