open-kimi-ppt-skill/skills/open-kimi-ppt/reference/design_system/academic/pastel-derivation/design.md

Pastel Derivation · Academic STYLE DESIGN SYSTEM

PART A — Category Baseline (General Minimum Standard)

A rigorous, formally typeset academic presentation benchmarked against top-tier research talks and thesis defenses (main-stage NeurIPS/CHI/SIGGRAPH quality), where clarity of argument takes precedence over decoration. The deck must read as a reproducible scientific claim, not as marketing collateral. Use academic language throughout, with disciplined wording for research questions, hypotheses, evidence, significance, contributions, and limitations; do not mix in marketing or business language. Shift the visual emphasis by field: theoretical derivations should devote most of the canvas to equations and proof steps; computer science and engineering should be structured around system diagrams, code, execution traces, and performance curves; experimental research should tightly organize equations, charts, and images around a single conclusion; humanities and social sciences should add narrative texture and textual evidence—such as archives, maps, and manuscripts—without compromising academic rigor.

Content and layout:

1. Use a fixed argument structure: research question → gap in related work → data and methods → main results → ablation/analysis → limitations → future work. A short deck may compress this structure, but must not change its order.
2. Build evidence-first slides: include result tables with baseline comparison columns, comparison charts, and architecture/process diagrams; present the core contribution as a formally typeset equation.
3. Maintain medium information density: one claim per slide, with evidence organized around that claim. Slides must be composed, not filled—white space is a valid way to separate stages of the argument.

Visual style:

1. Aesthetic: restrained, precise, and scholarly. Do not use promotional copy, decorative photography, or clip-art icons.
2. Typography: use a clear sans serif (render with Liter / Chinese: MiSans) for body text and labels; headings may use a serif if the source style does. Equations must be professionally typeset (KaTeX quality), with standard equation numbering and italic variables.
3. Color: use a two-color system for each deck—one primary structural color and one accent color—with gray establishing hierarchy. Use a white or near-white background. Use only one consistent warning color for negative or ablation signals.
4. Graphics: use thin, precise vector lines for diagrams and minimal horizontal rules for tables. Charts must include axis labels, units, and sample definitions. Do not use 3D, shadows, or gradients as decoration.

Structure and discipline (distilled from work recognized as benchmark quality):

1. Every body-slide title must state a claim—a finding, not a section label: “Distraction Can Be Predicted 10 Seconds in Advance,” not “Experimental Results.” The evidence supporting that claim must follow on the same slide.
2. Align claims with evidence: every quantitative statement on a slide must be supported on that slide by a chart, table, or cited result. No orphan numbers.
3. The main results table must include baseline comparison columns and identify the best value; show relative improvement alongside absolute values.
4. Method slides must include reproducibility details: dataset scale, annotation protocol, inter-annotator agreement, and key hyperparameters. Show only the central idea in equations; move derivations to appendix slides.
5. Every figure must be self-contained: provide axis labels with units, a legend or direct labels, and a caption area specifying sample size and confidence basis. A figure must not rely on its legend alone.
6. Ablations and limitations are required slides, not optional ones: ablations isolate the contribution of each component; limitations accurately state scope, failure modes, and ethical/anonymization status.
7. Every number must include its unit, time span, dataset, and statistical basis (mean ± variance, sample size). Express uncertainty through measurable quantities or confidence intervals, not vague language.
8. Open with the problem statement and a one-sentence contribution. Close with numbered conclusions—claims, not a summary—and an explicit list of open questions.
9. Place citations in a compact footnote line at the bottom of the relevant slide. Do not create a dense, full-slide bibliography wall unless the format requires it.

Default baseline (distilled from the layout skill; PART B may override): organize each slide around one evidence unit (figure/table/equation); give charts complete axis labels, units, and sample definitions; keep charts honest—do not truncate the vertical axis to exaggerate differences—and distinguish correlation from causation explicitly in the copy; ensure projection readability, with body text at ≥14pt equivalent and chart labels at ≥10pt equivalent on a 960×540 canvas.

Title read-through test: reading only the slide titles in sequence should tell the complete story. If the sequence does not flow, the structure is not working.

Density anchor: the specific minimum for each style—evidence modules per slide, chart count, and annotation conventions—is defined by the 【Density Baseline】 reverse-engineered from that style's source deck. This baseline does not impose one universal number.

Global Prohibitions (Mandatory · Prevent Default AI Patterns)

1. Do not default to cards: unless the user explicitly requests them, do not use rounded rectangles or rectangular cards to create hierarchy or alignment. Containers with a genuine information function—such as status indicators, hypothesis boxes, or quotation boxes—do not count as cards. Lines, white space, type size, and weight differences are preferable.
2. Do not default to equal-split compositions: unless no other layout is viable, do not use three equal columns, four equal columns, or a two-by-two matrix by default. Equal division must be justified by the content structure (three scenarios naturally warrant three columns; a framework explicitly named in the content is exempt).
3. Do not use ungrounded AI color schemes: an institutional palette specified in the content is exempt. Prohibited choices are generic, commonplace, and unjustified defaults—blue-purple gradients, cyan-purple neon, rainbow light blobs, glass cards, and glowing borders.
4. Do not use any element inconsistent with the overall style: styles from outside the system are prohibited, such as rounded icons or rounded rectangles in a sharp-edged visual language.
5. Keep chart backgrounds transparent: use transparency or the slide background color, except for reversed charts on dark backgrounds.
6. Do not leave regions blank: do not leave an entire planned region empty within the body content area. Every planned region must contain real content; failing to fill it is the most serious sign of an unfinished layout.

PART B — Signature System (Distinctive Template Signature; PART B Takes Precedence in Conflicts with PART A)

One-line style signature: approachable, rational, whiteboard-like derivations on white, with pale pink–mist blue–mint chapter colors; electric blue and magenta carry fixed variable semantics; low-to-medium density with progressive reveal.

【Color Palette】Body background: #FFFFFF. Use low-contrast gradients only on cover and chapter slides; observed corner colors are approximately #FEF2E8 at top left, #DBEDFE at top right, and #DEF8E9 at bottom right. Use #172C34 for all headings and readable text. Use the primary accent #0064E0 only on one key action phrase within a title. Use annotation blue #0085FD for arrows, links, newly introduced variables, and in-figure labels. Use supporting magenta #C32E8F only for the inverse operation, source term, or limiting phrase paired with blue. Use light gray #F1F4F7 only as a background frame for theorems, conclusions, or conditions, covering no more than 25% of a slide. Use almond #FCEEDC only as a “sandbox” for geometry, distributions, and trajectories. Chart series use #1F77B4, #FF7F0E, #2CA02C, and #D62728; use #0085FD for positive signals and #C32E8F for negative signals/caveats, with each applied to no more than a short phrase. Except in multi-series charts, use no more than two chromatic hues per slide. Pastels do not enter body-slide backgrounds, and accent colors never flood the slide.

【Layout Skeleton】Source canvas: 1920×1080, 16:9. Anchor body-slide titles at the top left, starting at approximately x=2.7%, y=2.8%, and occupying 9%–12% of the height. Begin content at y≈15%; reserve the bottom 4% for bright-blue citations, with no full-width footer. The default reading path is “left-side definition/starting point → central arrow or main equation → right-side geometry/result.” When a horizontal flow is unsuitable, use a centered equation with a conclusion strip at the bottom. Primary evidence occupies 55%–70%; supporting explanation occupies 25%–40%. Do not use equal-weight cards. The cover alone uses a “large center-left title + right-side call to action + contributor information at the bottom” layout. Chapter slides alone use a slightly left-of-center “number + chapter title” layout, with a large QR code on the right only when a genuine external entry point exists.

【Typography】PDF embedding has been verified: use OptimisticDisplay-Medium for headings, with a single key phrase optionally switched to OptimisticDisplay-Bold. Use OptimisticText-Regular/Bold for body text and labels. Use STIXGeneral-Regular/Italic for equations, with italic variables and upright operators. Taking the source deck's body text size of 40 as 1x, the heading size of 74 is approximately 1.85x. There is no consistent large-numeral tier. Main equations are usually 1.4–1.75x, with occasional use of 2.8x on minimal mechanism slides; figure captions are 0.6–0.75x and citations 0.5x. Use sentence case for headings, never all caps. Serif type is reserved for mathematics. Render with {latin: Liter, ea: MiSans} (headings Medium/Bold); use STIX Two Text for equations (via customFonts).

【Chart Language】The source uses single- and dual-line charts, paired bar comparisons, minimal comparison tables, sample matrices, scatter plots, vector fields, trajectory schematics, and process-node diagrams. Tables inherit PART A, with no consistent additional treatment. ⭐Almond sandbox: use contours, point clusters, fine trajectories, and a small number of samples to spatialize abstract objects. ⭐Color-linked equation and diagram: key equation terms share their color with the corresponding arrows, points, and labels in the figure. ⭐Fixed-base progression: retain the same base figure and viewpoint across 2–5 adjacent slides, adding only one trajectory, condition, or equation term per slide; only the new element receives color. Standard performance charts do not label every data point; retain axes, legend, and source.

【Signature Components】①Light-gray theorem/conclusion strip: spans the lower half of the slide or sits next to the relevant equation; use no more than 2 per slide, and only for a theorem, condition, conclusion, or caveat. ②Almond canvas: uses a 1–2 px dark outline; use 1–3 per slide, and only for geometry, distributions, trajectories, and corresponding samples. ③Blue downward arrow + magenta upward arrow: expresses only inverse operations or two directions of operation; use no more than 1 pair per slide. ④Within-chapter route strip: place 3 thumbnails side by side; add a #F1F4F7 backplate to the current stage and reduce the others to approximately 40% opacity. Use only on subsection transition slides. ⑤Progressive construction sequence: established graphics remain fixed in place; each new slide reveals only one new piece of evidence or transformation. Do not use this pattern for a single-slide summary.

【Prohibited】Do not use pastel gradients, almond, or light gray as a full-slide background for ordinary body slides. Do not use serif type for explanatory prose. Do not turn #F1F4F7 rounded frames into generic cards; every frame must serve a theorem, condition, or conclusion. Do not use blue or magenta as decoration without semantic meaning, or swap their roles within the same progressive sequence. Do not expand a complete multi-step derivation on a single slide; split it across progressive slides while keeping the base figure fixed. Do not use reversed text on a dark background for body slides.

【Slide Types and Layouts】 Equation–Conclusion Slide | Definitions, losses, theorems, equivalences | Title 12% + central equation 60% + bottom conclusion 20% | 1–2 blocks | Light-gray conclusion strip, color-linked variables | Unsuitable for photo matrices.
Schematic Derivation Slide | Mechanisms or paths that map algebraic concepts | Almond figure 55%–65% + equation/explanation 35%–45% | 1–3 blocks | Almond canvas, paired arrows, progressive sequence | Unsuitable for dense literature.
Comparative Evidence Slide | Comparison of methods, conditions, or results | Two columns at 48/48 or three evidence units in a row | 2–4 blocks | Blue phrase in title, light-gray conclusion strip | Unsuitable for long equation chains.
Within-Chapter Navigation Slide | Subtopic progress | Three thumbnails across the horizontal middle band | 3 blocks, 1 current state | Within-chapter route strip | Does not carry new evidence.
References/Limitations Slide | Literature, open questions, or caveats | 1–3 light-gray functional boxes stacked vertically | 1–3 blocks | Bright-blue citations, magenta limiting phrases | Unsuitable for main results.
Occasional · Proposition Pause Slide | Problem reframing or within-chapter turn | Large proposition on the left 40% + geometric metaphor on the right 55% | 2 blocks | Almond geometric figure | Unsuitable for dense data.

【Density Baseline】The source deck has low-to-medium density because it relies on progressive construction; no deck-wide hard minimum applies. An Equation–Conclusion Slide usually contains 1 main equation + 0–1 conclusion strip and 0–2 explanatory blocks, with a 15–50 words/characters equivalent per block. A Schematic Derivation Slide contains 1–3 figures and adds only 1 key node per slide; annotations follow variables and trajectories rather than labeling every data point. A Comparative Evidence Slide contains 2–4 modules, 0–2 quantitative charts, and optionally 1 conclusion; it uses 1–2 explanatory blocks, with a 25–60 words/characters equivalent per block. A References/Limitations Slide contains 2–3 boxes, with 3–6 lines per box. Preserve white space between the title and primary evidence, between two layers of equations, and around large figures. When content exceeds capacity, add slides for progressive reveal; do not shrink type to fill white space.