slide-critique

name	slide-critique
description	Negative-space catalog and checklist execution for slide-deck QA. Catalogs AI-generated slide antipatterns, defines the critique output format, and orchestrates execution through the @deck-critic visual-QA pipeline (render + structural-asserts + multimodal rubric). Keywords: critique, audit, review, design quality, visual QA, feedback loop, antipatterns.

Slide Critique

Content placement — This skill is the canonical home for the negative-space catalog (AI-generated slide antipatterns) and checklist execution (critique output format, verdict logic). The positive design rules ("what good looks like") live in presentation-design. The execution pipeline (render + structural assertions + per-slide multimodal rubric) lives in pptx-visual-qa and pptx-structural-asserts and is run by @deck-critic.

Table of Contents

1. When to Use This Skill
2. Core Principle
3. Execution Model — How Critique Actually Runs
4. AI-Generated Slide Antipatterns
5. Critique Checklist
6. Critique Output Format
7. Verdict Logic
8. References

When to Use This Skill

Load this skill when:

• You are @deck-critic and need the antipattern catalog + critique output format
• You are @deck-builder applying pre-flight checks before handoff
• You need to interpret a qa-report.json produced by the QA pipeline

Core Principle

Quality comes from iteration, not first-pass generation.

The first render is almost never correct. Treat critique as a bug hunt, not a confirmation step. If you found zero issues on first inspection, you weren't looking hard enough.

Execution Model — How Critique Actually Runs

This skill is not a "mentally walk through the code" exercise. The actual critique is executed by @deck-critic via three concrete tools:

1. Headless render — pptx-visual-qa/scripts/render_pptx.py converts output.pptx → output.pdf → slide-{N}.png via soffice --headless (LibreOffice). Each PNG is then viewed by the critic agent and graded against the per-slide rubric in pptx-visual-qa/references/visual-rubric.md.
2. Structural assertions — pptx-structural-asserts/scripts/check_pptx.py loads the deck via python-pptx and runs deterministic checks (overflow, contrast, alignment grid, layout-hash repeats, underline spam, dark/light run length, speaker-notes presence, duplicate titles).
3. Antipattern sweep — the @deck-critic agent applies the catalog below to the per-slide PNGs.

The verdict (pass | revise) and top-fixes-json are produced by @deck-critic and consumed by @story-orchestrator. Builders fix on retry.

AI-Generated Slide Antipatterns

The catalog below is the canonical home for AI-slide antipatterns. Do not duplicate this list in agent prompts or other skills.

❌ Antipattern 1: The Accent Line Under Every Title

Thin decorative lines under every headline are the #1 hallmark of AI-generated decks. They add no information and create visual monotony.

Detection: pptx-structural-asserts/check_pptx.py.title_underline_count > 2 across the deck.

Fix: Use whitespace to separate title from content. Reserve accent lines for 1–2 emphasis slides only — or eliminate entirely.

❌ Antipattern 2: Identical Layout Repetition

Every content slide uses the same headline-bullets-on-light-background template. The deck feels like a document.

Detection: repeated_layout_hash flags any consecutive pair of slides with matching shape-position + text-frame-count signature.

Fix: Vary layouts — split layouts (text+visual), big statement slides (one phrase, massive font), data metric grids (2–3 large numbers), full-bleed visual with text overlay, comparison columns.

Rule: No two consecutive slides share the same layout pattern.

❌ Antipattern 3: Text-Heavy Slides

More than 4 bullet points, or bullets exceeding 15 words.

Detection: overflow_violations or per-slide text_density: high.

Fix: Split into multiple slides. Replace text with a single bold statement, a 60–72pt number, a comparison, or a tension question.

❌ Antipattern 4: Generic Color Palette

Default blue accent, white background. Could be any presentation about any topic.

Detection: visual rubric — uniform palette across all slides with no section-color signaling.

Fix: Choose colors informed by topic; use section-specific accent colors to reinforce narrative structure (see slide-design-systems and presentation-design Section Colors).

❌ Antipattern 5: Uniform Visual Weight

Every element competes for attention equally. No clear focal point.

Detection: visual rubric focal_point_present: false.

Fix: One element dominates (60% visual weight). Subordinate the rest through size, color, or opacity.

❌ Antipattern 6: Decorative Over Functional Shapes

Shapes (circles, bars, stripes) that add no information — "making it look designed."

Detection: visual rubric flags decorative-only elements; structural hash often unchanged when removed.

Fix: Every non-text element must carry meaning — section identity, progress indication, data emphasis. If it doesn't, remove it.

❌ Antipattern 7: Missing Narrative Rhythm

Flat sequence of equally-weighted slides. No peaks, no valleys.

Detection: max_same_bg_run >= 3 or no Big-Statement / Question / Section-Divider slides at narrative turning points.

Fix: Use section dividers as pacing devices. Alternate between high-density evidence slides and low-density statement/question slides. Insert "pause" slides before major transitions.

❌ Antipattern 8: Missing Speaker Notes

A deck without speaker notes is a half-built deliverable.

Detection: speaker_notes_missing non-empty.

Fix: Add 2–3 sentences per slide with concrete data points and a transition cue.

❌ Antipattern 9: Duplicate Titles

Two slides with the same headline indicate either a layout dupe or incomplete authoring.

Detection: duplicate_titles non-empty.

Fix: Differentiate by adding the slide's specific takeaway, or merge slides if redundant.

❌ Antipattern 10: Low-Contrast Text

Text not readable against its background — common when accent colors are used for body text on similar-luminance backgrounds.

Detection: contrast_violations (luminance ratio <4.5:1).

Fix: Adjust text color or background to meet 4.5:1 minimum (WCAG AA).

Critique Checklist

Run these against every generated deck. Each item maps to a structural or visual check.

Layout Variety (CRITICAL)

• No two consecutive slides share the same layout pattern (repeated_layout_hash empty)
• At least 3 different layout types across the deck
• Big-statement / question slides at narrative turning points
• Split layouts used for at least one comparison or evidence slide

Visual Composition

• Each slide has a clear focal point (visual rubric)
• Whitespace exceeds 30% on every content slide
• Elements aligned consistently (alignment grid)
• No overlapping text or elements
• Margins ≥0.75" from edges

Text Compression

• No slide has more than 4 bullet points
• No bullet exceeds ~15 words
• Headlines 4–10 words, action statements
• At least 2 slides have NO body text

Color & Contrast

• All body text passes 4.5:1 contrast ratio
• Dark/light slide ratio roughly 30/60/10
• No max_same_bg_run >= 3

Storytelling Coherence

• Headlines alone tell a coherent story
• Emotional arc progresses (no two adjacent slides target the same emotion)
• Opening creates curiosity (no agenda openers)
• Closing has specific, concrete CTA

AI Antipattern Sweep

• Title underlines on ≤2 slides
• No identical layouts on consecutive slides
• No generic Agenda/Overview slides
• No text-only slides without visual element
• Shapes serve function, not decoration
• Every slide has speaker notes
• No duplicate titles

Critique Output Format

@deck-critic produces qa-report.json (schema: .story-telling-stm/schemas/qa-report.schema.json) and, when verdict is revise, a top-fixes.json with ≤5 ordered actionable fixes. Both formats are normative — the orchestrator parses them.

The legacy markdown form is deprecated; only the JSON form is consumed.

Verdict Logic

pass requires all:

• All structural checks pass.
• ≤2 non-blocking findings.
• No critical antipatterns (1, 2, 7, 8, 9, 10).
• ≥3 distinct layouts present.
• Headline sequence reads as a coherent story (see Slide-Sorter Test below).
• At least one fix-and-verify cycle has occurred this session OR the first-pass deck cleanly cleared every check.

Any failure → revise with ordered top-fixes. The orchestrator applies retry policy (max qa_iteration = 2 before user surfacing).

Numeric score does NOT auto-flip the verdict. Per Q3 decision in session 2026-05-04-5707a9ef, the boolean pass / revise / error remains the canonical gate. The 0–100 weighted score (next section) is a prioritization signal only.

Slide-Sorter Test (Deck-Level)

Source: research §A lines 44–45 + narrative-craft/SKILL.md "Throughline" section. Promoted to a deterministic critic check in session 2026-05-04-5707a9ef.

The slide-sorter test reads slide titles in deck order, ignoring section dividers, and asks one question: "Does this parse as a coherent executive summary of the proposal's throughline?"

Procedure (deterministic):

1. Load proposal.md; extract ## Throughline (one sentence).
2. Load deck-spec.json; extract slides[].title in order, skipping slides whose type ∈ {section-divider, title}.
3. Concatenate the titles into a paragraph.
4. Apply the Punch Test (narrative-craft/references/headline-craft.md) per title: every title that is a topic label (not a claim / question / stat) is a finding.
5. Apply the throughline test: does the paragraph's argument map onto the throughline's current-state → tension → resolution → action structure? If a beat is missing or out-of-order, that's a finding.
6. Record findings as antipattern slide_sorter_failure (per-title failure → severity: warn; whole-deck failure → severity: fail).

A whole-deck slide-sorter failure is a revise trigger. Per-title failures rank in top_fixes ordered by their per-slide score (see next section).

Numeric Quality Score (0–100, prioritization-only)

Source: research §8 lines 367–376; ported in session 2026-05-04-5707a9ef. Per Q3 decision the score is a prioritization signal, never a gate. Boolean verdict still wins.

For each slide and for the deck as a whole, the critic emits a weighted 0–100 score. Use it to:

• Rank top_fixes — the orchestrator surfaces fixes for the lowest-scoring non-appendix slides first, capped at 5.
• Track convergence — across qa_iterations, deck-level score should rise. A flat or falling score across two iterations means the retries aren't converging; surface to the user.

The Weighted Rubric

Dimension	Max points	What it measures
Narrative clarity	20	Slide carries one assertion that fits the throughline; AEI triad complete
Assertion title quality	15	Punch Test pass; ≤14 words; specific + quantified when possible
Evidence strength	15	Evidence type matches assertion's relationship; sourced; visual over textual
Layout & hierarchy	20	One focal point; visual hierarchy clear in 3s; layout doesn't repeat the prior slide
Visual polish & brand consistency	15	Design-system tokens applied consistently; no decorative-only shapes; whitespace ≥30%
Chart / image quality	10	Chart relationship → chart type per pptx-engine/references/chart-selection.md; image earned its place per presentation-design/references/image-direction.md
Accessibility & technical correctness	5	Contrast ≥4.5:1; alt text on meaningful images; speaker notes present; no overflow
Total	100

Deck-level score = average of non-appendix per_slide_scores[].total.

Reference Targets (Diagnostic, Not Gates)

• Deck score ≥90: deck is healthy; surface as positive.
• Deck score 80–89: ship-able; recommend a follow-up polish pass.
• Deck score <80: investigate even if verdict == pass — likely one or two outlier slides dragging the average.
• Per-slide score <85 for non-appendix slides: the slide is weakest in the deck and goes into top_fixes first.

These are diagnostic targets, not auto-revise triggers. A deck with verdict: pass and score: 78 ships if structural + visual checks pass; the score is reported to the user in qa-summary so they can decide whether to ask for another iteration.

Recording the Score

The critic records the score under the score and per_slide_scores fields of qa-report.json (schema v2.1.0, see .story-telling-stm/schemas/qa-report.schema.json). When the score is omitted (e.g. on error verdict or when the critic has not implemented the rubric), absence is allowed and the orchestrator falls back to verdict-only behaviour.

References

• AI Slide Antipatterns — Expanded catalog with concrete examples
• pptx-visual-qa — Render pipeline + per-slide visual rubric
• pptx-structural-asserts — Programmatic python-pptx checks
• presentation-design — Positive design rules (the "what good looks like" companion)