PPTX

/pptx — brand-agnostic, design-library-bound PowerPoint renderer

Apply any resolved brand kit to a slide outline and render a .pptx that meets the design library's quality bar. Brand-agnostic: same code path produces a ClientCo deck, an ClientCo deck, or a Genesys deck — only the resolved BrandConfig differs.

This skill is renderer-only, not author. The outline (titles, body, speaker notes, source-attributed claims) comes from /sales-deck, /proposal, /webinar, /product-launch, /case-study, or you. This skill turns that outline into a brand-applied, design-reviewer-clean deck.

Doctrine inherited (Step 7 — 0626 rollout, locked 2026-06-04)

Renderer skill — inherits doctrine from consuming author skills (sales-deck, proposal, webinar, etc.). Applies [[feedback_execution_doctrine_refinements_step6]] R3 (template caption library defaults to product-update tone — no "thrilled" defaults), R9 (slide-title templates verb-led). R1/R2/R5/R6/R7/R8 cascade from authoring skills per slide content.

The skill sits in primitives/design/ next to /dashboard, /figma-prototype, /vibe-coding and inherits the design library at composition time:

• Layout / typeset / distill / polish / cognitive-load principles via the wrapper references

Adapted from Anthropic's document-skills:pptx example skill — see ATTRIBUTION.md. Python tooling (XML pack/unpack, ECMA/ISO validators, soffice → pdf → jpg pipeline) preserved unchanged.

For full brand-resolution waterfall, parser spec, and token → pptxgenjs field map → the premium reference. For the 12-type slide grammar → the premium reference. For the locked-zone grid → the premium reference.

Triggers

Invoke when user says:

• /pptx (with an outline path or upstream skill output)
• "render this deck as pptx"
• "make a powerpoint for [client]"
• "build a branded pptx"
• "apply [client] brand to this deck"
• "convert this outline to slides"

Do NOT invoke when:

• User wants to write a deck from positioning/messaging (use /sales-deck)
• User wants Google Slides output (use cd.claude/mcp/gdrive && node create-slides.mjs)
• User wants a PDF (use document-skills:pdf)
• User wants to read/extract text from an existing pptx without rendering (use python -m markitdown file.pptx directly)

Inputs

Required:

• slide_outline — markdown file (or upstream skill output) with H1 = deck title, H2 = slide titles, bullets/paragraphs = body, optional speaker notes blocks (> notes: or HTML comments)

Recommended:

• client_slug — drives brand resolution. Any active client, archived client, genesys-growth, or whatever's defined in gdrive-config.json.
• brand-kit — pass --brand-kit-path for one-off external clients or non-canonical kits

Optional:

• template.pptx — existing deck to use as layout source (overrides default 12-type grammar)
• output_path — defaults to projects/consulting/active/{client}/sales/execution/{MMYY}-{topic}.pptx (or projects/genesys/... for Genesys)
• --no-strict — fall through to generic palette when no brand source resolves

Prerequisites

1. Python deps — pip install "markitdown[pptx]" Pillow defusedxml
2. pptxgenjs — npm install -g pptxgenjs (set NODE_PATH=$(npm root -g) when running)
3. LibreOffice — soffice on PATH (for jpg renders during QA loop)
4. Poppler — pdftoppm on PATH (PDF → jpg)

The skill fails loud, not silent, when any of these are missing.

Workflow at a glance (6 phases)

For full per-phase detail → the premium reference.

Brand resolution waterfall (5 layers, brand-agnostic)

Smoke-tested against ClientCo, ClientCo, ClientCo, Pivot, ClientCo, Genesys, ClientCo. Any new client onboarded via /brand-kit resolves immediately with zero /pptx code change.

For the parser spec + DESIGN.md token mapping → the premium reference.

The 12-type slide grammar

Each type is a first-class artifact named for the verb of the slide's job. One type per slide.

Full specs (zones + token application + banned-pattern guards) → the premium reference.

The locked grid system

Every slide places eyebrow / title / subtitle / body / footer / wordmark in identical zones. Zones in inches assuming LAYOUT_WIDE (13.333 × 7.5):

• Eyebrow row — y=0.55, h=0.30 (0.45 on cover/cta), 12pt body-font, secondary color, charSpacing 4–6, uppercase
• Title row — y=0.95, h=1.15, serif heading-font, 38pt content / 54pt cover-cta / 32pt section-divider
• Hairline — y=2.05, w=1.2", weight=2pt, color=secondary
• Subtitle row — y=2.20, h=0.45 (optional), 18pt body italic, muted
• Body zone — y=2.55–6.85, 12-col grid (col=0.92" + gutter=0.10")
• Footer — y=7.00, h=0.30, 10pt body-font muted, slug left + idx/total right
• Wordmark — top-right (content slides) or bottom-right (cover/cta)

No zone wanders ±0.05" between slides of the same type. Full diagram + composition budgets → the premium reference.

Pre-render checklist (run before EVERY render)

• Brand resolved — Resolved layer: N printed (where N ≠ 5 unless explicit --no-strict)
• BrandConfig has full token coverage: colors + fonts + spacing + rounded + components + signature_elements + presentation_guidance + voice_overrides (warnings field flags any defaults supplied)
• Outline parsed — slide count matches user expectation
• Grammar plan written — exactly one type per slide
• No client name appears in render code (brand-agnostic check)
• No hardcoded hex codes / font names / radii / margins in render code — all values reference BrandConfig

Anti-Hallucination + design-library guardrails

1. No invented colors / logos / fonts. Render fails loud if brand resolution falls back to defaults; never substitute "close enough" values.
2. No invented metrics, customer logos, or quotes in slide content. Mark unverifiable claims [UNVERIFIED] per .claude/rules/ontology.md.
3. Voice overrides apply to ALL auto-text (slide titles, eyebrows, footers, source citations). E.g., no-em-dashes strips " — " → ", ".
4. Banned visual patterns refused at composition time (per .claude/rules/design-production.md + impeccable rules in the premium reference): no gradient text, no side-stripes ≥2px, no generic drop shadows on cards, no glassmorphism, no hero-metric template, no icon-tile-above-heading template, no centered prose paragraphs, no bounce easing, no accent line under titles (the AI tell), and 7 more.
5. Quantitative budgets enforced: ≤2 brand colors per slide, ≤2 font weights per slide, 1 primary color per slide for the most important element, ≤7±2 visible options per decision point, ≤10% accent area per slide, mono font ONLY for numerical/technical content.

Composition with other skills

Design cycle (post-authoring phases)

After rendering, walk these phases. Run /design-reviewer as the final ship-ready gate.

• Final review — /design-reviewer (P0–P3 severity findings + remediation)

Completion report

When done, output:

• Output path of rendered.pptx
• Brand source resolved (layer 1–5 + path)
• Slide count + grammar-type breakdown (e.g., "1 cover + 3 narrative + 1 proof + 1 cta")
• Logo / signature elements / voice overrides applied
• Validation status (scripts/office/validate.py)
• QA loop iterations to clean pass
• Brand-agnostic audit result (zero client-slug references in render code)
• Warnings (e.g., "fell through to gdrive-config layer", "no logo found")
• Suggested next: /design-reviewer for final ship-ready gate

External References

• .claude/rules/design-production.md — DESIGN.md token format + shadcn integration contract + banned visual patterns + skill authorship contract
• .claude/mcp/gdrive/gdrive-config.json — brand-source layer 4
• .claude/mcp/gdrive/create-slides.mjs — sibling renderer (Google Slides flavour)
• .claude/mcp/gdrive/upload-pptx.mjs — auto-upload pptx to Drive folder, converts to native Google Slides
• LICENSE.txt — Anthropic proprietary on inherited Python tooling
• ATTRIBUTION.md — lineage from document-skills:pptx

Final ship gate

Run /premortem --output before ship. See /premortem skill for the 5 execution domains (will-it-resonate / will-it-convert / will-it-stay-on-brand / will-stakeholder-push-back / will-it-degrade-over-time) and output template.

Trivial-case escape: ## Premortem\nNo failure modes — trivial change satisfies the contract for genuinely trivial outputs.