carousel-pdf

name	carousel-pdf
description	Produce a LinkedIn 4:5 portrait carousel (PDF document post) from the active company's knowledge pack. This skill should be used when the user says "make a carousel about [topic]", "turn this into a carousel", "lead-magnet PDF for [topic]", "make a myth-bust carousel", "make a how-to carousel", "make a case-study carousel", "make a framework carousel", "make a light carousel", or "post about [topic]". Takes a topic or signal → selects arc shape (myth-bust, how-to, case-study, framework, or default) → drafts the 8–10 card copy arc with swipe momentum → shows for approval → renders an on-brand PDF + per-slide PNGs → outputs caption (2 variants) and a close (compliant recap + single ask by default; opt-in trigger-word lead magnet with DM copy). Dark (default) or light theme per carousel.
license	MIT
metadata	{"version":"0.5.0","phase":"2B","capability_tier":"core"}

carousel-pdf Skill

Produce a professional LinkedIn carousel — a 4:5 portrait PDF document post — from any topic, market signal, or source document. Dark-cinematic brand, the active company's design system, Slidev engine. Every carousel ships with a full publish package: PDF + PNGs + caption + trigger word + DM copy.

Resolve the active profile (the agent provides it; everything company-specific loads from profiles/<active>/, never from plugin/). Read brand identity from profiles/<active>/PROFILE.md (brand_name, social_handle, deck_byline) and all content facts from profiles/<active>/knowledge/.

---

When to use this skill

• User says: "make a carousel about X", "turn this blog into a carousel", "lead-magnet PDF for X", "post about [topic] from the knowledge pack", "make a light carousel", "make a myth-bust carousel", "make a how-to carousel", "make a case-study carousel", "make a framework carousel"
• market-scan output lists a carousel as an output type for a signal
• User wants a LinkedIn document post on any product, use-case, or industry angle
• A builder-studio article needs repurposing ("make a carousel from this article"): treat content/<active>/journey/assets/<item-id>.article.md as the source document — same facts, one more surface; every claim stays traceable to the article's evidence pack

Arc shapes available (see references/sample-carousels.md): myth-bust · how-to · case-study · framework · default/product-led

Theme options: dark (default, cinematic) · light (white background — set colorSchema: light in the global frontmatter; see references/light-theme.md)

What this skill does NOT do

• Does NOT auto-send DMs or post to LinkedIn (draft only — human gate required)
• Does NOT generate images/visuals (handled by carousel-visuals skill)
• Does NOT invent product facts — all claims pulled from profiles/<active>/knowledge/

---

Workflow — exact sequence

Step 0 — Read current product facts

The carousel must be built on current product facts, not stale snapshots. There is no sync step on the sidecar render path: the brain composes every card's copy inline from profiles/<active>/knowledge/ (the catalog the next step reads), and the deck-renderer sidecar bakes the deck-theme into its own image. Nothing to refresh — go straight to reading the sources.

Step 1 — Read the knowledge sources

Mandatory reads before drafting any copy:

File	What to pull
profiles/<active>/knowledge/product.md	Feature names, pain owners, solution themes, discovery questions
profiles/<active>/knowledge/case-studies.md	Proof points — verbatim quotes, metrics, customer names
profiles/<active>/knowledge/icp-personas.md	Buyer context — who this carousel is for
profiles/<active>/knowledge/company.md	Brand voice, tagline, language conventions, "never say" list
references/platform-playbook.md	Platform mechanics + attention craft — hook formulas, swipe momentum, CTA compliance, retention curve. The shared authority for every LinkedIn/X fact below.

Never invent product claims. If a stat or feature name isn't in profiles/<active>/knowledge/, omit it or ask.

Step 2 — Choose theme and arc shape

Theme: Ask (or infer from context) whether the colleague wants dark (default, cinematic) or light (white background, editorial). Default to dark if unspecified.

• Dark: colorSchema: dark in global frontmatter. No other changes.
• Light: colorSchema: light in global frontmatter. The sidecar's baked deck-theme honours it — see references/light-theme.md for the light palette details.

Arc shape: Check references/sample-carousels.md for the four arc templates (myth-bust, how-to, case-study, framework). If the topic maps clearly to one shape, say which shape you're using and proceed. If ambiguous, ask: "This could work as a [myth-bust] or a [framework] — which angle do you want?"

Step 3 — Draft the copy arc (show for approval before rendering)

From the topic and chosen shape, draft:

1. Three hook options (different formulas — pick from the six-shape library in references/platform-playbook.md §4: factual gap, contrarian, numbered promise, outcome + timeframe, stop-doing-X, how-[entity]-does-it). Judge each against the 3-second test. See references/sample-carousels.md for shape-specific templates, and docs/hook-craft.md for the cross-surface discipline every option must still pass: zero-context self-containment, one real fact from profiles/<active>/knowledge/, and a belief (never a named vendor) as any fault-line.
2. The arc: 8–9 cards following the chosen shape from references/sample-carousels.md. Apply the momentum rules (playbook §5): ≤50 words per card, every middle card ends with a forward pull, re-hook at card 4–5.
3. The close mode (playbook §2): default = compliant close (recap + ONE ask — save, follow, or a genuine question — + identity block). Offer the trigger-word lead-magnet close ("Comment [WORD] → DM") only as an explicit opt-in when the colleague wants lead capture — auto-pick the word (TRUST / GATEWAY / AGENTS / IDENTITY / FABRIC / PROOF, logic below) and note the cadence cap (≤1 lead-magnet post per 3–4 weeks, manual DM fulfilment only).
4. Per-card summaries (one line each: card role, headline, component/layout, and the card's forward pull)

Show this draft arc to the user. Do not render until the arc is approved. This is the most important step — the script is the product.

Default arc (standard/product-led — use when no shape fits cleanly)

Card 1  — HOOK     : [hook option chosen]
Card 2  — PROBLEM  : [specific pain — mapped to a persona]
Card 3  — RISK     : [consequence if unaddressed]
Card 4  — STAT     : [hard evidence — from profiles/<active>/knowledge/, not invented]
Card 5  — BRIDGE   : [reframe — "this isn't X, it's Y"]
Card 6  — SOLUTION : [product answer — feature name + one sentence]
Card 7  — HOW      : [mechanism — one proof point or detail]
Card 8  — QUOTE    : [verbatim customer voice — from profiles/<active>/knowledge/case-studies.md]
Card 9  — CLOSE    : [recap + ONE ask + identity block — or Comment TRIGGER → promise in lead-magnet mode]

For myth-bust, how-to, case-study, and framework arcs — use the templates in references/sample-carousels.md instead.

Minimum 8 cards, maximum 10. Default: 9.

Step 5 — Compose the deck source

Create the carousel working folder under the resolved content root (resolve it via gtm_core.paths.resolve_content_root() — never the repo root, never the deck workspace): content/<active>/carousels/<topic-slug>/. Write into it:

• slides.md — the approved arc, composed with the layout conventions below. The file must be named exactly slides.md (the deck-renderer tool requires that name). No package.json and no scaffold step — the sidecar bakes the deck-theme; the brain only authors slides.md.
• outputs/ — for caption.md, dm.md

In the global frontmatter of slides.md, set colorSchema to match the chosen theme:

colorSchema: dark   # default — dark cinematic
# OR
colorSchema: light  # white background — honoured by the baked deck-theme (see references/light-theme.md)

Do not add a theme: field — the deck-renderer sidecar pins the theme to its baked copy at render time.

Critical structure rule: No bare --- separators between cards. Each slide's frontmatter block (---\nlayout:...\n---) IS the separator. See references/layout-conventions.md.

Step 6 — Render (export via the deck MCP tool)

Render by calling the deck-renderer sidecar — this is the render path; the brain never runs node/npm/npx.

Call mcp__deck__export_deck with:

• slides_md_path = the absolute path to the slides.md you wrote in Step 5 (e.g. <content-root>/<active>/carousels/<topic-slug>/slides.md)
• format='pdf' → the carousel PDF (the post deliverable)

For per-slide PNGs (social preview), call it again with format='png'.

The tool returns the output file/folder path. On a [deck-error] string, DO NOT retry — surface it to the operator verbatim. No Chromium install, no npm — the sidecar owns the toolchain.

Operator-run fallback (only when the deck tool is absent). If the deck MCP tool is not available — the sidecar isn't configured (deck_renderer_url unset), so the deck server is omitted — fall back to a local export by the operator. Hand off the deck source (slides.md) and tell the operator to render it from their Terminal in their local Slidev deck workspace. Do not attempt node/npm/npx yourself — they are denied by the least-privilege policy; retrying only burns budget. The local commands (operator runs these, not the brain):

# From the operator's local Slidev deck workspace directory:
npm run export -w "@<deck-scope>/<deck-slug>"                                    # → outputs/carousel.pdf
npm run export -w "@<deck-scope>/<deck-slug>" -- --format png --output outputs/slides   # per-slide PNGs
npx playwright install chromium   # first export only — install headless Chromium once

Step 7 — Publish package

After export, write to outputs/:

File	Contents
caption.md	2 caption variants: Hook-led (cold audience) + Story-led (warm audience). First ~140 chars carry the hook; close with a genuine question; 0–3 hashtags (never stacked); alt-text line for the cover.
dm.md	Lead-magnet mode only. Delivery DM + one follow-up. Draft only — fulfilment is manual, never automated. Omit this file in the default close.

See references/script-conventions.md for copy rules and references/platform-playbook.md §7 for caption mechanics.

Then append ⟦FILE:…⟧ sentinels so the Telegram cockpit delivers the PDF automatically:

⟦FILE:/absolute/path/to/content/<active>/carousels/<topic-slug>/outputs/carousel.pdf⟧

Put this on its own line at the very end of your response, after all prose. Use the real resolved absolute path.

---

Layout catalog (carousel family)

All layouts live in .engine/deck-theme/layouts/. They are additive — keynote decks are untouched.

Layout	Use for	Key elements
carousel-hook	Card 1 — opening (every carousel)	Full-bleed ambient, oversized h1 (5.5rem), sub-headline, swipe prompt
carousel-point	Cards 2–3, 5–7 — value & bridge	Numbered pill, h2 (3.4rem), body (1.25rem), optional left accent bar
carousel-stat	Card 4 — proof stat	Giant gradient number (7.5rem), context label, source attribution
carousel-quote	Card 8 — customer voice	Glass card, quote-body (1.75rem), attribution with — separator
carousel-cta	Card 9 — close + trigger	trigger-word class (7.5rem gradient), cta-verb, cta-body, cta-promise

Persistent footer on every layout (no extra code needed):

[handle]    [cardNo / cardTotal]    [brand symbol]

The [brand symbol] is the active company's mark (resolve from profiles/<active>/PROFILE.md → brand_name; brand assets in profiles/<active>/knowledge/brand/). Set handle, cardNo, cardTotal in each slide's frontmatter.

Frontmatter fields (all layouts)

layout: carousel-hook    # or carousel-point / carousel-stat / carousel-quote / carousel-cta
page: fabric             # product theme — sets --product-accent, --product-mesh, etc.
handle: '<social_handle>'  # footer left — resolve from profiles/<active>/PROFILE.md social_handle
cardNo: 1                # footer center left
cardTotal: 9             # footer center right
halo: true               # carousel-hook only — show PulseHalo (default: true)
accent: true             # carousel-point only — left gradient bar (default: false)
triggerWord: 'TRUST'     # carousel-cta only — used in caption.md

CSS classes (slot content)

These go in the slide's markdown/HTML content, not the layout itself:

<!-- carousel-hook -->
<span class="eyebrow">Label here</span>
<h1>Big claim</h1>
<p class="lede">Sub-headline</p>

<!-- carousel-point -->
<span class="card-num">02</span>  <!-- or <span class="chip-tag">[Product name]</span> -->
<h2>Point headline</h2>
<p>Body copy. <strong>Emphasis.</strong></p>

<!-- carousel-stat -->
<span class="eyebrow">Label</span>
<div class="divider" />
<p class="stat-number">#1</p>
<h2>Stat headline</h2>
<p class="stat-label">Description</p>
<p class="stat-source">Source: ... · Year</p>

<!-- carousel-quote -->
<div class="glass-quote">
  <p class="quote-body">"Quote text here."</p>
  <p class="attribution">Customer name · context</p>
</div>

<!-- carousel-cta -->
<p class="cta-verb">Comment</p>
<p class="trigger-word">TRUST</p>
<p class="cta-body">...and I'll DM you the <strong>Asset name</strong>...</p>
<p class="cta-promise">↓ Save this for [use case]</p>

---

Copy rules

(Attention-craft authority: references/platform-playbook.md §4–§7 — the rules below are the working summary.)

One idea per card. No exceptions.

If a card needs two ideas, it's two cards. If it needs a paragraph, it's two sentences. ≤50 words per card (target 25–40); max 6–8 words per line, broken at phrase boundaries.

Every card earns the next swipe

Middle cards (2–8) end with a forward pull — an open loop, a transition phrase ("Here's why →", "That's not the worst part."), or the next visible number in a sequence. Re-hook at card 4–5: readers decay along the retention curve (playbook §1), so place a second strong beat mid-deck rather than front-loading everything.

Headings

• h1 (hook): 3–8 words. Factual, specific, provable. No corporate buzzwords. Must pass the 3-second test — the cover competes as a feed thumbnail, not a page.
• h2 (point/stat/quote): 5–8 words. Statement, not question (unless the question IS the hook).
• Never generic: "AI is changing everything." Always specific: "Every agent looks like a ghost."

Language conventions

Match the active company's language/spelling conventions (e.g. British English: organisations, authorised, recognise, personalised). Consistent with profiles/<active>/knowledge/company.md.

Never invent stats

Every number must come from profiles/<active>/knowledge/. If no stat exists for the topic, use a qualitative claim or a verbatim quote. No fabricated percentages.

The hook card (Card 1) — six formulas

From the playbook §4 library:

1. Factual gap: "X doesn't have Y. Yet." (specific, provable absence)
2. Contrarian: "Everyone is building [X]. Nobody is building [what matters]."
3. Numbered promise: "5 things your AI governance stack is missing." (specific numbers lift CTR ~36%; odd numbers read credible)
4. Outcome + timeframe: "[Customer] fixed [problem] in [timeframe]. Here's how."
5. Stop doing X: "Stop [common practice]. Do [better thing] instead."
6. How [known entity] does it: "How [recognisable company/framework] handles [problem]."

Always write 3 options (different formulas). Judge each against the 3-second test. Show to user. Pick together.

The close card (Card 9) — recap + ONE ask

Default close: one-line recap of the takeaway + exactly one ask (save / follow / genuine question) + identity block. Trigger-word lead-magnet close only on explicit opt-in — see playbook §2 for the compliance rules. Never stack asks on one card.

---

Product → page theme mapping

The deck-theme ships six fixed page: token sets (engine slots with baked accent colours — the data-page tokens in references/layout-conventions.md). Map the active company's products onto them by capability, resolved from profiles/<active>/PROFILE.md → products[] and profiles/<active>/knowledge/product.md — never by any one tenant's product names:

Capability of the product the carousel is about	page: value
Trust / identity / governance (typically the flagship)	fabric
Safety / moderation / guardrails	elements
Data / developer platform	radix
Messaging / payments / transactional	messaging
Creation / builder tooling	forge
Corporate / multi-product / platform story	website

Flagship-product carousels default to page: fabric; multi-product stories to page: website.

---

Trigger word auto-pick logic (lead-magnet mode only)

Applies only when the colleague opts into the lead-magnet close (playbook §2: grey-zone tactic — ≤1 per 3–4 weeks, manual DM fulfilment, never automated). In the default compliant close there is no trigger word; the big gradient word on the carousel-cta layout carries the single ask instead (e.g. FOLLOW, SAVE, or the asset name).

Pick the most on-theme trigger word from:

The words below are exemplars — when the active company's own product vocabulary (profiles/<active>/knowledge/product.md) offers a more on-theme single word, prefer it.

Word	Use when
TRUST	Identity, governance, compliance, audit — when trust is the thesis
GATEWAY	Product-led posts for a gateway/proxy-shaped flagship, developer audience — swap for the active product's own category word
AGENTS	Multi-agent, agentic AI, "what to do about agents" framing
IDENTITY	DID, verifiable credentials, identity-first positioning
FABRIC	Full-stack / platform story — swap for the active platform's own name-word
PROOF	Case study carousels, customer evidence, ROI

In lead-magnet mode the trigger word must appear on the CTA card AND in both caption variants AND in dm.md — one word, everywhere identical.

---

Export — quick reference

Render is one MCP call (see Step 6 for full detail):

• PDF (the post): mcp__deck__export_deck(slides_md_path=<abs path to slides.md>, format='pdf')
• Per-slide PNGs: mcp__deck__export_deck(slides_md_path=<abs path to slides.md>, format='png')

Returns the output path; on a [deck-error] string DO NOT retry — surface it to the operator. The operator-run local fallback (npm/npx in their Slidev deck workspace) applies only when the deck tool is absent — see Step 6.

---

Example carousel (validated reference)

content/<active>/carousels/ai-agent-identity/ — "Why AI agents need verifiable identity"

• 9 cards · dark · page: fabric · trigger word: TRUST
• Sources: profiles/<active>/knowledge/product.md, profiles/<active>/knowledge/case-studies.md
• Caption + DM: content/<active>/carousels/ai-agent-identity/outputs/

---

Pre-flight checklist

Before rendering:

• slides.md written under the content root (content/<active>/carousels/<slug>/slides.md), named exactly slides.md
• Arc approved by user — script confirmed before render
• Cover passes the 3-second test — insight-stating headline, 3–8 words
• Every card ≤50 words; every middle card ends with a forward pull; re-hook present at card 4–5
• Every stat has a source line (stat-source class or profiles/<active>/knowledge/ reference)
• No invented claims — all facts traceable to profiles/<active>/knowledge/
• Close card carries exactly ONE ask; lead-magnet mode only if explicitly opted in
• Lead-magnet mode: trigger word identical across CTA card, caption A, caption B, and dm.md
• cardNo / cardTotal correct on every slide
• No bare --- separators (creates blank slides in the PDF)
• handle set — matches PROFILE or user instruction

After export:

• PDF opens — 9 pages, portrait orientation
• Background matches chosen theme: dark #020617 or light #F8FAFC
• Fonts rendered — FatFrank display, Figtree body
• Footer visible on every card — handle · count · symbol
• Close card: the single big gradient word (ask or trigger word) renders large and legible
• No blank pages in the PDF