Deck — presentation orchestrator (4-stage)

When the user asks for a "presentation", "slides", "deck", "presentazione", "slide" — go through four stages in order. This skill is the orchestrator: stages 1-2 are content discovery, stage 3 chooses the output format, stage 4 delegates the actual rendering to a format-specific skill (or to the cerase-deck-renderer MCP for the fast HTML/PDF path).

Stage 1 — brief (interview)

Ask the user, one question at a time:

1. Audience: who will read this? (board / sales prospect / team / customer)
2. Objective: what should the audience think or do after reading?
3. Roughly how many slides + orientation? (e.g. 10 slides 16:9)
4. Content: paste or describe the source material (bullet points, doc URL, free text).
5. Brand (optional): should the deck follow a visual brand? Capture whatever the user offers — a primary/accent colour, a background/text colour, a heading/body font, or a ready-made CSS snippet they already have. If they have none, skip it: the deck renders on the default theme.

When answered, write presentation-brief.md in the workspace with the structured fields, including a Brand field:

• record the brand colours / fonts exactly as given, or the raw CSS verbatim if the user pasted one;
• write Brand: default when the user wants no brand override.

Confirm with the user before moving on.

Stage 2 — content draft (format-agnostic markdown)

Read presentation-brief.md from the workspace. Produce presentation.md, format-agnostic:

• Cover: # <title> + 1-line subtitle

• Slides separated by --- on its own line (blank lines above and below)

• Slide title: ## <title>

• Slide bodies: ≤ 7 bullets max, parallel structure, concrete numbers over adjectives

• No HTML, no inline scripts, no remote images (slide tools render them inconsistently)

• Chapter cover (optional — long, multi-part decks): to open a major section/act, wrap a whole slide in a :::chapter fence — a # H1 chapter title + optional subtitle line(s) — for a full-height act divider. One per major act (a 3-4 act deck gets 3-4); it's an act break, not a per-topic transition. Example:

  :::chapter
  # Part 1 — The problem
  an optional subtitle line
  :::
  

  Rendered only by the HTML/PDF path (cerase-deck-renderer / md2 ≥ 0.2.1); other output formats ignore the fence.

Show the user the slide titles + 1-line outline. Wait for green light or revisions.

Stage 3 — format chooser

Ask the user explicitly which output they need. Don't assume.

Options:

Pick one primary format. Offer to render a second format afterwards if the user wants a backup copy.

Stage 4 — delegate to the right backend

Path A — HTML / PDF (fast, native to Cerase)

Read presentation.md (the deck) and the Brand field from presentation-brief.md.

Brand → template_css. When the brief carries a brand override, turn it into a small CSS snippet and pass it as template_css. It is applied on top of the default md2 theme (appended after the default template's own CSS, so your rules win):

• raw CSS pasted by the user → pass it verbatim;
• brand colours / fonts → derive a minimal override, e.g. the brand colour on headings + accents and the brand font on the deck. Keep it to the few brand tokens the user actually gave — don't invent a full theme.

When Brand: default (no override), omit template_css entirely — the deck renders on the default theme.

Call (include the template_css argument only when there is a brand override):

call_recipe("cerase-deck-renderer.render", {markdown_content: <full file contents>, output_filename: "presentation.pdf", template_css: <brand CSS>})

(or output_filename: "presentation.html" for HTML responsive).

The recipe returns {filename, size_bytes, contents_base64}. Decode the base64 and write to the workspace. Attach to the reply.

Path B — PPTX / ODP / Google Slides

Hand off to the pptx skill (system-opt-in, attached by template). Input it the presentation.md workspace path + the chosen output format. The pptx skill produces the artefact and writes it back to the workspace.

Path C — DOCX (when the user actually wants a document, not slides)

Hand off to the docx skill. Same contract: pass workspace path + chosen output format.

If the corresponding format-specific skill is not attached to your Agent template (admin didn't opt in), tell the user politely, in their language, that exporting to needs the skill, and offer to ask the admin to enable it. Don't try to bash + python it yourself.

Language rules

• Chat: in the user's language.
• Brief + draft artefacts: in the user's language.
• File names: use the title slug + extension, e.g. q3-results-presentation.pdf.