Slide Deck from a Document

What you'll build

A repeatable pipeline that turns a written document into a structured slide deck outline — titles, bullets, speaker notes, and visual suggestions — ready to paste into Google Slides, Canva, or PowerPoint.

Similar in spirit to Meeting Notes to Action Items — structured extraction from unstructured input — but the output is a presentation, not a task list.

The flow

Source document
  → Define audience + slide count
  → LLM extracts slide outline (JSON)
  → Human review (reorder, cut, sharpen)
  → Paste into Slides / Canva
  → Apply template + visuals

The LLM does the thinking transfer. You do the taste and layout.

Step 1 — Prepare the source

Not every paragraph becomes a slide. Before extraction:

• Remove appendices, raw data tables, and internal-only sections.
• Mark sections that are context vs. conclusions — the LLM should weight conclusions higher.
• Split docs over ~30 pages by section if using Claude or GPT. Gemini handles longer docs in one pass.

A 15-page strategy brief typically becomes 10–12 slides. A 40-page report becomes 15 slides + 3 appendix slides.

Step 2 — Lock constraints

Tell the model who will see this deck:

Max slide count prevents the LLM from dumping every section header into a slide. 12 slides for a 30-minute talk is a good default.

Step 3 — Extract the outline

Use structured outputs or JSON schema mode. The shape:

{
  "title_slide": { "title": "Q2 Product Strategy", "subtitle": "Board Review — June 2026" },
  "slides": [
    {
      "slide_number": 1,
      "title": "Revenue grew 23% driven by enterprise upsell",
      "bullets": [
        "Enterprise ARR up 31% QoQ",
        "Self-serve flat — pricing test in July",
        "Churn stable at 2.1% monthly"
      ],
      "speaker_notes": "Open with the headline number. Pause on enterprise — that's the story. Self-serve flat is expected until the pricing experiment.",
      "visual_suggestion": "Bar chart: ARR by segment, Q1 vs Q2"
    }
  ],
  "key_takeaways_slide": {
    "bullets": [
      "Double down on enterprise sales motion",
      "Launch pricing experiment in July",
      "Hire 2 solutions engineers by August"
    ]
  }
}

Title rule: slides titles are conclusions, not labels. "Revenue grew 23%" not "Revenue Update."

Bullet rule: max 5 bullets, max 12 words each. If you need more, split into two slides.

Step 4 — Human review (non-negotiable)

Spend 10 minutes on the outline before touching layout:

• Reorder — is the "so what" on slide 2, not slide 9?
• Cut — are any slides repeating the same point?
• Sharpen — are titles conclusions or vague topic labels?
• Verify — does every stat trace to the source doc? LLMs invent plausible numbers.

This is where When NOT to Use AI applies — don't trust unverified data on a board slide.

Step 5 — Paste into your slide tool

Google Slides

1. Create a new deck from your brand template.
2. One slide per outline entry — paste title and bullets.
3. Paste speaker notes into the notes panel.
4. Use Explore → Images for quick visuals, or insert charts from Sheets.

Canva

1. Start from a presentation template.
2. Paste bullets into each slide — Canva Magic Write can reformat.
3. Use Magic Design to auto-suggest layouts from your text.
4. Swap icons and photos from Canva's library based on visual_suggestion.

PowerPoint

1. Paste outline — Designer suggests layouts automatically.
2. Speaker notes go in the notes pane.
3. Best for orgs with existing .pptx brand templates.

Step 6 — Visuals and polish

The LLM suggests visuals; you create them:

• Charts — pull data from the source doc into Google Sheets or Excel, chart it, embed.
• Diagrams — simple flowcharts in Slides/Canva for process slides.
• Icons — one icon per bullet in Canva for visual rhythm.
• Quotes — pull a key quote from the doc for a full-slide callout.

Optional second pass: run the speaker notes polish prompt per slide to match your delivery style.

When to use Gemini vs Claude vs GPT

For most business users, Claude or GPT with a 15-page doc is sufficient. Reach for Gemini when the source doc exceeds 50 pages.

Extensions

• Auto-sync from Google Docs — Apps Script triggers extraction when a doc is marked "ready for slides."
• Template-aware extraction — pass your firm's slide template structure as few-shot examples.
• Multi-language — extract in English, generate a second outline in Spanish/French for regional audiences.
• Gamma / Beautiful.ai integration — export JSON outline directly to auto-visual deck generators for internal decks.

Cost reality

At ~$0.15 per deck, this is effectively free. The time saved is 1–3 hours per presentation — the LLM outline takes 2 minutes, human review takes 10 minutes, layout takes 30–60 minutes. Compare to 3–4 hours starting from a blank slide.

Ship the outline pipeline first. Fancy auto-visual generation is a v2 optimization.