Deck Prompt

Purpose

Decompose a markdown or PDF report into a structured deck specification JSON file that the nano-banana --deck mode can consume to generate presentation slide images. Claude reads the source document, understands its structure and content, selects appropriate slide types, writes structured slide fields (heading, visual metaphor, labels, text panel) conforming to a strict schema, and outputs a complete deck spec ready for image generation.

Path Resolution

When this skill is loaded, Claude Code displays a "Base directory for this skill:" header. Use that absolute path to resolve file paths in all commands below. In the examples, $BASE_DIR represents this path. The sibling nano-banana skill lives at $BASE_DIR/../nano-banana.

Workflow

Step 1: Read the Source Report

Use the Read tool to load the full content of the report file. Supported formats:

• Markdown (.md): Read directly.
• PDF (.pdf): Read using the PDF reader (specify pages for large documents).

Parse the document structure: title, abstract/summary, section headings, body content, tables, statistics, quotes, conclusions, and recommendations.

Step 2: Load the Presentation Config

Read the presentation config JSON to understand available slide types, prompt prefixes, and brand identity:

$BASE_DIR/../nano-banana/presets/presentations/{preset}.json

Available presets:

If the user does not specify a preset, default to consulting.

What to extract from the config:

• global_prompt_prefix -- prepended to every slide prompt by nano-banana
• slide_types -- per-type prompt_prefix values (also prepended automatically)
• default_aspect, default_size, default_model -- inherited defaults

You do NOT need to include the global or type prompt prefixes in your slide fields. Nano-banana assembles the full prompt automatically from the structured fields.

Step 3: Decompose the Report into Slides

Map report sections to slide types using these rules:

Slide count guidance:

• If --slides <count> is specified, target that number.
• If not specified, use the report length as a guide: roughly 1 slide per major point, typically 6-12 slides for a standard report.
• Every deck must have exactly one title slide (first) and one closing slide (last).
• Use section-divider slides to create visual breaks between major topics; do not use them for every minor heading.

Step 5: Write Structured Slide Fields

For each slide, write four structured fields: heading, visual, labels, and text_panel. Follow these authoring rules strictly.

Rule 1: Every slide starts with a physical metaphor. Before writing any field, answer: "What concrete object or scene would a viewer see?" The visual field must describe something you could build with your hands or photograph. Not "two columns comparing X and Y" but "a 1D number line dissolving into a 3D coordinate space."

Rule 2: The visual field is the hero -- 60% of the slide area. Describe the object, its materials, its spatial arrangement. Use engineering/architectural vocabulary: isometric, cross-section, exploded view, dimension lines, blueprint grid. ~30 words max.

Rule 3: Labels are short annotations ON the visual. Each label is 1-4 words. They go on or next to the visual element they describe. Think engineering drawing callouts. Array of strings, not sentences.

Rule 4: text_panel is max 2 sentences (~40 words). This is the "so what" at the bottom. If you can't say it in 2 sentences, the slide is trying to do too much -- split it.

Rule 5: heading is the claim, not the topic (~8 words). Not "Market Analysis" but "Transportation Leads at $4.2B." The heading states what the viewer should conclude.

Rule 6: One metaphor per slide, one insight per slide. If you're describing two different visuals, it's two slides.

Metaphor mapping table:

Good example (structured):

{
  "slide_number": 3,
  "slide_type": "content",
  "heading": "Deconstructing Human Impact",
  "visual": "four isometric 3D blocks in a row connected by large multiplication signs: blue block (Direct Effect), copper block (Leverage Effect), green block (Time Effect), equals dark block (Organizational Output), with engineering dimension arrows",
  "labels": ["Direct Effect", "Leverage Effect", "Time Effect", "Organizational Output"],
  "text_panel": "Employee impact is not merely additive. It is additive (Ai, Di), multiplicative (Li), and exponential (Ci, Fi).",
  "style_overrides": {}
}

Bad example (old freeform -- DO NOT USE):

{
  "slide_number": 3,
  "slide_type": "content",
  "layout": "split-horizontal",
  "prompt": "Top half heading 'What Conventional Labels Miss'. Two columns: Left column labeled 'Standard View' showing linear spectrum from Subtractor to Adder to Multiplier..."
}

Step 6: Write the Deck Spec JSON

Write the complete JSON file to:

output/nano-banana/decks/{auto_name}.json

Auto-naming convention: YYYY-MM-DD_{title_slug}.json where title_slug is the deck title in lowercase with spaces replaced by underscores, truncated to 40 characters. Example: 2026-02-27_infrastructure_ai_market_analysis.json

Step 7: Report and Offer Generation

After writing the deck spec:

1. Report the absolute file path of the deck spec.
2. Summarize the deck: total slides, breakdown by type, preset used.
3. Offer to run nano-banana to generate the slide images:

python3 $BASE_DIR/../nano-banana/scripts/nano_banana.py --deck path/to/deck_spec.json

If the user wants to regenerate specific slides (e.g., after prompt edits), use --force:

python3 $BASE_DIR/../nano-banana/scripts/nano_banana.py --deck path/to/deck_spec.json --force

Deck Spec JSON Schema

The deck spec is the contract between this skill and nano-banana's --deck mode. Every deck spec must conform to this structure:

{
  "title": "string",
  "presentation_config": "string -- consulting | workshop | pitch | creative | notebooklm",
  "total_slides": "integer",
  "output_dir": "string",
  "slides": [
    {
      "slide_number": "integer -- 1-indexed",
      "slide_type": "string -- title | section-divider | content | data-chart | quote | closing",
      "heading": "string -- Bold title text, ~8 words, states the claim",
      "visual": "string -- Dominant physical metaphor, ~30 words, concrete object/scene",
      "labels": "array of strings -- Short annotations on the visual, 1-4 words each. Optional.",
      "text_panel": "string -- Explanatory text, max 2 sentences. Optional.",
      "reference_image": "string -- Path to style reference image. Optional, null if unused.",
      "style_overrides": "object -- Optional: { aspect, size, model }"
    }
  ]
}

Required fields: title, presentation_config, total_slides, output_dir, slides Required per-slide fields: slide_number, slide_type, heading, visual Optional per-slide fields: labels (defaults to []), text_panel (defaults to ""), reference_image (defaults to null), style_overrides (defaults to {})

Backward compatibility: The old prompt and layout fields still work but should NOT be used in new deck specs.

output_dir convention: Use output/nano-banana/decks/YYYY-MM-DD_{title_slug} for relative paths. Nano-banana resolves relative paths from the current working directory (project root).

Slide Types Reference

Example Deck Spec

A 6-slide consulting deck generated from a hypothetical market analysis report:

{
  "title": "Canadian Infrastructure AI Opportunity",
  "presentation_config": "consulting",
  "total_slides": 6,
  "output_dir": "output/nano-banana/decks/2026-02-27_canadian_infrastructure_ai",
  "slides": [
    {
      "slide_number": 1,
      "slide_type": "title",
      "heading": "Canadian Infrastructure AI Opportunity",
      "visual": "subtle circuit-board pattern fading into cityscape silhouette with blueprint grid lines",
      "labels": [],
      "text_panel": "Market Analysis Q1 2026 | Tenebrus Capital",
      "style_overrides": {}
    },
    {
      "slide_number": 2,
      "slide_type": "section-divider",
      "heading": "Market Landscape",
      "visual": "abstract visualization of interconnected infrastructure nodes -- roads, bridges, power lines rendered as glowing isometric network graph",
      "labels": [],
      "text_panel": "Understanding the $4.2B Canadian infrastructure AI market",
      "style_overrides": {}
    },
    {
      "slide_number": 3,
      "slide_type": "data-chart",
      "heading": "Transportation Leads at $1.8B",
      "visual": "bar chart showing four vertical bars of decreasing height, with the tallest bar highlighted in coral accent",
      "labels": ["Transportation $1.8B", "Energy $1.4B", "Water $0.6B", "Telecom $0.4B"],
      "text_panel": "42% CAGR since 2023. Transportation sector alone exceeds combined water and telecom spending.",
      "style_overrides": {}
    },
    {
      "slide_number": 4,
      "slide_type": "content",
      "heading": "Three Forces Driving Convergence",
      "visual": "three overlapping translucent circles forming a Venn diagram, each containing a distinct isometric icon: government building, digital twin cube, AI chip",
      "labels": ["Federal $14B bill", "Municipal digital twin mandates", "60% private AI adoption"],
      "text_panel": "Federal infrastructure spending, municipal digital twin mandates, and private sector AI adoption are converging to create a structural opportunity.",
      "style_overrides": {}
    },
    {
      "slide_number": 5,
      "slide_type": "quote",
      "heading": "By 2028, 75% of Canadian infrastructure inspections will involve AI-assisted analysis",
      "visual": "subtle geometric pattern of intersecting lines forming an abstract bridge structure",
      "labels": [],
      "text_panel": "-- Canadian Infrastructure Report Card 2025",
      "style_overrides": {}
    },
    {
      "slide_number": 6,
      "slide_type": "closing",
      "heading": "Recommendations & Next Steps",
      "visual": "three ascending platforms connected by arrows, each containing an isometric icon: target, handshake, calendar",
      "labels": ["Target transportation vertical", "Build municipal pilot", "Secure partnerships by Q3"],
      "text_panel": "Contact: reggie@tenebruscapital.com | Tenebrus Capital Corp",
      "style_overrides": {}
    }
  ]
}

Argument Handling

Error Handling

• If the report file does not exist or cannot be read, inform the user and stop.
• If the requested preset does not have a matching config file, list available presets and ask the user to choose.
• If the report is too short to fill the requested slide count, inform the user and suggest a lower count.
• If the report has no clear structure (no headings, no sections), do your best to decompose by paragraph and inform the user that results may vary.