/generate-image [prompt]

Generate AI images using Gemini with interactive option selection.

Usage

/generate-image
/generate-image "A diagram showing the treaty adoption process"

Step 1: Gather Context

Read source files for available options:

• •scripts/images/generate-image.ts → ImageTypes, AspectRatios
• •scripts/lib/image-prompts.ts → VisualStyles

Determine target QMD file (priority order):

1. •Explicit file reference (e.g., @knowledge/appendix/file.qmd)
2. •Current conversation context
3. •Git status: git diff --name-only HEAD~3 | grep "\.qmd$" | head -1

Derive output folder from QMD filename:

• •knowledge/problem/regulatory-capture.qmd → assets/images/regulatory-capture/
• •knowledge/appendix/invisible-graveyard.qmd → assets/images/invisible-graveyard/
• •No target file → assets/images/generated/ (fallback)

For charts, get actual numbers:

.venv/Scripts/python.exe scripts/preview-qmd-with-variables.py <file.qmd> --numbers-only

IMPORTANT: Include confidence intervals when available. The preview output shows values with 95% CIs in format X (95% CI: Y-Z). Always include these uncertainty ranges in chart prompts - request error bars or uncertainty visualization.

Step 2: Ask User for Options

Use AskUserQuestion. Order options by inferred likelihood, putting recommended choice first with "(Recommended)" suffix.

Inference heuristics:

Aspect ratio hints: QMD chapters → 16:9, Social media → 1:1

Step 3: Build Prompt and Confirm

Extract comprehensive context from the QMD file (with variables resolved). Include enough surrounding text for the image generator to understand the full concept.

Run preview script with broad line range (e.g., 30-50 lines) to capture full context:

.venv/Scripts/python.exe scripts/preview-qmd-with-variables.py <file.qmd> --line-range "X-Y"

Prompt Writing Rules

CRITICAL: Gemini renders most text literally. Avoid meta-labels and editorial commentary.

Write prompts as plain descriptive prose:

• •Describe what you want to see, not metadata about it
• •State facts and numbers directly without labeling them
• •Avoid bullet points, numbered lists, or structured formatting in prompts
• •Don't editorialize ("Key insight", "Important", "Note that")

Do NOT prescribe specific layouts (e.g., "left bar should be X, right bar should be Y"). Provide the content and let the generator be creative.

For charts:

• •Append "Linear scale." to the prompt
• •Include error bars if confidence intervals are available (e.g., "Show error bars: X ranges from Y to Z")

Confirmation (MANDATORY)

NEVER generate without showing the prompt and getting explicit user confirmation. This applies to initial generation AND regeneration/edits.

Present the full prompt in a code block:

Prompt: [full prompt text]
Type: [type]
Aspect: [aspect]
Style: [style]
Alt text: [alt]

Use AskUserQuestion to confirm: "Generate with this prompt?" with options:

• •"Yes, generate"
• •"Edit prompt"
• •"Cancel"

Do NOT proceed to Step 4 until user confirms.

Step 4: Run Generation

After user confirms:

npx tsx scripts/images/generate-image.ts '<prompt>' \
  --type <type> --aspect <aspect> --style <style> --alt '<alt_text>' \
  --output 'assets/images/<qmd-basename>/'

Examples:

• •Target: regulatory-capture.qmd → --output 'assets/images/regulatory-capture/'
• •Target: invisible-graveyard.qmd → --output 'assets/images/invisible-graveyard/'
• •No target → --output 'assets/images/generated/'

IMPORTANT: Use single quotes around prompt and alt text to prevent bash from interpreting $ as variable expansion (e.g., $8 trillion would break with double quotes).

Step 5: Insert at Optimal Location

Insert where the content is discussed, NOT at the end.

1. •Search QMD file for the section discussing the image's subject
2. •Insert image markdown immediately AFTER the relevant paragraph

Use Edit tool with correct relative path from QMD to assets/images/<qmd-basename>/.

Step 6: Report Result

Report: image file path, where inserted, remind user to preview.

Environment

Requires GOOGLE_GENERATIVE_AI_API_KEY environment variable.