GFX-301c · Module 1

Template-Based Generation

3 min read

For recurring asset types — social cards, blog headers, course thumbnails, OG images — the prompt should be a template with variable slots, not a unique composition every time. The template locks down the style-carrying elements (palette, composition grid, typography, effects) and exposes only the content-carrying elements (headline text, subject imagery, data values) as variables.

Template anatomy: "[LOCKED: background #000000, layout: centered-hero, font: Exo 2 Bold, glow: cyan text-shadow] [VARIABLE: headline = {headline_text}] [VARIABLE: subject = {subject_description}] [LOCKED: aspect ratio 1200x630, no photographic elements]." The generator receives the template plus a CSV of variable values. Row 1: headline = "Agent Team Architecture," subject = "abstract network diagram." Row 2: headline = "Quality Loop Engineering," subject = "circular feedback arrows." Each row produces one asset from the same template.

Templates reduce generation cost by 60-70% — the research and prompt-architecture stages run once, not per asset. The style specification is embedded in the template. The generator only needs to execute the variable substitution and generate.

Do This

  • Build templates for every recurring asset type — social, headers, thumbnails, OG images
  • Lock style elements, expose only content variables — consistency is built into the template
  • Feed templates with CSV batches — one template, many outputs, identical style

Avoid This

  • Write unique prompts for recurring assets — that is artisan workflow at factory volume
  • Leave style elements as variables — the model will reinterpret them differently each time
  • Generate one at a time from a template — batch the variable inputs and parallelize