Complete Guide to gpt-image-2 Prompt Template Categories in Garden Skills

The gpt-image-2 skill organizes its prompt templates into 18 distinct categories—ranging from UI Mockups and Product Visuals to Academic Figures and Technical Diagrams—each stored as subdirectories under the references/ folder and enumerated in SKILL.md.

The gpt-image-2 skill in the ConardLi/garden-skills repository provides a structured framework for AI image generation through categorized prompt templates. Understanding these gpt-image-2 prompt template categories is essential for developers who want to programmatically select the right visual context for generation tasks or extend the skill with new template types.

How Template Categories Are Defined

The category system is declared in skills/gpt-image-2/SKILL.md, which serves as the skill's central manifest. According to the source code, the runtime uses this manifest to determine which subdirectory under references/ to load for a specific task, avoiding unnecessary file system scans across the entire template tree.

Each category maps to a dedicated subdirectory following the pattern references/<category-folder>/. This hierarchical structure allows the skill to load only the most relevant template files required for the current execution mode.

The 18 gpt-image-2 Prompt Template Categories

The skill defines exactly 18 top-level categories, each targeting a specific visual use case:

• UI Mockups (references/ui-mockups/): Templates for interface designs including live-commerce screens and chat UIs.
• Product Visuals (references/product-visuals/): Product-centric renders such as exploded-view posters and white-background shots.
• Maps (references/maps/): Geographic and route visualizations including food maps and travel itineraries.
• Slides & Visual Docs (references/slides-and-visual-docs/): One-page explainer slides and dense policy documents.
• Poster & Campaigns (references/poster-and-campaigns/): Branding-focused marketing materials and campaign layouts.
• Portraits & Characters (references/portraits-and-characters/): Professional headshots, founder portraits, and character designs.
• Scenes & Illustrations (references/scenes-and-illustrations/): General illustration contexts including healing scenes and concept art.
• Editing Workflows (references/editing-workflows/): Image manipulation templates for background replacement and object removal.
• Avatars & Profile (references/avatars-and-profile/): Profile image templates including style-transfer selfies and 3D icons.
• Storyboards & Sequences (references/storyboards-and-sequences/): Sequential storytelling formats like comics and cinematic storyboards.
• Grids & Collages (references/grids-and-collages/): Multi-panel layouts including banner grids and look-book arrangements.
• Branding & Packaging (references/branding-and-packaging/): Brand identity systems and packaging design templates.
• Typography & Text Layout (references/typography-and-text-layout/): Text-centric posters and bilingual typography layouts.
• Assets & Props (references/assets-and-props/): Collections of isolated assets such as icon sets or game screenshot mock-ups.
• Academic Figures (references/academic-figures/): Scientific visualizations including method pipelines and neural-network architectures.
• Infographics (references/infographics/): Data-driven visualizations like KPI dashboards and step-by-step diagrams.
• Technical Diagrams (references/technical-diagrams/): Engineering schematics including system architecture and ER diagrams.

Directory Structure and Template Files

Each category subdirectory contains individual markdown files representing specific prompt templates. As documented in references/prompt-writing.md, every template follows a standardized "single-template file" structure that includes parameter definitions and prompt formatting rules.

For example, the UI Mockups category contains files like live-commerce-ui.md, which can be loaded directly by the runtime scripts to generate specialized prompts for e-commerce interfaces.

Programmatic Access to Template Categories

Developers can extract category metadata directly from SKILL.md using standard file system operations. The manifest uses a consistent naming convention where each category line matches the pattern ### N. Category Name (references/folder-name/).

To list all 18 categories programmatically:

const fs = require('fs');
const path = require('path');

const skillRoot = path.resolve(__dirname, '../../skills/gpt-image-2');
const skillMd = fs.readFileSync(path.join(skillRoot, 'SKILL.md'), 'utf8');

// Extract lines that match the pattern "### N. <Name> (`references/...`)"

const catLines = skillMd
  .split('\n')
  .filter(l => /^\s*###\s+\d+\./.test(l));

const categories = catLines.map(l => {
  const [, name, folder] = l.match(/###\s+\d+\.\s+(.+)\s+\(`([^`]+)`\)/);
  return { name: name.trim(), folder };
});

console.log('gpt-image-2 template categories:');
categories.forEach(c => console.log(`- ${c.name} → ${c.folder}`));

To load a specific template file for use in generation workflows:

const templatePath = path.join(
  skillRoot,
  'references',
  'ui-mockups',
  'live-commerce-ui.md'
);
const template = fs.readFileSync(templatePath, 'utf8');
console.log(template);

Key Implementation Files

The template category system relies on several critical files within the skills/gpt-image-2/ directory:

• SKILL.md: The central manifest enumerating all 18 categories and defining runtime execution modes.
• references/prompt-writing.md: Methodology document specifying folder-tree conventions and template formatting standards.
• references/<category>/<template>.md: Individual prompt template files (e.g., ui-mockups/live-commerce-ui.md).
• scripts/check-mode.js: Runtime detection script that determines which execution mode (A, B, or C) to invoke based on input parameters.
• scripts/generate.js and scripts/edit.js: API wrapper scripts that utilize categorized templates when calling OpenAI-compatible image generation or editing endpoints in Mode A.

Summary

• The gpt-image-2 skill organizes templates into 18 distinct categories spanning UI design, academic visualization, and technical documentation.
• Category definitions live in skills/gpt-image-2/SKILL.md, with each category mapped to a dedicated subdirectory under references/.
• Individual templates are stored as markdown files within their respective category folders, following standardized formatting rules defined in prompt-writing.md.
• The runtime uses scripts/check-mode.js to determine execution context, then loads only the relevant category templates to optimize performance.

Frequently Asked Questions

How many prompt template categories does gpt-image-2 support?

The skill supports exactly 18 categories, ranging from UI Mockups and Product Visuals to Academic Figures and Technical Diagrams. This number is hard-coded in the SKILL.md manifest and represents the complete taxonomy of visual generation tasks the skill handles.

Where are the prompt template files physically stored?

Template files reside in the references/ directory within the skill root (skills/gpt-image-2/references/). Each category has its own subdirectory (e.g., references/infographics/), containing individual markdown files for specific template variants. The folder structure mirrors the category enumeration in SKILL.md.

How does the skill determine which template category to use?

The runtime evaluates input parameters through scripts/check-mode.js to determine the execution mode (A, B, or C). Based on the detected mode and user intent, the system loads the appropriate category subdirectory from references/, selecting the specific template file that matches the requested visual output type.

Can I add custom template categories to gpt-image-2?

While the source code defines 18 standard categories, extending the system requires modifying SKILL.md to register the new category and creating the corresponding subdirectory under references/. New templates must follow the "single-template file" structure documented in references/prompt-writing.md to ensure compatibility with the generation scripts (generate.js and edit.js).

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}