How to Automate Prompt Generation with the GPT-Image2 Agent Skill

The GPT-Image2 agent skill automates prompt generation by transforming a curated style library into structured prompt templates through a validation and rendering pipeline.

The freestylefly/awesome-gpt-image-2 repository provides a complete automation framework that converts static style definitions into dynamic, context-aware image prompts. By combining a structured JSON knowledge base with human-readable documentation and a Node.js generation script, the system enables the GPT-Image2 agent to automatically select and compose optimal prompts based on user intent without manual intervention.

Understanding the GPT-Image2 Agent Skill Architecture

The automation workflow relies on three integrated components that work together to produce the runtime reference consumed by the agent skill.

The Style Library Data Layer (data/style-library.json)

At the core of the system sits data/style-library.json, which stores structured data for templates, categories, styles, and scenes. Each template entry contains an ID, multilingual titles, category classification, style tags, scene identifiers, cover image paths, and guidance metadata.

The JSON structure enables the agent to perform semantic matching between user requests and available prompt templates. When you modify this file to add new templates or update existing styles, the automation pipeline detects these changes during the next generation cycle.

Human-Readable Documentation (docs/templates.md)

Parallel to the JSON data, docs/templates.md provides comprehensive markdown documentation for every template defined in the library. Each template section includes an HTML anchor (<a name="…"></a>) that corresponds to the anchor field in the JSON entries.

The generation script cross-references these anchors to ensure every JSON entry points to a valid documentation block. This validation step prevents broken references and guarantees that the agent skill always has access to complete template descriptions during prompt selection.

The Generation Script (scripts/generate-style-skill.mjs)

The scripts/generate-style-skill.mjs Node.js script (requires Node ≥ 18) serves as the automation engine. It performs three critical validation steps before rendering the final reference:

• JSON Schema Validation: Ensures all required fields are present and ID uniqueness is maintained
• Asset Verification: Confirms that every cover image referenced in the JSON exists in the repository
• Anchor Mapping: Validates that JSON anchors correspond to actual documentation sections in docs/templates.md

Upon successful validation, the script renders agents/skills/gpt-image-2-style-library/references/style-library.md, the comprehensive reference file that the GPT-Image2 skill reads at runtime.

How the Selection Rules Automate Prompt Building

When the GPT-Image2 agent skill processes a user request, it applies a three-tier matching algorithm defined in the generated reference:

Step 1 – Category Match: Product type keywords (e.g., poster, UI, icon) map to a template's category field. The agent filters the library to templates matching the requested content type.

Step 2 – Style Match: Visual adjectives (e.g., realistic, illustration, minimalist) map to the template's styles array. This narrows the selection to templates capable of producing the desired aesthetic.

Step 3 – Scene Match: Context keywords (e.g., commerce, travel, education) map to the scenes tags, ensuring the template supports the intended use case.

If the user request remains ambiguous after these steps, the skill presents 2–3 strong template directions and requests clarification before composing the final prompt.

Setting Up the Automation Pipeline

To initialize or update the prompt generation system, execute the generation script after ensuring your style library and documentation are current:

node scripts/generate-style-skill.mjs

The script outputs confirmation when successful:

Generated GPT-Image2 style skill reference at agents/skills/gpt-image-2-style-library/references/style-library.md

After updating data/style-library.json or docs/templates.md—whether adding new templates, revising guidance, or updating cover images—simply rerun the script. The GPT-Image2 skill automatically consumes the refreshed reference without requiring code changes:

git pull  # Retrieve latest library updates

node scripts/generate-style-skill.mjs

Example Template Structure and Usage

A typical entry in data/style-library.json follows this structured format:

{
  "id": "t01",
  "title": { "en": "Modern Poster", "zh": "现代海报" },
  "category": "poster",
  "styles": ["realistic", "vibrant"],
  "scenes": ["commerce"],
  "cover": "/images/poster-modern.png",
  "anchor": "modern-poster",
  "useWhen": {
    "en": "When a clean, bold poster is needed.",
    "zh": "需要干净、醒目的海报时。"
  },
  "guidance": {
    "en": ["Use bold typography", "Keep background simple"],
    "zh": ["使用粗体排版", "保持背景简洁"]
  }
}

At runtime, the agent skill consumes the generated reference to automate prompt construction:

import { loadReference } from 'agents/skills/gpt-image-2-style-library';

const ref = loadReference();                      // reads the generated markdown
const prompt = ref.selectPrompt(userMessage);      // applies the selection rules
await sendToGPTImage2(prompt);

The automated output always includes the chosen template name, a copy-paste-ready GPT-Image2 prompt string, and concise constraints specifying aspect ratio, layout requirements, and negative prompts.

Summary

• The GPT-Image2 agent skill automates prompt generation through a three-part pipeline: JSON data in data/style-library.json, documentation in docs/templates.md, and the Node.js generator at scripts/generate-style-skill.mjs.
• Selection Rules automate template matching through category, style, and scene keyword analysis.
• Running node scripts/generate-style-skill.mjs validates data integrity and produces the runtime reference at agents/skills/gpt-image-2-style-library/references/style-library.md.
• The system supports continuous expansion—new templates require only JSON updates and documentation anchors before regeneration.
• According to the freestylefly/awesome-gpt-image-2 source code, the agent handles ambiguity by suggesting 2–3 template options when automatic selection criteria overlap.

Frequently Asked Questions

What Node.js version is required to run the generation script?

The scripts/generate-style-skill.mjs script requires Node.js 18 or higher. Earlier versions may lack support for the modern JavaScript features and APIs used in the file validation and markdown rendering logic.

How does the agent skill handle ambiguous user requests?

When the Selection Rules cannot determine a single best template, the GPT-Image2 skill identifies 2–3 strong template candidates based on partial matches. It presents these options to the user with brief descriptions from the useWhen fields, requesting clarification before generating the final prompt string.

Can I add custom templates without modifying the agent code?

Yes. Add your template to data/style-library.json with a unique ID, create the corresponding anchor section in docs/templates.md, place the cover image in the specified path, and run node scripts/generate-style-skill.mjs. The agent skill automatically incorporates the new template into its selection logic without requiring changes to the runtime code.

What validation does the generation script perform?

The script validates JSON schema compliance, ensures ID uniqueness across all templates, verifies that referenced cover images exist in the repository, and confirms that every JSON anchor field corresponds to an actual HTML anchor in docs/templates.md. These checks prevent runtime errors and broken references in the generated skill reference file.

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