How the GPT‑Image2 Agent Skill Processes User Requests in freestylefly/awesome‑gpt‑image‑2

The agent skill processes user requests through a deterministic, rule‑based pipeline: language detection, target output classification, template matching against a curated style library, optional candidate presentation, and structured prompt assembly with copy‑ready output.

The GPT‑Image2 style‑library skill transforms vague user intent into production‑ready image generation prompts according to the open‑source repository freestylefly/awesome-gpt-image-2. This agentic skill combines a curated dataset of reverse‑engineered GPT‑Image2 cases with explicit workflow rules defined in SKILL.md. Understanding how the agent skill processes user requests helps developers extend the system or debug prompt generation failures.

Agent Registration and Configuration

The skill registers itself in the OpenAI agent ecosystem through agents/openai.yaml. This YAML file supplies metadata that agent frameworks consume:

• display_name – human‑readable skill label
• short_description – quick capability summary
• default_prompt – fallback instruction when user input is ambiguous

# From agents/skills/gpt-image-2-style-library/agents/openai.yaml

display_name: "GPT-Image2 Style Library"
short_description: "Generate optimized prompts for GPT-Image2 using reverse-engineered style templates"
default_prompt: "Analyze the user's image request and match it to the best-fitting template from the style library."

The default_prompt value seeds the agent's reasoning before any user‑specific processing begins.

Step‑by‑Step Request Processing Workflow

Step 1: Language Detection

The skill commits to response language parity. It analyzes the user's input to determine Chinese or English, then generates all subsequent output—including prompts, rationales, and constraints—in that language. This rule appears verbatim in SKILL.md: "Detect the user's language and answer in that language."

Step 2: Target Output Identification

The skill classifies the user's desired deliverable into a Category from the style library. Valid categories include UI design, poster, product photography, illustration, scenes, and storytelling. This classification determines which subset of templates the matching algorithm prioritizes.

From SKILL.md: "Identify the user's target output … This determines which Category from the style library to prioritize."

Step 3: Template Matching Hierarchy

The skill applies fixed‑priority selection rules from style-library.md:

1. Category — explicit product type or domain
2. Visual style tag — descriptive aesthetic keywords
3. Scene tag — context, environment, or narrative setting
4. Closest example cases — nearest neighbors from >500 reverse‑engineered samples

This ordering guarantees deterministic behavior: a "futuristic product render" request always matches product category templates before considering scene‑based alternatives.

Step 4: Candidate Presentation (Optional Fork)

When multiple templates score equally, the skill pauses for user disambiguation. It returns 2–3 candidates with concise rationales, then awaits selection. This prevents arbitrary prompt construction when user intent remains ambiguous.

Workflow rule: "If several are plausible, present 2‑3 options … and ask the user to choose."

Step 5: Prompt Construction with Structured Blocks

Upon template selection, the skill assembles the final prompt from six explicit blocks defined in SKILL.md:

Block	Purpose
Subject & task	Core entity and action
Composition & layout	Framing, perspective, spatial arrangement
Visual style & materials	Aesthetic direction, textures, rendering approach
Text & label requirements	Copy, typography, readability constraints
Aspect ratio & output format	Technical specifications
Constraints & negative details	Exclusions, avoided elements, quality guardrails

This block structure ensures prompts remain complete yet modular—developers can override single blocks without rewriting entire prompts.

Step 6: Output Formatting with Language Defaults

The skill always emits content in this order:

1. Copy‑ready prompt first — immediately usable in GPT‑Image2
2. Concrete constraints — exact text strings, aspect ratios, readability rules
3. Modification offer — explicit yes/no gate for refinement

For Chinese requests, the skill applies Chinese‑specific defaults (full‑width punctuation, Simplified Chinese technical terms, local aspect‑ratio conventions). For English requests, it defaults to ASCII punctuation and imperial/metric agnostic specifications.

Code Example: End‑to‑End Interaction

Below is a representative user interaction demonstrating all processing stages.

User input:

用 gpt-image-2-style-library 技能生成城市生命系统图谱

Skill output:

### Selected Template

**Template:** City life system map  
**ID:** scene-storytelling  
**Category:** Scenes & Storytelling  
**Styles:** Scenes, Illustration  
**Scenes:** Story, Social  

### Prompt (copy‑ready)

/imagine prompt: 城市生命系统图谱, 俯视视角, 细节丰富的街道与建筑, 霓虹灯光, 未来城市氛围, 4:3, 高分辨率, --ar 4:3 --v 5

### Constraints

- Text: 必须包含 "城市" 与 "系统图谱"。  
- Aspect Ratio: 4:3（宽高比）。  
- Layout: 采用俯视构图，主体位于画面中心。  
- Negative Details: 避免过度写实的阴影、无关的文字标签。

Would you like to modify any part of the prompt? (yes/no)

The output demonstrates: Chinese language parity, category‑driven template selection (Scenes & Storytelling), visual/style tag matching (Illustration, Story), and full block‑structured assembly with copy‑first formatting.

Maintenance and Synchronization Pipeline

The skill remains synchronized with its data source through automated regeneration. The scripts/generate-style-skill.mjs script consumes data/style-library.json and produces references/style-library.md:

# Regenerate reference from source data

npm run generate:skill

This script executes three operations:

1. Parses data/style-library.json containing >500 GPT‑Image2 cases
2. Rebuilds references/style-library.md with updated categories, tags, and constraints
3. Ensures SKILL.md references remain valid

Developers install the updated skill into local Codex environments via:

npm run install:skill

The package.json in the skill folder defines both scripts, making the maintenance loop reproducible.

Key Source Files and Their Roles

File	Responsibility
agents/skills/gpt-image-2-style-library/SKILL.md	Canonical workflow, selection rules, output formatting, and maintenance instructions
agents/skills/gpt-image-2-style-library/agents/openai.yaml	Agent registration metadata for OpenAI‑compatible frameworks
agents/skills/gpt-image-2-style-library/references/style-library.md	Human‑readable template reference used during matching
scripts/generate-style-skill.mjs	Build script syncing JSON data to markdown reference
data/style-library.json	Source‑of‑truth dataset with reverse‑engineered cases
agents/skills/gpt-image-2-style-library/package.json	NPM script definitions for generate:skill and install:skill

Summary

• The agent skill processes user requests through six deterministic stages from language detection to copy‑ready output.
• Template matching follows strict priority: Category → Visual style → Scene → Example cases.
• All behavior is explicitly documented in SKILL.md and data‑driven via style-library.json.
• Language parity is enforced throughout—Chinese inputs yield Chinese prompts, constraints, and rationales.
• Automated regeneration via generate-style-skill.mjs keeps the skill synchronized with its underlying dataset.

Frequently Asked Questions

How does the skill handle ambiguous user requests?

The skill pauses and presents 2–3 candidate templates with brief rationales, then requires explicit user selection. This prevents arbitrary prompt construction when target output classification yields multiple valid categories or style combinations.

Can I modify the template matching priority?

The matching order—Category, then Visual style, then Scene, then Example cases—is hardcoded in SKILL.md and style-library.md. To change this behavior, fork the repository and edit these files, then regenerate the skill with npm run generate:skill.

What happens when the underlying style data updates?

Running npm run generate:skill executes scripts/generate-style-skill.mjs, which rebuilds references/style-library.md from data/style-library.json. The skill immediately uses updated templates without code changes, then reinstall via npm run install:skill to propagate to your agent environment.

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