Multi-Agent Image

multi-agent-image is a standalone Hermes skill for image generation workflows.

It is designed for cases where a simple one-line prompt is not enough. Instead of sending raw user input directly to an image model, this skill:

1. analyzes the request,
2. compiles it into a design-aware prompt,
3. generates through gpt-image-2,
4. archives the result,
5. and optionally reuses successful outputs as future style references.

This skill is independent at runtime. The design compiler is built into this repository and does not require an external skill.

When to Use

Use this skill when the user wants one or more of the following:

• Design-oriented poster generation
• Product images or ad visuals
• PPT cover visuals or chapter art
• Infographic-like or teaching/demo visuals
• Style reference reuse from prior generations
• Interactive “show examples first, then generate” flow
• Batch generation for multiple directions or aspect ratios
• Series generation where multiple images should share one visual language

Do not use this skill for:

• pixel-accurate UI recreation
• editable charts
• exact typography output inside the image
• tasks that require vector, HTML, or PPT-native assets rather than raster images

Architecture

User Request
    ↓
[Prompt Engineer]
    ↓
[Style Scout]
    ↓
[Internal Design Compiler]
    ↓
[GPT-Image-2 Generation]
    ↓
[QA + Archive]
    ↓
[Case Library]

Optional layers on top of the main path:

• Interactive reference selection
• Batch generation
• Series generation

Setup

1. Deploy the skill

The skill source lives in:

~/.hermes/skills/multi-agent-image/

Install runtime files into the working directory:

python3 ~/.hermes/skills/multi-agent-image/scripts/install.py

This prepares:

• ~/.hermes/agents/multi-agent-image/output/
• ~/.hermes/agents/multi-agent-image/case_library/
• agent role folders and memory files
• local runtime scripts copied from the skill

2. Install Python dependencies

pip install openai requests

3. Set API key

export OPENAI_API_KEY="sk-..."

This key is used with the apimart-compatible GPT-Image-2 endpoints in this skill.

Core Components

scripts/design_compiler.py

Internal prompt compiler.

Responsibilities:

• detect task type
• choose defaults for aspect and quality
• build design_reasoning
• compress it into compiled_brief
• produce the final generation prompt

This is the core logic that makes the skill independent.

scripts/design_image.py

CLI entrypoint for the internal compiler.

Use it when you want:

• prompt-only output
• a local design compilation test
• direct generation without the full multi-agent workflow

Example:

cd ~/.hermes/agents/multi-agent-image
python3 design_image.py \
  --task poster \
  --brief "AI训练营招生海报，强调速度、增长、实战" \
  --direction balanced \
  --aspect 3:4 \
  --prompt-only

It prints:

• design_reasoning
• compiled_brief
• prompt
• settings

scripts/orchestrator_v2.py

Main workflow entrypoint.

Responsibilities:

• run prompt analysis
• choose task and generation parameters
• optionally select a reference from the case library
• call the internal compiler
• call GPT-Image-2
• archive outputs
• auto-save successful results into the case library

scripts/gpt_image2_generator.py

Low-level GPT-Image-2 client.

Responsibilities:

• submit async generation tasks
• poll task status
• download image results

Use this when you want direct API access without the full workflow.

scripts/case_library.py

Persistent library of past generations.

Responsibilities:

• save outputs by task type
• store metadata and rating
• search by brief, prompt, or tags
• return image paths for reuse as references

scripts/case_selector.py

Interactive helper for Hermes dialogue flows.

Responsibilities:

• render user-facing selection text
• parse replies like 1, n, case_001, or 搜索蓝色

scripts/interactive_run.py

Two-phase dialogue wrapper.

Use it when the workflow needs to ask the user before generating.

scripts/batch_generator_v2.py

Batch generation entrypoint.

Supports:

• same brief, multiple directions
• same brief, multiple aspect ratios
• multiple briefs in one run

scripts/series_generator.py

Style-consistent series generator.

Workflow:

1. generate a master image
2. extract style signals from its compiled brief
3. generate child images that follow the same visual system

templates/linear_batch.py

Editable template for resumable sequential runs.

Useful when you want:

• explicit scene lists
• filesystem-based progress monitoring
• style propagation from the first generated image

Internal Design Compiler

The internal compiler produces three layers:

1. design_reasoning

This captures design intent before generation.

Typical fields:

• task
• communication_goal
• audience
• channel
• visual_system
• hierarchy_strategy
• safe_zone_strategy
• lighting_strategy
• palette_strategy
• anti_filler_rules
• anti_slop_rules

2. compiled_brief

This is a compressed design brief for generation.

It includes:

• what the image is for
• what should dominate visually
• what space should remain available
• what to avoid

3. prompt

Final model-facing prompt used for GPT-Image-2.

The prompt is generated from design logic, not just from a list of style keywords.

Supported Tasks

The built-in compiler understands these task classes:

• poster
• product
• ppt
• infographic
• teaching
• auto

Default aspect assumptions:

• poster → 3:4
• product → 1:1
• ppt → 16:9
• infographic → 4:3
• teaching → 16:9

Direction modes:

• conservative
• balanced
• bold

Quality modes:

• draft
• final
• premium

Current generation channel:

• gpt-image-2

Usage

Quick start

cd ~/.hermes/agents/multi-agent-image
python3 quick_start.py "AI训练营招生海报，强调速度、增长、实战"

Prompt-only compilation

cd ~/.hermes/agents/multi-agent-image
python3 design_image.py \
  --task product \
  --brief "高端陶瓷咖啡杯电商首图，温暖晨光，突出釉面质感" \
  --prompt-only

Full orchestrated generation

from orchestrator_v2 import run

run("AI训练营招生海报，强调速度增长实战")

Force task and visual settings

from orchestrator_v2 import run

run(
    "高端咖啡杯商品图",
    task="product",
    direction="balanced",
    aspect="1:1",
    quality="final",
    use_reference=False,
)

Interactive Workflow

Use the two-phase pattern when Hermes should ask before generating.

Phase 1: prepare text for the user

from interactive_run import prepare

text = prepare("帮我做张 AI 训练营海报", task="poster")
print(text)

Phase 2: execute after the user chooses

from interactive_run import execute

result = execute("帮我做张 AI 训练营海报", user_choice="1", task="poster")

Supported reply patterns:

• 1, 2, 3
• n
• y
• case_001
• 搜索蓝色

Batch Generation

Same brief, multiple directions

from batch_generator_v2 import batch_styles

batch_styles("AI训练营海报", task="poster")

Same brief, multiple aspect ratios

from batch_generator_v2 import batch_aspects

batch_aspects("AI训练营海报", task="poster", aspects=["1:1", "16:9", "9:16"])

Multiple briefs

from batch_generator_v2 import batch_briefs

batch_briefs(["海报A", "海报B", "海报C"], task="poster")

Series Generation

Use this when several outputs should feel like the same campaign or product family.

from series_generator import SeriesGenerator

sg = SeriesGenerator()
sg.create_series(
    master_brief="AI训练营系列视觉，科技蓝，专业商务感",
    items=[
        {"name": "主海报", "brief": "AI训练营招生主海报", "aspect": "3:4"},
        {"name": "Banner", "brief": "官网 Banner", "aspect": "16:9"},
        {"name": "朋友圈", "brief": "朋友圈推广方形图", "aspect": "1:1"},
    ],
    task="poster",
    direction="balanced",
)

Case Library

Case library directory:

~/.hermes/agents/multi-agent-image/case_library/

Output directory:

~/.hermes/agents/multi-agent-image/output/

Typical case structure:

case_library/
├── poster/
│   └── case_001_example/
│       ├── image.png
│       └── metadata.json

Typical metadata fields:

• case_id
• task
• brief
• prompt
• params
• tags
• rating

Validation Guidance

Before generating at scale, test prompt quality first:

python3 design_image.py \
  --task poster \
  --brief "AI训练营招生海报，强调速度、增长、实战" \
  --direction balanced \
  --aspect 3:4 \
  --prompt-only

What to check:

• Does design_reasoning state a clear communication goal?
• Is there an explicit safe zone?
• Is hierarchy obvious?
• Do anti_slop_rules remove HUD overlays, fog, and generic clutter?
• Does the prompt describe a single strong visual idea rather than a pile of elements?

Current Limits

• Current image provider is centered on gpt-image-2
• QA scoring is intentionally lightweight
• Series generation is heavier than one-off generation
• The skill is optimized for raster outputs, not editable assets
• Some reference documents remain longer than necessary, but the main runtime path is consistent

Version History

• v1.0.0 Initial multi-agent workflow for GPT-Image-2 generation
• v2.0.0 Added case library, interactive reference selection, and image-to-image style reuse
• v2.1.0 Added stronger download retry logic, batch workflows, and series generation
• v2.2.0 Packaged as a reusable Hermes skill with install script and runtime layout
• v3.0.0 Internalized the design compiler and removed external runtime dependency