agent-image

name	agent-image
description	Local-first image generation helper. Use when you need to create or batch-generate images (portraits, mock art, docs images) as part of implementing a feature/spec, or when the user asks for images.
license	MIT
compatibility	Requires local Ollama and/or local ComfyUI scripts in this repo.

Agent Image Generation

Generate images locally and reproducibly using the repository pipelines.

When To Use

Use this skill when:

• A spec/task requires images (e.g., kin portraits, NPC portraits, docs screenshots/mock art).
• The user asks you to create images or a batch of images.
• You need deterministic filenames/locations for generated images.

Do NOT use this skill to call remote image APIs unless the user explicitly requests that.

Backend Selection

Default policy:

• Prefer Ollama for quick, single-shot or simple batches.
• Prefer ComfyUI for advanced workflows, specific nodes/models, or higher fidelity/controllability.

If one backend is unavailable, fall back to the other when appropriate.

Output Conventions

Default: choose an output directory appropriate to the current task context.

Common conventions (use when applicable):

• Character/NPC portraits: assets/portraits/npcs/<slug>.png
• Kin portraits (if generating a full set): assets/portraits/kins/<slug>.png
• Docs/supporting images: docs/images/<slug>.png

If the right destination is ambiguous, ask the user (or follow existing nearby patterns in the repo).

Slug rules: lowercase; spaces -> _; strip non-alphanumerics; collapse _.

Overwrite policy: do NOT overwrite existing images unless the user requests overwrite.

Guardrails

• Do not commit generated images unless the user explicitly asks.
• Never commit model weights or local generator state (tools/ComfyUI, Ollama model store).
• Keep generation local-first. If a remote provider is requested, ask for confirmation and clarify licensing.

Ollama Workflow (Quick)

Prereqs:

• Ollama installed (macOS-first for image generation).

Commands:

# Install/verify + pull default model
pnpm ollama:setup

# Verify image generation support
pnpm ollama:doctor

# Generate a batch from a job config (example)
python3 scripts/ollama/ollama.py generate --job <job.json>

Ad-hoc single image:

python3 scripts/ollama/ollama.py generate \
  --name "mallsing" \
  --prompt "<your prompt>" \
  --out assets/portraits/npcs

Model selection:

python3 scripts/ollama/ollama.py generate \
  --model x/flux2-klein \
  --job <job.json>

ComfyUI Workflow (Advanced)

Prereqs:

• Local ComfyUI checkout in tools/ComfyUI (gitignored)
• A checkpoint in tools/ComfyUI/models/checkpoints/

Commands:

pnpm comfyui:setup
pnpm comfyui:server
pnpm comfyui:doctor

# Config-driven batch (example)
python3 scripts/comfyui/comfyui.py generate --job <job.yaml>

Examples

Single portrait (Ollama):

• Prompt: "A fantasy character profile portrait in watercolor and ink, side profile, no text"
• Output: assets/portraits/npcs/river_guard.png

Batch generation (Ollama):

• Provide a job config (.json) with items[] or a source{type=..., path=...}.
• Keep configs data-only; avoid embedding scripts.

High fidelity batch (ComfyUI):

• Use ComfyUI when you need a specific checkpoint, LoRA stack, or a custom workflow.