Ai Image CreatorCAUTION

Shared starter template configuration and CLAUDE.md memory bank system for Claude Code

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: ai-image-creator
description: Generate, edit-from-reference, or analyze images with AI via OpenRouter (Gemini, GPT Image, Seedream, Qwen, MAI, Grok, FLUX.2, Recraft, Muse, Riverflow; Cloudflare AI Gateway BYOK). Also analyze a video (--analyze-video, read-only — no video generated) into a text description for video prompts. Use when the user asks to generate an image, create a PNG, make an icon, make it transparent, edit with a reference, design a logo/banner, describe/analyze/explain an image ("what's in this image"), or describe/analyze a video ("what happens in this video").
allowed-tools: Bash, Read, Write
compatibility: Requires uv (Python runner) and network access. Environment variables for CF AI Gateway or direct API keys must be configured in shell profile (~/.zshrc on macOS, ~/.bashrc on Linux, or System Environment Variables on Windows).
metadata:
  tags: image-generation, ai, openrouter, cloudflare, gemini, flux2, riverflow, seedream, gpt54, gpt-image, qwen, mai, grok, recraft, muse
---

# AI Image Creator

Generate PNG images via multiple AI models, routed through Cloudflare AI Gateway BYOK or directly via OpenRouter/Google AI Studio.

## Model Selection

When the user mentions a model keyword in their image request, use the corresponding `--model` flag:

| Keyword | Model | Use When User Says | Best For (measured cost · time per image) |
|---------|-------|--------------------|-------------------------------------------|
| `gemini` | [Google Gemini 3.1 Flash](https://openrouter.ai/google/gemini-3.1-flash-image) | "gemini", "nano banana 2" | Previous default; still marginally closer on the Opus benchmark; up to 4K ($0.067 · 12s at 1K) |
| `gemini-lite` | [Google Gemini 3.1 Flash Lite](https://openrouter.ai/google/gemini-3.1-flash-lite-image) | "gemini lite", "nano banana lite", "fast draft" | Fast cheap iteration; **1K only** ($0.034 · 5s) |
| `nano-banana-2.1` | [Google Nano Banana 2.1](https://openrouter.ai/google/gemini-nano-banana-2.1) (default) | "nano banana 2.1", "nb 2.1", "generate an image" (no model specified) | Versatile default, Flash-tier successor to `gemini`: refs, `--analyze`, up to 4K; near-identical fidelity for ~44% less ($0.038 · 12s at 1K) |
| `geminipro` | [Google Gemini 3 Pro](https://openrouter.ai/google/gemini-3-pro-image) | "geminipro", "gemini pro", "use gemini pro" | Highest-quality Gemini (~$0.17 at 2K) |
| `riverflow` | [Sourceful Riverflow v2 Pro](https://openrouter.ai/sourceful/riverflow-v2-pro) | "riverflow", "use riverflow" | Artistic/illustration ($0.15) |
| `flux2` | [FLUX.2 Max](https://openrouter.ai/black-forest-labs/flux.2-max) | "flux2", "flux", "use flux" | Illustration, clean lines (~$0.07/MP) |
| `seedream` | [ByteDance Seedream 5.0 Lite](https://openrouter.ai/bytedance-seed/seedream-5-0-lite) | "seedream", "use seedream" | **2K/4K only**, web-connected knowledge, 14 refs ($0.035 · 40s at 2K) |
| `gpt5.4` | [OpenAI GPT-5.4 Image 2](https://openrouter.ai/openai/gpt-5.4-image-2) | "gpt5.4", "gpt-5.4 image", "use gpt5.4" | Multimodal GPT; also `--analyze` (token-billed) |
| `gpt-sunburst` | [OpenAI GPT Image 2.5 Sunburst](https://openrouter.ai/openai/gpt-image-2.5-sunburst) | "gpt image", "gpt image 2.5", "sunburst" | Precision editing, 16 refs, `--quality` up to `max`, native `-t` ($0.015 · 26s at default quality; token-billed, rises with `--quality`) |
| `gpt-flare` | [OpenAI GPT Image 2.5 Flare](https://openrouter.ai/openai/gpt-image-2.5-flare) | "gpt flare", "fast gpt image" | Same features as Sunburst, speed tier ($0.015 · 19s at default quality) |
| `mai` | [Microsoft MAI-Image-2.6](https://openrouter.ai/microsoft/mai-image-2.6) | "mai", "microsoft image" | Multi-reference compositing of people/products/styles, 5 refs ($0.041 · 25s) |
| `mai-flash` | [Microsoft MAI-Image-2.6 Flash](https://openrouter.ai/microsoft/mai-image-2.6-flash) | "mai flash" | Faster MAI, same 5-ref editing ($0.020 · 15s) |
| `grok` | [xAI Grok Imagine Image 2.0](https://openrouter.ai/x-ai/grok-imagine-image-2.0) | "grok", "grok imagine" | 1K/2K, `--quality low\|medium`, 3 refs ($0.060 · 66s — billed above its $0.04 list price; +$0.01/ref) |
| `qwen` | [Qwen Image 3](https://openrouter.ai/qwen/qwen-image-3) | "qwen", "qwen image" | Small legible text (posters, UI, infographics), 1K/2K, 4 refs ($0.030 · 63s) |
| `qwen-pro` | [Qwen Image 3 Pro](https://openrouter.ai/qwen/qwen-image-3-pro) | "qwen pro" | Qwen Image 3 with richer world knowledge ($0.040 · 57s) |
| `muse` | [Meta Muse Image](https://openrouter.ai/meta/muse-image) | "muse", "meta image" | Complex multi-part prompts; reasons first, may search the web; prompt only ($0.010 · 18s, 1600px) |
| `recraft-flash` | [Recraft V4.1 Flash](https://openrouter.ai/recraft/recraft-v4.1-flash) | "recraft", "cheapest", "quick draft" | Cheapest/fastest ~1K drafts; no `-r` ($0.007 · 5s) |

Models from `seedream` down to `recraft-flash` (except `gpt5.4`) use the OpenRouter **Images API** (`/v1/images`). Each accepts only the `-a`/`-s`/`--quality`/`-r` options it supports, and the script rejects anything else before calling the API. Run `--list-models` to see each model's limits.

**Cost and time** are real OpenRouter charges and end-to-end times (through the Cloudflare gateway) at default settings. They come from the Claude Opus robot benchmark, one sample per model on 2026-09-28 (`nano-banana-2.1` on 2026-10-07). `geminipro`, `riverflow`, `flux2` and `gpt5.4` were not benchmarked; their figures are list prices or earlier cost-log values. For per-model output format, resolution and quality notes, read `references/model-benchmarks.md`. Re-run the benchmark with the `ai-image-test-run` skill.

## Instructions

> **Routing check:** If the user asks to **describe, analyze, or explain an existing image** (not generate a new one), skip directly to the **Image Analysis (`--analyze`)** section below. No prompt enhancement or output path needed.
>
> **Video routing:** If the user asks to **describe, analyze, or expla

CAUTIONgrade B · trust 89/100 Install with care. The audit found things worth knowing before you trust its output.

What the source does

Filesystem

UNDECLARED (5 observation(s))

Network

declared (3 observation(s))

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (11)

shutil.rmtree(tmp_dir, ignore_errors=True)

shutil.rmtree(tmp, ignore_errors=True)

shutil.rmtree(grounding_tmpdir, ignore_errors=True)

shutil.rmtree(grounding_tmpdir, ignore_errors=True)

shutil.rmtree(grounding_tmpdir, ignore_errors=True)

image_bytes = base64.b64decode(b64_data)

image_bytes = base64.b64decode(b64_data)

image_bytes = base64.b64decode(b64_data)

Edit `~/.zshrc` (macOS) or `~/.bashrc` (Linux):

- **uv** (Python runner): Install via `curl -LsSf https://astral.sh/uv/install.sh | sh` or `brew install uv`

Install uv: `curl -LsSf https://astral.sh/uv/install.sh | sh` or `brew install uv`

Gates applied: no_behavioural_pass.