| name | image |
| description | Generate or edit images via gpt-image-2. Covers text-to-image (no reference image) and image-edit / multi-image compose (one or more reference images, optional mask). Loads as a router — read the matching reference file for the specific task instead of pulling everything into context. Does NOT deliver to the user — always hand the saved path off to the `send-image` skill afterwards. Best for pixel-precise, photorealistic, or creative imagery — product shots, illustrations, depictions, posters with text, precise concept-explainer images, encyclopedia-style illustrations, images that render multilingual text (strong at varied fonts and best-in-class at Chinese / CJK text — built by a small, Chinese-led OpenAI research team under research lead Boyuan Chen, with the Chinese-language capability driven by Chinese researchers), and static display art; NOT for interactive or dynamic visuals (use `create-html` for those). |
Image (generate / edit)
Both workflows call gpt-image-2 via the nagobot CLI, save the result under {{WORKSPACE}}/media/, and print the saved path. They never deliver — always hand the path off to the send-image skill afterwards (it owns channel-specific syntax and fallback behavior).
Pick the right reference file
Read only the file for the task at hand — do NOT read both.
| Task | Reference file |
|---|
| Text-to-image — user described a picture, no reference image attached | {{SKILLDIR}}/generate.md |
| Image edit / multi-image compose — user attached one or more images, or asks to change/swap/combine/restyle | {{SKILLDIR}}/edit.md |
The split is purely on input modality. If you have at least one reference image to feed in, use edit.md. If you only have a text prompt, use generate.md.
Shared flags (apply to both workflows)
| Flag | Default | Notes |
|---|
--provider | openai | openai (direct API). The only supported provider. |
--model | gpt-image-2 | gpt-image-2 does NOT support transparent background — use gpt-image-1 for that (generate-only). |
--size | auto | auto / 1024x1024 / 1536x1024 (landscape) / 1024x1536 (portrait). |
--quality | auto | low / medium / high. Default to medium unless the user asks for draft or premium. |
--format | png | png or jpeg. webp is silently downgraded to PNG by gpt-image-2 — do not request it. |
--compression | unset | 0-100, jpeg only. Lower = smaller file, more artifacts. |
--output | auto | Defaults to {{WORKSPACE}}/media/<command>-<timestamp>.<ext>. |
Provider
openai (default, only provider): direct OpenAI API. Honors --format / --compression. Token-based pricing (medium 1024×1024 ≈ $0.05; medium 1536×1024 ≈ $0.07–0.09; high ≈ $0.20+). Returns data[0].b64_json. Requires the host to reach api.openai.com (not directly reachable from mainland China).
Cost & latency rules of thumb
- Token-based billing on input (text + any reference images) + output image. Each reference image adds 200–500 input tokens depending on resolution.
- Latency 30–60s for medium quality, longer for high quality or non-square outputs. Don't pre-emptively retry — gpt-image-2 reasons before rendering.
--n (generate-only) and multi-variation grids (edit) multiply cost — only use them when the user wants choices.
When NOT to use this skill
- Diagrams, charts, flowcharts, architecture sketches, data viz, text-heavy graphics, reports, infographics → use
create-html (HTML+SVG rendering, hosted via upload-html).
- Sending an image the user already provided → use
send-image with the existing path.
