# BananaHub Skill 🍌 [简体中文说明](./README.zh-CN.md) ![BananaHub Skill hero: agent-native image workflows](./docs/assets/github/readme-hero-gpt-5-4.png) BananaHub is an agent-native image workflow skill. Describe what you want in Chinese, English, or mixed language; BananaHub cleans up the prompt, detects the available runtime, routes the task to the right image path, and keeps generation, editing, templates, and reusable prompts under one `/bananahub` command. It is not just a prompt helper. BananaHub acts as a workflow layer between agents and image providers: GPT Image 2 by default, plus Gemini / Nano Banana, OpenAI official, compatible gateways, and host-native image tools when configured. When installed in hosts that support implicit skill loading, BananaHub can also appear for ordinary image requests. In that case it asks whether you want prompt optimization first; if you decline, the agent can continue with its normal image tool. ## Why BananaHub - **Works in more environments**: direct provider calls when keys are configured, host-tool delegation when the client already has image generation, prompt-only output when no image runtime is available. - **Less rework**: locks text, layout, preserved elements, and edit boundaries before generation. - **Multi-model without guesswork**: prompt cleanup and archiving are cross-model; mask edits, native sizes, reference limits, quality presets, and output formats stay provider/model-scoped. - **Reusable by default**: archive final prompts for handoff, QA, reruns, and template authoring. - **Template-ready**: use built-ins, install more from BananaHub, or create repeatable prompt/workflow modules. ## Quick Start ```bash # Open Agent Skills / skills.sh npx skills add https://github.com/bananahub-ai/bananahub-skill --skill bananahub # Or install directly in Claude Code claude skill install https://github.com/bananahub-ai/bananahub-skill /bananahub init /bananahub 一只橘猫趴在键盘上打盹 ``` `init` defaults to GPT Image 2 through an OpenAI-compatible image endpoint. If your Codex or agent host already exposes a built-in image generation tool, BananaHub can use it as a host-native channel without storing a local API key. Check the current execution path: ```bash python3 scripts/bananahub.py check-mode --pretty ``` Test the host-native image channel: ```text /bananahub test-host-imagegen ``` ## Workflow ![BananaHub Workflow: describe, optimize, route, generate, reuse](./docs/assets/github/user-flow-infographic-gpt-5-4.png) The core loop is simple: describe the image, let BananaHub optimize and route the request, generate through the configured provider, then reuse the archived prompt or template. ## Runtime Modes | Mode | When It Applies | What BananaHub Does | |---|---|---| | `provider-backed` | A supported provider and key are configured | Optimizes the prompt, calls the provider, and saves outputs | | `host-native` | Local provider config is incomplete, but Codex or the host agent has an image tool | Optimizes/archives the prompt, then delegates image generation to the host | | `prompt-only` | No provider and no host image tool are available | Returns a reusable prompt; never pretends an image was generated | Codex OAuth and some API login modes expose a built-in image generation tool. For CLI checks, use `BANANAHUB_HOST_IMAGEGEN=1` or `check-mode --host-imagegen` to mark host-native image generation as available. ## Setup Flow ![Set up BananaHub: install, init, configure, check, generate](./docs/assets/github/setup-guide-gpt-5-4.png) Configuration is intentionally profile-based: initialize once, persist provider credentials locally, check the runtime path, and keep future runs predictable. ## Best Fit - **Everyday visual work**: product clean shots, background replacement, article one-pagers, infographics, and repo explainers. - **Knowledge visuals**: turn processes, long-form content, or codebase context into readable one-image summaries. - **Provider-aware teams**: use the same template IDs across Gemini / Nano Banana and OpenAI GPT Image without hiding model-specific limits. - **Reusable workflows**: archive prompts and install richer templates from BananaHub instead of bloating the skill package. ## Commands | Command | Use | |---|---| | `/bananahub ` | Optimize the prompt and generate an image | | `/bananahub edit --input ` | Edit an existing image | | `/bananahub optimize ` | Optimize only, no generation | | `/bananahub generate ` | Generate directly from an English prompt | | `/bananahub models` | List available models | | `/bananahub check-mode` | Inspect runtime mode and capability layers | | `/bananahub test-host-imagegen` | Verify Codex/host built-in image generation | | `/bananahub templates` | List built-in and installed templates | | `/bananahub use ` | Use a prompt template or start a workflow template | | `/bananahub discover ` | Search BananaHub for matching templates | | `/bananahub capture-workflow` | Turn the current image iteration into a reusable workflow draft | | `/bananahub init` | Check and initialize the environment | For ordinary image requests without `/bananahub`, BananaHub is advisory by default: it prompts before optimizing and does not silently replace the host's image generator. ## Common Flags | Flag | Meaning | |---|---| | `--direct` | Fewer confirmations; still conservative about creative additions | | `--raw` | Translate/clean only, no extra enhancement | | `--model ` | Select a model, e.g. `gemini-3-pro-image-preview`, `gpt-image-2` | | `--provider ` | Override provider for one run, e.g. `openai`, `google-ai-studio` | | `--aspect ` | Aspect ratio, e.g. `16:9`, `1:1`, `9:16` | | `--image-size ` | Gemini native size preset: `1K`, `2K`, `4K` | | `--n ` | OpenAI Images API number of generated/edited outputs | | `--openai-size ` | OpenAI Images API native size option | | `--quality ` | Provider-native quality preset | | `--background ` | Provider-native background option | | `--output-format ` | Provider-native output format, e.g. `png`, `jpeg`, `webp` | | `--output-compression ` | Provider-native output compression when supported | | `--response-format ` | Optional legacy OpenAI Images response override: `url` or `b64_json` | | `--timeout ` | OpenAI image request timeout override | | `--resize ` | Resize after generation/editing | | `--output ` | Image output path | | `--save-prompt` | Archive the final prompt under `bananahub-prompts/` | | `--prompt-output ` | Save the final prompt to a file or directory | | `--input ` | Source image for edit commands | | `--ref ` | Reference images for edit commands | | `--mask ` | OpenAI-native mask edit | Set `BANANAHUB_SAVE_PROMPTS=1` to archive final prompts by default for generate/edit runs. Set `BANANAHUB_IMAGE_TIMEOUT=` to override the OpenAI image request timeout; the default is `900`. Use `--timeout ` for a single command. ## Prompt Archive Prompt archive is useful for QA, handoff, reruns, provider switching, and template design. ```bash python3 scripts/bananahub.py generate \ "A clean product photo of a blue wireless earbud case" \ --save-prompt python3 scripts/bananahub.py generate \ "A launch poster for BananaHub" \ --prompt-output docs/prompts/launch-poster.md ``` The archive file stores command metadata, provider, model, timestamp, and the final prompt. It is written before the provider call, so a failed API request still leaves a reusable prompt behind. ## Provider Paths BananaHub supports several provider routes. Advanced capabilities are not assumed globally; use `check-mode` and the capability registry to verify what is available. | Provider | Best For | Generate | Edit | Mask | Notes | |---|---|---:|---:|---:|---| | `openai-compatible` | Recommended default for OpenAI-style image gateways | ✅ | ✅ | ✅ | GPT Image 2 path; standard Images API support depends on the gateway | | `openai` | Official OpenAI GPT Image APIs | ✅ | ✅ | ✅ | GPT Image native route; supports multi-output and multi-image edits through Images API | | `google-ai-studio` | Gemini / Nano Banana users | ✅ | ✅ | — | Gemini Developer API route | | `gemini-compatible` | Gemini-style relays or proxies | ✅ | ✅ | — | Depends on the relay | | `vertex-ai` | Enterprise GCP / Vertex AI | ✅ | ✅ | — | ADC or API key auth | | `chatgpt-compatible` | Chat/completions endpoints that return images | ✅ | — | — | Best-effort URL/base64 extraction | Setup examples: ```bash # Agent/script-readable diagnosis before provider calls python3 scripts/bananahub.py config doctor --json # Recommended GPT Image gateway profile python3 scripts/bananahub.py config quickset --provider openai-compatible --profile gpt --default-profile \ --base-url https://your-openai-compatible-endpoint --api-key --model gpt-image-2 # OpenAI official GPT Image profile python3 scripts/bananahub.py config quickset --provider openai --profile gpt --default-profile \ --api-key --model gpt-image-2 # Google AI Studio / Gemini Developer API profile python3 scripts/bananahub.py config quickset --provider google-ai-studio --profile nano --default-profile \ --api-key --model gemini-3-pro-image-preview # Gemini-compatible relay profile python3 scripts/bananahub.py config quickset --provider gemini-compatible --profile nano --default-profile \ --base-url https://your-gemini-endpoint --api-key --model gemini-3-pro-image-preview # Chat/completions-compatible image endpoint python3 scripts/bananahub.py config quickset --provider chatgpt-compatible --profile chat --default-profile \ --base-url https://your-chat-endpoint --api-key --model gpt-5.4 # Vertex AI ADC profile python3 scripts/bananahub.py config quickset --provider vertex-ai --profile vertex --default-profile \ --auth-mode adc --project --location global # Human-terminal fallback when you want guided prompts python3 scripts/bananahub.py init --wizard ``` Agents should use `config doctor --json`, collect only missing provider fields, and persist them with `config quickset`. Direct API-key entry is supported when the user chooses it; otherwise provide a quickset command with placeholders for the user's local terminal. When an agent receives the key directly, prefer `--api-key-stdin` over `--api-key` so the secret is not placed in command arguments. Manual validation for an OpenAI-compatible `gpt-image-2` gateway: ```bash BANANAHUB_PROFILE=gpt python3 scripts/bananahub.py generate \ "Create a cute sticker of a tiny cheerful robot holding a banana, rounded kawaii vector style, thick clean outlines, soft pastel colors, plain white background, no text." \ --model gpt-image-2 \ --aspect 1:1 \ --no-fallback \ --output /tmp/bananahub-gpt-image-2-retry-test.png ``` The command should return `status: "ok"` with `actual_model: "gpt-image-2"`; telemetry `HTTP 403` warnings do not indicate generation failure. Config priority: `--config ` → selected profile in `~/.config/bananahub/config.json` → environment variables fill missing fields. Use `BANANAHUB_PROFILE=` to pick another persisted profile. Set `BANANAHUB_ENV_OVERRIDE=1` only when environment variables should temporarily override profile fields. Agent-facing diagnostics are provider-scoped: `resolved_from` only points to active values, `ignored_config_sources` lists values inactive for the selected provider, and `env_shadowed_config_sources` lists environment values ignored because the selected profile already defines that field. ## How Prompt Optimization Works ```text User input │ ├─ Extract constraints: exact text / keep / avoid / platform / edit invariants ├─ Clean up: keyword dumps, SD/MJ syntax, negative phrasing, buried subjects ├─ Translate smartly: descriptive content → English; in-image text and proper names stay locked ├─ Detect intent: photo / product / diagram / sticker / text-heavy / 3d / concept-art ... └─ Enhance conservatively: add only useful detail; confirm high-impact creative choices ``` This layer is cross-model. Provider-dependent features such as mask edit, exact size, transparent background, quality presets, and reference-image limits are routed through provider/model metadata. ## Templates Templates come in two shapes: ![BananaHub Skill Architecture: user to outputs through prompt pipeline, templates, and provider router](./docs/assets/github/skill-architecture-gpt-5-4.png) - **Prompt templates** assemble one reusable prompt for stable jobs such as product clean shots, infographic cards, or background replacement edits. - **Workflow templates** guide multi-step jobs such as article one-page summaries or repository explainer diagrams. ```bash /bananahub templates /bananahub templates article-one-page-summary /bananahub use info-diagram "PR review flow: Change, Review, Test, Fix, Merge" /bananahub use background-replace-edit --input product.png /bananahub discover logo system /bananahub create-template /bananahub capture-workflow ``` Built-ins are a lean starter pack: `product-white-bg`, `info-diagram`, `article-one-page-summary`, `repo-explainer-diagram`, and `background-replace-edit`. They are practical, low-maintenance, sample-free, and support both Gemini/Nano Banana and OpenAI GPT Image through provider-specific prompt variants. After a multi-turn image session, use `/bananahub capture-workflow` to summarize the accepted result, useful prompts, failed attempts, locked invariants, allowed deltas, and model/channel choices into a `type: workflow` template draft. It asks before writing files and can save to user templates, the current project, or a publishable template folder. Install richer official templates from `bananahub-ai/templates` through BananaHub discovery when you need style packs, logo systems, character workflows, campaign visuals, or heavier sample galleries. Community templates remain available through the broader catalog. User-installed templates live in `~/.config/bananahub/templates/` and take precedence over built-ins with the same ID. ## References - Prompt rules: `references/prompt-guide.md` - Capability layers: `references/capability-registry.md` - Model registry: `references/model-registry.json` - Template spec: `references/template-format-spec.md` - Official sources: `references/official-sources.md` ## Requirements - Claude Code / Open Agent Skills runtime with skill support - Python 3.8+ - Optional: `google-genai`, `pillow` - Optional: Gemini, OpenAI, or compatible gateway API key ## Contributing Code and docs contributions are accepted under MIT. Built-in starter templates under `references/templates/` should stay lightweight and explicitly licensed; heavier sample galleries belong in official or community template repositories. ## License - Code, scripts, and general repo docs: MIT - Built-in starter templates under `references/templates/`: CC BY 4.0 unless a template directory states otherwise

bananahub

Installation

SKILL.md