Grok Image Generation Skill

Workflow

1. Receive description — the user describes what they want in any language
2. Craft prompt — translate to English, add style-appropriate quality keywords
3. Pick model — choose variant based on need (see Model Selection below)
4. Generate — run the script; it handles provider discovery and failover
5. Display — Read the output image file to show the user
6. Deliver — ALWAYS send to IM after display using send.sh with auto routing. Only skip if the user explicitly says they don't need delivery, or if running in direct CLI mode with no IM channels configured.

Commands

Generate

bash ${CLAUDE_SKILL_DIR}/bin/generate.sh "<english_prompt>" [size] [model]

The script:

• Reads providers from ~/clawd/proxy/config.json
• Tries each matching provider, falling back to other grok-imagine variants
• Skips auth errors (401/403), retries on rate limits (429) and timeouts
• Outputs image path to stdout on success, exits 1 on failure

Send to IM

bash ${CLAUDE_SKILL_DIR}/bin/send.sh <image_path> [channel]

Source-aware routing: In auto mode (default), the script reads the claude-to-im bridge's binding data to detect which channel sent the current request, and routes the image back to that same channel. If the request came from Telegram, the image goes to Telegram; if from Feishu, it goes to Feishu. If detection fails (e.g. direct CLI usage), it falls back to the first available channel.

When delivering images, prefer using auto (omit the channel argument) so the image goes back to wherever the request came from. Only specify a channel explicitly when the user asks for a specific destination (e.g. "发到飞书" → feishu, "send to all" → all).

Config in ~/.claude-to-im/config.env:

• Feishu/Lark: CTI_FEISHU_APP_ID, CTI_FEISHU_APP_SECRET, CTI_FEISHU_ALLOWED_USERS
• Telegram: CTI_TG_BOT_TOKEN, CTI_TG_CHAT_ID
• Discord: CTI_DISCORD_WEBHOOK_URL

Model Selection

• grok-imagine-1.0 — best quality, use by default for most requests
• grok-imagine-1.0-fast — faster but slightly lower quality, good when speed matters
• grok-imagine-1.0-edit — for image modification concepts

Prompt Crafting

Translate the user's idea to fluent English, then add keywords matching the style:

• Photo / Portrait: cinematic photography, ultra realistic, soft bokeh, golden hour lighting
• Illustration / Cartoon: digital illustration, vibrant colors, detailed, trending on artstation
• Concept Art: concept art, highly detailed, dramatic lighting, 8k resolution

Keep the user's creative intent intact — enhance, don't override.

Example

User: "画一张赛博朋克风格的城市夜景"

1. Prompt: "A cyberpunk cityscape at night, neon lights reflecting on wet streets, towering skyscrapers with holographic ads, flying vehicles, cinematic photography, ultra realistic, 8k resolution"
2. bash ${CLAUDE_SKILL_DIR}/bin/generate.sh "A cyberpunk cityscape..." "1024x1024"
3. Read the output file to display
4. ALWAYS deliver: bash ${CLAUDE_SKILL_DIR}/bin/send.sh /tmp/grok_image_xxx.jpg (auto-routes back to the channel that sent the request) User can specify destination: "发到飞书" → feishu, "send to all" → all

Notes

• Generation takes 5-30 seconds depending on provider and model
• Images save to /tmp/grok_image_<timestamp>.jpg (cleaned on reboot)
• If all providers fail, suggest trying again later