Communitygithub.com

talki-io/img-video-skill

A Claude Code Agent skill that lets Claude generate images and videos via the jiekou.ai multi-model relay, supporting text-to-image, image-to-video, and models like Midjourney, Veo 3, Sora 2, Seedance, and Flux. Install to ~/.claude/skills to use.

What is img-video-skill?

This skill bridges the gap between Claude's text-only capabilities and visual content generation by wrapping external model services into a simple CLI script. It accepts a prompt, generates an image or video, saves the asset into an `assets/` directory, and prints the relative path for direct use in HTML, PPT, Markdown, or design files. The model layer is pluggable: by default it routes through the jiekou.ai relay, which aggregates providers like Midjourney, Flux, Veo 3, Sora 2, and Seedance, but it can fall back to direct providers like ByteDance's Jimeng, Kuaishou's Kling, or Google Gemini/Veo when keys are missing or requests fail. The skill supports text-to-image and image-to-video workflows, automatic fallback across providers, and configurable output directories. It is designed as a Claude Code Agent skill (install to ~/.claude/skills) and requires Python 3 with the requests library. It is ideal for generating banners, hero images, product shots, social media graphics, UI mockups, and short video clips directly from Claude's conversations.

Works with✓Claude Code✓Codex CLI~Cursor✓Gemini CLI
npx skills add talki-io/img-video-skill

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Preview

From the skill README

img-video-skill preview

Documentation

Img-Video — Asset Generation Bridge (Multi-Provider)

Claude cannot generate pixels directly, but it can call external model services. This skill encapsulates that pipeline into a single script: Provide a prompt → Get an image/video → Save to assets/ → Print the relative path, ready for use in HTML, PPT, Markdown, or design files.

The model layer is pluggable: by default it routes through Interface AI (jiekou.ai) relay, and if a key is missing or a call fails, it automatically falls back to providers like Jimeng / Kling / Gemini, etc. See the full list in references/providers.md.

When to Use

Use this skill whenever your final deliverable should contain an image or a video and a placeholder or plain text is not appropriate. Typical scenarios:

  • Web pages / landing pages / email templates: Banners, hero images, illustrations, icon backgrounds
  • PPT / slides: Covers, section pages, content images
  • Graphic design: Posters, covers, social media cards, thumbnails
  • UI prototypes: Diagrams, product screenshot placeholders
  • Short videos: Dynamic backgrounds, product showcases, animations driven by a single image

Principle in one sentence: Never leave a blank or write "(insert image here)" — just generate it.

Prerequisites: Configuration

Before first use, prepare a config.json in the skill root directory (the script will error with instructions if it doesn't exist):

  1. Copy the template: cp config.example.json config.json
  2. Configure at least one provider — API keys must be obtained from the user, never fabricated. For each provider, set:
    • type: Provider type (see table below)
    • base_url: API endpoint (has defaults, can be omitted)
    • Credentials: Usually api_key; for Kling, use access_key + secret_key
    • image_model / video_model: Model name (leave null if not needed)
  3. Set default_provider (primary) and fallback (fallback chain, an array).

Keys are stored only in the local config.json (ignored by .gitignore). Do not include them in code, SKILL.md, or commit them to version control.

Before each session, check availability (confirm keys/capabilities are ready, see the attempt order):

python <skill>/scripts/generate.py providers

Usage

Run this in your project working directory so that the output assets/ relative paths can be directly referenced by web pages or PPT.

Text-to-Image

python <skill>/scripts/generate.py image \
  --prompt "Flat illustration style tech company homepage hero image, blue-purple gradient, clean white space" \
  --size 1024x1024 --n 1 --output-dir assets/images
ParameterRequiredDefaultDescription
--prompt✅—Be specific (style + subject + color + composition) for best results
--size1024x1024Some relays scale proportionally, converting to 1:1 internally
--n1Number of images to generate
--output-dirassets/imagesOutput directory
--providerFollows fallback chainForce a specific provider, skipping fallback

Image-to-Video

python <skill>/scripts/generate.py video \
  --image assets/images/xxxx.png \
  --prompt "Camera slowly pushes in, background light spots gently floating" --output-dir assets/videos
ParameterRequiredDefaultDescription
--image✅—Local path or http(s) URL
--promptEmptyMotion/scene description; recommended
--output-dirassets/videosOutput directory
--providerFollows fallback chainForce a specific provider

Image-to-video tasks are mostly asynchronous; the script automatically polls until completion.

Specify / Switch Providers

  • Without --provider, it automatically tries default_provider → fallback in order (skipping providers without keys or without the required capability).
  • With --provider <name> (e.g., kling/ark/gemini/relay), force a specific provider, skipping the fallback chain. Example: if the user says "use Kling" → --provider kling.

Supported Model Services

typeServiceText-to-ImageImage-to-Video
jiekouInterface AI Relay (aggregates Midjourney/Flux/Seedance/Kling/Sora/Veo, default)✅✅
ark / jimengByteDance Jimeng Seedream/Seedance (direct)✅✅
klingKuaishou Kling (direct)✅✅
geminiGoogle Imagen/Veo (direct)✅✅
relayOther OpenAI-compatible relay✅⚠️ Depends on gateway

Recommended combinations (jiekou): Text-to-image: mj-txt2img (Midjourney, generates 4 at once), flux-2-pro; Image-to-video: veo-3.1-generate-img2video (Veo 3.1, with native audio), sora-2-img2video, seedance-v1-pro-i2v.

⚠️ Model switch important: Different models have different prompt field names, duration/resolution enums, and required fields (e.g., Midjourney uses text, Veo uses duration_seconds+generate_audio). Changing only the model name without adjusting these fields will result in a 400 error. The config already provides switches (prompt_field / send_aspect_ratio / duration / resolution / video_extra), refer to the "Model Field Differences" quick reference in references/providers.md.

Output

  • Images are saved to assets/images/, videos to assets/videos/.
  • Naming: {timestamp}_{first_words_of_prompt}.{png|jpg|mp4|…} (identifies actual format from file header, automatically avoids overwriting).
  • The script prints a list of relative paths to the generated files — ready for use in <img src="assets/images/…">, PPT placeholders, or Markdown.

Typical Workflow

  1. Use the providers subcommand to confirm available services and the attempt order.
  2. Plan which images/videos your final deliverable needs, their purpose, and style.
  3. Generate each image (or image then video) one by one, using --provider when you need a specific service.
  4. Collect the printed relative paths and incorporate them into your deliverable.

Example: User says "Help me build a coffee brand landing page" → image --prompt "Warm tone pour-over coffee close-up, morning light, shallow depth of field" to generate hero → Generate two more product images → Use those assets/images/*.png to write the HTML.

Exceptions and Fallback (Built into Script)

  • Missing config.json / missing providers section: Error with instructions to configure.
  • Primary provider missing keys / unsupported model: Automatically skips to the next.
  • Runtime failure (4xx/5xx/timeout): 5xx and network errors retry with exponential backoff; if still failing, degrades to the next provider in the fallback chain; reports an error only if all providers fail, with a summary of failure reasons from each.
  • Video timeout: Exceeds poll_timeout, reports a clear timeout.
  • Input image not found (image-to-video): Direct error, check --image.

Dependencies & Extensibility

  • Dependencies: Python 3 + requests (pip install requests). Kling JWT authentication uses the standard library, no extra dependencies.
  • Adding new providers: Add a new Provider subclass in scripts/providers.py, implementing generate_image / generate_video, then register it in the REGISTRY — fallback, saving, naming, etc., require no changes. See the end of references/providers.md for details.

Related Skills