run-fal-generation

Run FAL Generation

Execute a saved prompt against the official fal-ai MCP server (https://mcp.fal.ai/mcp), then save the output into the project's visuals/ directory and update the prompt YAML.

When to use

• The user has a finalised prompt saved in prompts/<slug>.yaml and wants to generate the visual.
• The user wants to iterate on a generation or upload a reference for image-to-image.
• The user wants to look up pricing or pick a model before generating.

Prerequisites: official fal-ai MCP

This skill targets fal's official remote MCP server. If it isn't connected, install it once:

claude mcp add --transport http fal-ai https://mcp.fal.ai/mcp \
  --header "Authorization: Bearer $FAL_KEY"

(Source: https://fal.ai/docs/documentation/setting-up/mcp — auth is a bearer token in the header; no npm package or env-var-only config.)

If mcp__fal-ai__* tools are not exposed in the current session, stop and instruct the user to run the command above with their fal API key.

Tools exposed by the official server

Discovery:

• mcp__fal-ai__search_models — find models by keyword/category
• mcp__fal-ai__get_model_schema — full input/output schema for a model
• mcp__fal-ai__get_pricing — cost of running a model before you call it
• mcp__fal-ai__search_docs — search fal docs
• mcp__fal-ai__recommend_model — describe goal, get model suggestions

Execution:

• mcp__fal-ai__run_model — run a model and wait for the result (images, video, audio)
• mcp__fal-ai__submit_job — submit a long-running job, return a request ID
• mcp__fal-ai__check_job — poll status / fetch result / cancel

Utility:

• mcp__fal-ai__upload_file — upload a local file (or URL) to fal's CDN; returns a cdn_url usable as image_url/audio_url input

Outputs from run_model / check_job are returned as CDN URLs, not local paths — this skill is responsible for downloading them.

Inputs to gather

• Project name (or active project)
• Prompt slug — which prompts/<slug>.yaml to execute
• Model — from the prompt YAML, or override; if unsure, call recommend_model or search_models
• Iteration notes — any tweaks to merge into the prompt before submission
• Reference image — for image-to-image; local path that needs upload_file first

Procedure

1. Verify the MCP is connected. Check that mcp__fal-ai__* tools are available. If not, direct the user to the install command above and stop.

2. Load project + prompt. Read ${CLAUDE_PLUGIN_DATA:-$HOME/.local/share/claude-plugins}/visual-communications/projects/<project>/project.yaml and prompts/<slug>.yaml.

3. Pick the model (if not set). If the prompt YAML lacks a concrete fal model id (e.g. fal-ai/flux/dev, fal-ai/flux-pro, fal-ai/minimax/video-01), call recommend_model with a one-line description, or search_models with a keyword. Confirm the choice with the user if there's ambiguity.

4. Sanity-check inputs and price. Call get_model_schema to confirm parameter names (image models commonly take prompt, image_size / aspect_ratio, num_images, negative_prompt; video models often take prompt, image_url, duration). Optionally call get_pricing and surface the cost to the user before running.

5. Upload references if needed. For image-to-image or video-from-image, call upload_file on the local reference and use the returned cdn_url as image_url in the model input.

6. Run the model.

   • Default: mcp__fal-ai__run_model (synchronous; preferred for short jobs)
   • Long-running (most video, large batches): mcp__fal-ai__submit_job → poll with mcp__fal-ai__check_job until status is completed. Use submit_job whenever the model's typical latency exceeds ~60s.

7. Download the outputs. Parse output URLs from the response (typical fields: images[].url, video.url, audio.url). For each, download to:

   ${CLAUDE_PLUGIN_DATA:-$HOME/.local/share/claude-plugins}/visual-communications/projects/<project>/visuals/<slug>-vN.<ext>
   

   Use curl -L -o <dest> <url> (or wget). Auto-increment the vN suffix; never overwrite existing files without confirmation.

8. Update the prompt YAML. Append every saved file to generated_files::

   generated_files:
     - ../visuals/hero-vN.png
   model_used: fal-ai/flux/dev
   generated_at: 2026-04-30T22:15:00+03:00
   

9. Report. Show the saved paths, the model id used, and the price (if get_pricing was called). Offer next steps: regenerate with tweaks, run a new prompt, or upscale an output (call recommend_model with "upscale" if asked).

Tips

• For Flux: fal-ai/flux/dev is a good default for stills; fal-ai/flux-pro/v1.1 for higher fidelity.
• For video: most models are async — prefer submit_job/check_job.
• For aspect ratios: pass the model's expected enum (square_hd, landscape_16_9, etc.) — get_model_schema is the source of truth.
• Negative prompts: only some models accept negative_prompt; check the schema before passing it.

Constraints

• Do not invent fal model ids or tool parameters — always verify via get_model_schema / search_models first.
• Never overwrite existing files in visuals/ without explicit user confirmation.
• If a job fails (rate limit, quota, NSFW filter, model error), surface the exact fal error message; don't retry silently.
• Pricing varies — for batch generations, confirm with the user before kicking off jobs that will run >5 images or any paid video.