fal.ai Skills

fal.ai Skills are agent-ready knowledge layers designed for platforms like Claude.ai Projects and Claude Code. They provide agents with the necessary information to discover endpoints, inspect schemas, run models, poll status, and download outputs.

Most skills function as knowledge layers that instruct an agent on:

• Which endpoint to use.
• How to prompt the model.
• How to chain multiple calls together.

The primary execution path for these skills is the genmedia CLI, which provides the following core capabilities:

• models: Discover available models.
• schema: Inspect endpoint schemas.
• run: Execute a model.
• status: Poll for execution status.
• upload: Upload assets.
• pricing: Check costs.
• docs: Access documentation.

Use the ugc skill to plan and run User Generated Content (UGC) style creator ads and social proof videos. This is intended for content that feels like a creator's authentic video (e.g., TikTok, Reels, Shorts) rather than polished studio advertising.

Use cases include:

• Direct-to-camera creator scripts and talking-head ads.
• Product demos, unboxings, and testimonials.
• Founder clips, reaction videos, and before-after comparisons.
• Faceless voiceovers and short vertical social videos.

Important Guardrails:

• Claims: Do not write fake testimonials, invented reviews, medical claims, or financial claims. Use observable statements (e.g., "the texture looks lighter") instead of unsupported claims (e.g., "this cures acne").
• Fidelity: Ensure product shape, packaging, and logos remain stable. Avoid hallucinating captions or text overlays inside the video.

When designing workflows in fal.ai, you can utilize several structural patterns to manage data flow and dependencies:

1. LLM Prompt → Image → Video

Sequential generation where an LLM generates prompts for subsequent media generation steps. [Input] → [LLM: Image Prompt] → [Image Gen] → [LLM: Video Prompt] → [Video Gen] → [Output]

2. Parallel Processing (Fan-Out)

Executing multiple independent processes from a single hub node. All parallel nodes depend on the hub, but do not depend on each other. [Hub Node] → [Process A/B/C] → [Merge] → [Output]

3. Multi-Destination Campaign

A complex pattern for generating multiple variations (e.g., for different marketing destinations) and merging them into a single output. [Input 1..N] → [Vision LLM/Edit/Upscale/Video Gen] → [Merge Videos] → [Output]

Use the Channel Matrix to determine the required aspect ratios and asset types for specific distribution channels:

When planning the video sequence using multi_prompt, follow these constraints to ensure valid Kling execution:

Hard Rules:

• Beat Duration: Each multi prompt must be at least 3 seconds.
• Total Duration: The total video duration must be 15 seconds or less.
• Prompt Count: Use between 2 and 5 multi prompts.
• Duration Sync: Set the top-level Kling duration equal to the sum of all beat durations.
• Prompt Length: Keep Kling prompts concise (250-430 characters).

Recommended Patterns:

• Simple cutaway: 2 beats, 6 seconds total.
• Standard reaction: 3 beats, 9 seconds total.
• Rich fan-cam moment: 4 beats, 12 seconds total.
• Full story beat: 5 beats, 15 seconds total.

Use this pattern for generating character poses, product angles, style variants, or dataset panels. Instead of generating multiple separate images, you generate a single grid (contact sheet) and then process the individual panels.

Rules:

• Generate all panels in a single image to ensure visual consistency.
• Prompt for an exact grid layout and specify no visible borders or gaps.
• If the crop utility requires percentage coordinates, crop by percentage.
• Perform upscaling after the cropping step to ensure final quality.

generate grid -> crop panel 1..N -> upscale panel 1..N -> optional video

Kling provides two distinct modes for video-to-video tasks:

Edit Mode (/edit)

Changes specific content while preserving the original motion. Best Practice: Explicitly state what to change and what to preserve. Example: Change: the sky becomes a starry night with a visible moon. Preserve: the subject, all motion, the foreground composition, the camera move.

Reference Mode (/reference)

Restyles the entire video to match a specific style cue (e.g., "transform into watercolor").

The Image Restoration recipe acts as a smart dispatcher. To achieve the best results, identify the dominant defect in the source image and select the corresponding specialist endpoint:

If multiple defects are present, you should chain the processes (e.g., denoise $\rightarrow$ deblur $\rightarrow$ upscale).

The generation process follows three main steps:

1. Brief extraction: Uses openrouter/router to transform your input context into a structured JSON brief containing brand details, taglines, sections, tone, palette hints, and hero image prompts.
2. Variant generation: For each design direction, an LLM generates a complete, single-file index.html using Tailwind via CDN.
3. Hero image: fal-ai/gpt-image-2 renders a cinematic hero image for each variant. The image is downloaded and inlined as a data: URI to ensure the HTML file is standalone.

A Planner node is responsible for producing the workflow graph, shot table, or variant matrix. To prevent interpretation drift during execution, the output must be highly structured.

Ensure the Planner node includes:

• node_ids
• Dependencies
• Intended endpoint role
• Inputs
• Expected outputs

Keep creative planning logic separate from the executable parameters.