video-model-selection

Choosing a generation model

fal-ai and google-ai teach how to call a model once you've picked one. This skill teaches how to pick it. Read references/model-capabilities.md whenever quoting model duration, resolution, aspect ratio, audio, language, or reference support. Read references/cost-guidance.md whenever the user compares models, asks about price, or must approve a paid plan. Neither this skill nor either of those pins a model catalog or a "best model for X" table -- both providers' catalogs, capabilities, and pricing change independently and on their own schedule, and a fixed ranking here would go stale exactly the way a fixed model ID would.

Resolve production-level unknowns before the first paid call

video-creation already asks one compact question only when a missing choice would substantially change the result or cost (see "Understand the request"). A long-form AI-generated production has its own unresolved-question shape, and getting it wrong is more expensive to unwind than a typography choice: a wrong shot is a paid regeneration, not a free re-render. Before the first generation call, resolve whatever the brief hasn't already answered:

• Provider/model preference. Does the user already know they want fal.ai, Google, or a specific model? If not, state the tradeoff briefly (see "Decide by requirement" below) rather than picking silently on their behalf for a paid call.
• Character/subject consistency requirements. Does any character, presenter, or product need to recognizably recur across multiple shots? This determines which models are even viable (see video-cinematography's consistency section) and must be known before the first shot of that subject is generated, not discovered after three independently-drifting attempts.
• Shot count vs. budget. A true long-form (8-15+ minute) production is assembled from dozens of separately-generated clips. Do the arithmetic for the actual model in hand rather than quoting a fixed number: a 10-minute piece is 600 seconds of finished runtime, so a model producing 5-second clips needs ~120 calls, one producing 15-second clips ~40, and one producing 30-second clips ~20 -- before accounting for re-prompts, which the regeneration tolerance below multiplies on top. Per-call duration is therefore a budget decision as much as a creative one. If the user hasn't stated a cost ceiling or a maximum acceptable number of paid generation calls, ask once, compactly, showing the call-count arithmetic for the model actually being considered -- do not silently generate an open-ended number of variants across a production this large.
• Regeneration tolerance. If a shot doesn't match the brief on the first attempt, how many re-prompts is reasonable before stopping to ask the user rather than continuing to spend on the same shot? State a default (e.g. two re-prompts) and get confirmation rather than assuming unlimited budget for one stubborn shot.