Can an open-source prompting skill direct a multi-shot AI film with zero re-rolls?

Hypothesis

Video-model prompting is folklore — scattered across Discord threads and X posts. If that folklore is packaged as an agent skill (a structured, loadable playbook), an AI agent should be able to direct a coherent multi-shot film on the first attempt, not iterate its way there.

What we built

We took a community skill for ByteDance's Seedance 2.0 — 200+ files of distilled prompting grammar — security-reviewed it, and had an agent apply it to our in-house character pipeline: Pam, Thrivbe's openly-AI digital colleague.

The brief: a ~37-second film where Pam introduces the AI OS and her agent colleagues. The skill contributed the directing discipline — one action + one camera move + motivated light + explicit sound per shot, Shot 1/2/3: labels to force real editorial cuts inside a single generation, and single-take contracts for the talking-head clips. Our own canon contributed the identity anchor, the painterly style register, and native-speech dialogue written verbatim into the prompt.

Learnings

Three renders, three keepers — zero re-rolls. The 15-second tour clip executed three distinct shots with clean cuts in one generation: Pam walking toward the Kernel (a blue orb of orbiting clocks), a tracking move past Hermes and Lem (her media and outreach colleagues, rendered as glowing entities), and a vista over the OS "city" at dusk. Whisper transcripts confirmed every scripted line was spoken word-for-word by the model's native audio. The finished film was published straight to Instagram through our social pipeline.

One new pipeline capability emerged: first-frame identity locking and native speech work in the same call — talking-head clips no longer need a weaker reference mode.

Packaged prompting skill beats accumulated folklore. The skill's shot grammar is now part of Pam's production pipeline.

Log

• 2026-07-17 — Security-reviewed the seedance-2.0 community skill and applied it to the Pam pipeline. Three renders, three keepers, zero re-rolls; the 15-second tour clip cut three distinct shots inside one generation. Confirmed first-frame identity locking and native speech work in the same call. Film published to Instagram via the social pipeline.