One script in, a produced short out.broll-pipeline

The problem, in business terms. A comms or newsroom team needs to ship short explainers on a daily cadence without burning out editors. Today that means a scripter, a voice artist, an editor cutting b-roll (the supporting footage laid over a narration), and a captioning pass, measured in hours per finished minute. The question a forward-deployed creative gets asked is not "can AI make a video," it is "can you turn our existing pipeline into one a small team runs before lunch, and can you prove what it costs."

The shape of the answer. One input file, input/script.md, holds the whole brief: one beat per block, each with a line of voiceover, a visual prompt, and a target duration. The orchestrator parses that into beats, fans the audio across the ElevenLabs modalities, renders or generates each beat's visual by its declared mode, then ffmpeg cuts everything to the voiceover's timing and muxes a toggleable caption track. Every stage caches its output keyed by a content hash, so a re-run only regenerates what actually changed. That is the difference between a demo and a workflow a team will keep using.

Model choices, and one honest boundary. The audio stack runs end-to-end through the ElevenLabs API: Text-to-Speech for narration, Eleven Music v2 for a licensed, commercially-cleared score, the Sound Effects API for accent beats, and Dubbing for the Spanish cut. Video generation sits behind a pluggable provider adapter: one thin module the pipeline calls per beat, routed today to a text-to-video model (Veo 3.1, via fal.ai); swapping providers is a config change, not a rewrite. And the beats that carry the pipeline's own evidence, the script, the cost manifest, the reroll command, the receipt, are not generated at all. They are rendered deterministically in code from the real artifacts, because for those shots the proof reads better than a prompt. Knowing which beats deserve a model and which deserve the actual data is the judgment call the role is about.

The manifest is the evidence. Every API call appends to output/run-manifest.json with model, params, cost, and latency; cached artifacts carry their original cost forward, so the receipt reflects the whole cut, not just the last re-run. The artifacts in this cut carry $9.51 of logged cost (the number the on-screen receipt bakes): $0.28 for the complete audio stack, $5.36 for the ten generated shots including the review board's retakes, $3.50 for creative direction, $0.37 for clip review. Getting to the shipped version added the board's extra re-rolls of one stubborn text-bearing shot, a full score re-render, the foley experiments, and the Spanish dub renders: $14.20 all-in, every line item logged, including the retries. The manifest is committed in the public repo, and beat twelve of the video is that manifest, rendered as a thermal receipt. That single file answers the customer's real question, "what does one finished minute cost," without a spreadsheet or a guess.

The creative council. Prompting shots one at a time produces b-roll that reads as disconnected stock, so generation is directed by a council of seven specialist personas (cinematographer, story producer, video researcher, sound designer, motion designer, music composer, audience-engagement expert) running in parallel on the Anthropic API. A creative director synthesizes their notes into one edit decision brief: a single visual system, a hook for the first two seconds, pacing built for short attention spans, and sound design motivated by what is on screen. After generation, the council's review board inspects frames from every generated clip against the brief, rejects unacceptable takes, and re-prompts and regenerates them. The ten shots in this cut share one visual system the council wrote, and several are retakes the review board demanded. Sound gets the same treatment: stems are gain-staged to measured loudness targets and mixed with the voice always on top, and a blind audio reviewer that can actually hear the cut scores the mix with timestamps until it is fit to ship. The pipeline also carries a per-shot video-to-audio foley stage (a model watches each shot's frames and scores the natural sound a live microphone would have caught); the sound-design judgment for this cut was to leave it out, because screen-heavy footage with little physical action does not motivate foley, and invented sound reads as artifact. Knowing when not to add sound is part of the craft.

Where a human still touches it. Script approval, brand and legal sign-off, voice selection, and a caption QA pass all stay human. Naming those seams is not a weakness to hide; it is what separates an enterprise-realistic workflow from a toy. The pipeline removes the mechanical hours and leaves the judgment where it belongs.

A forward-deployed creative is a creative solutions architect and an enablement lead, not a designer for hire. The job is to embed with a customer, understand their production pipeline, prototype an AI-native replacement, package it as a template, and teach it. That is the loop this artifact runs. It is also the loop I ran for eight years in live newsrooms and then at Google, building production AI enablement since 2024 for a community of 1,000+ senior engineers.