cnemaiGo to Studio

Guide · workflow

The AI video workflow, stage by stage

An AI video workflow for films, commercials and music videos: plan, cast, look, shoot, cut, review. What each stage needs, what it costs, where it breaks.

October 25, 2026

An AI video workflow is the order you make decisions in when the footage comes from video models instead of a camera. It runs in six stages: plan, cast, look, shoot, cut, and review. The models change every few months. The stages don't, and a workflow that holds them together is what turns five-second clips into a film someone will watch to the end.

This guide walks through each stage, with what it needs, what it costs at Cnemai's published prices, and the mistakes that sink most AI video projects. The numbers come from two films made start to finish in the studio: Face Masks Off, a 3:58 music video with 102 cuts, and Tony's Noodles, a 50-second commercial.

6stages, from idea to final cut
229clips rendered for one 3:58 film
102cuts on Face Masks Off
$5.58of renders for a 50 s spot

Why a clip generator isn't a workflow

A video model gives you a clip: a few seconds, one prompt, one try. A film needs twenty to sixty shots that agree with each other, with the same faces, the same light, the same product and a cut that lands on the beat or the line. Every problem in AI video lives in the gaps between clips, and that's exactly where a workflow does its work.

The six stages at a glance

StageWhat goes inWhat comes outTypical cost
1. PlanA script, a brief or a songScenes and shots, each with one action and a lengthIncluded
2. CastReferences for the people, places and propsOne locked sheet for each, with a hero versionA few stills
3. LookEach shot's line and its sheetsA first frame for every shot3¢ a still
4. ShootThe chosen first frameTwo or three takes per shot6¢ a second on the fast lane
5. CutThe takes, the voice, the soundA timeline with saved versions12¢ per 1,000 characters of voice
6. Review and deliverA proposed cutAn approved cut and a 1080p masterIncluded

1. Plan: turn the idea into shots

Start from whatever you have, a script, a client brief or a finished song, and break it into scenes and shots. Give each shot one action, one camera move and a length. One action per shot is the rule that matters most: a prompt that asks for three things comes back as a blur of all three.

A song brings its own grid. Read its sections, bars and drops first, then plan the shots on the bars, so the cut has somewhere to land. Face Masks Off was planned that way, 55 shots across a 3:58 song.

In Cnemai the script is the shot list: change a line and the shot changes, and every revision is kept. Paste a brief and the studio proposes scenes and shots with cast, staging and length, for you to keep or change.

A song read bar by bar: its sections, the beat grid, and a suggested cut every few bars
Planning on the song: sections, the beat grid and where to cut.

2. Cast: lock every face, place and prop

Consistency is the problem everyone hits first. The fix is the one film crews already use: casting. Make one sheet per character, place and prop, pick the hero version, and have every take draw from it.

Keep sheets to identity: the face, the wardrobe, the pack, the set, on a plain background. The action belongs to the shot. A sheet that tries to show a pose or a mood leaks it into every shot that uses it.

3. Look: find it on stills before you pay for video

A still costs cents and a second of video costs more than a still. Try three stills for a shot, keep the one that's right, and animate only that. The chosen still becomes the shot's first frame. When two shots need to flow, the last frame of one can open the next.

4. Shoot: takes, not tries

Render two or three takes of every shot and keep the best. Plan to shoot several times what you use: Face Masks Off rendered 229 clips, about 21 minutes of video, to keep 3:58; Tony's Noodles rendered 93 seconds to keep 52.

Two habits save money. See the price before every render; in Cnemai it shows on the take before it starts, and a failed take is never charged. And when a model keeps missing a shot, try the shot on another model rather than rewording the prompt a tenth time.

5. Cut: one timeline for picture and sound

Lay the takes on a real timeline: picture on the main lane, voice-over, dialogue, foley, ambience, score and captions on their own lanes. Trim to the bar or the read. Save a version whenever a pass is worth keeping. Face Masks Off is on version 61, and all 60 before it are still there to restore.

The Cnemai timeline on Face Masks Off: takes on the Main lane, the song's waveform on Score, lyrics below
Face Masks Off mid-cut: takes on Main, the song on Score, every lyric timed below.

6. Review and deliver: decide on the record

On a team, the review round is part of the workflow, not an email thread after it. Branch the cut to try an ending, propose it, let every member vote and leave notes, compare it against the live cut, and have the producer approve it. Then export a 1080p master of the approved cut. Solo, the same history matters for a different reason: it's how you get back to the version that worked.

Where AI video workflows break

  • Tool-hopping. The script in a doc, stills in one app, clips in another, the edit in a fourth, the review in email. Each handoff loses a prompt, a reference or a version.
  • Prompting instead of casting. Describing the hero again in every prompt drifts. A locked sheet doesn't.
  • Overloaded shots. Three actions in one prompt make one muddy clip.
  • One take per shot. The edit has nothing to choose from.
  • No versions. "final_v3_REAL.mp4" isn't a history.

One studio, or a chain of apps

You can run this workflow across separate tools: a doc for the script, an image model, a video model, a voice app and an editor, joined by downloads and uploads. It works, and it's how most people start. Cnemai runs all six stages on one project instead, so the shot, its sheets, its takes, its sound and its place on the cut stay connected, and the team works in the same place at $99 a month for the whole studio.

Directing it from a chat

The workflow doesn't have to live in one app's buttons. Through the Cnemai connector, Claude or ChatGPT can open the same project: read the script and the cast, render takes, voice a line and lay clips on the cut, and every change lands in the studio as a saved version. Tony's Noodles was directed that way, from a Claude chat.

Workflows by kind of film

Questions

What is an AI video workflow?

The order you make decisions in when the footage comes from video models: plan the shots, cast the people and places, find the look on stills, render takes, cut, then review and deliver. The tools change every few months; the stages don't.

How many takes should I render per shot?

Plan on two or three, and expect to render several times what you keep. Face Masks Off rendered 229 clips, about 21 minutes, to keep 3:58. Tony's Noodles rendered 93 seconds to keep 52.

Can an AI video workflow be fully automated?

Pieces of it can: shot lists, prompts, renders and rough assembly. Choosing the take, the read and the cut point is still the director's job, and it's what separates a film from a pile of clips.

Which tools do I need?

Something to plan shots, image and video models, a voice, and an editor with versions and review. You can chain separate apps, or use one studio like Cnemai that keeps every step on the same project.

Can I run the workflow from ChatGPT or Claude?

Yes. Through the Cnemai connector an assistant can read the script and cast, render takes, voice lines and lay clips on the cut, and all of it lands in the studio where the team sees it.

Keep reading

Your film starts with one line.

Cnemai is opening by invitation. Leave your email and your invite arrives in your inbox. 14 days free, with $100 of renders on us.

Then $99 a month for the whole studio