ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›THE CRAFT›PLAN BEFORE YOU GENERATE
THE CRAFT · METHODINTERMEDIATE

Plan before you generate: AI storyboard and shot list template

Block the camera in 3D, lock the faces, then generate. The habit that cuts attempts — which is the only thing that cuts cost.

READ14 min
WORDS2,756
SECTIONS8
TYPEGUIDE
CHECKED6 SEP 26

To plan an AI video before you generate it, block the camera and the action somewhere cheap first (rough 3D boxes, a storyboard and a shot list), lock your characters in a reference sheet, and only then spend credits, because every re-roll costs the full clip. This page walks through that order and gives a shot-list template to copy.

Three separate AI video tutorials, three different tools, one idea underneath all of them: make the expensive decisions somewhere cheap, before you spend a credit.

TL;DR — THE SHORT VERSION
  • Every re-roll costs the full clip. So the goal is not a better prompt — it is fewer attempts.
  • Block the camera in 3D first. Rough boxes are enough. Moving a camera in a 3D scene is free; regenerating a shot is not.
  • Then generate over that render — video-to-video, so the previz drives the camera path and the model only dresses the world.
  • Lock your characters before any scene, in a reference sheet, or faces drift between shots.
  • The face trick worth stealing: keep the generated body and costume, paste the real photo into the portrait panel.
  • Connectors are model-agnostic. The endpoint is the thing; whichever assistant you use is a detail that changes monthly.
IN PLAIN ENGLISH

Film crews do not work out where the camera goes while the actors stand around and the lights burn. They storyboard, they walk it through, they block it — because the expensive part is shooting.

Generation put the expensive part back. Every attempt costs money and minutes, so the old habit becomes valuable again: decide it somewhere cheap first.

Why this beats better prompting

01

The arithmetic nobody runs

Video is billed by duration, so a re-roll costs the full length again. A thirty-second shot that takes five attempts costs you two and a half minutes of generation to deliver thirty seconds.

ONE SHOT · FIVE ATTEMPTS
The worked example above, drawn to scale: every re-roll is billed at the full clip length.
Generation you pay fortwo and a half minutes
Video you deliverthirty seconds
Reasoning — the arithmetic in this page’s first section, from how duration billing works; not a measured result, page checked 6 Sep 2026.

Now split those attempts by cause. Some are aesthetic — the light is wrong, the mood is off. Many are structural: the camera went the wrong way, the subject left frame, the cut landed early, two things moved when one should have. Structural failures are decidable in advance. They do not need a model to resolve them, and paying a model to resolve them by chance is the expensive way.

See what generation actually costs for the full sum. The short version: your real unit is cost per usable result, and previz attacks the multiplier rather than the price.

Step one: block it in 3D

02

Grey boxes are enough

This is the part people assume requires being a 3D artist. It does not, because nothing you build here appears in the final video. A corridor is a box with walls. A person is a capsule. What matters is where they are, where the camera is, and when each thing moves.

Higgsfield's own Blender workflow, published 28 August 2026, does exactly this: a plugin plus a bridge connector lets an assistant build the blocking inside Blender — a textured corridor, a proxy character, a continuously animated camera — and the viewport render then drives the generation.1

What to settle in the blocking: camera position and path, subject placement, what enters and leaves frame, the timing of each beat, and where cuts fall. That list is most of what goes wrong.

03

Then hand the render to the model

Export the viewport render as an ordinary video file — 1920×1080, 24fps is the setting their workflow uses, verbatim "Render the final blocking as a video file (1920x1080, 24fps, MPEG-4 viewport render) and attach it together with the character and location references" — and feed it as a video-to-video pass with your character and location references attached.

The important consequence: the previz drives the camera one-to-one, and the model's job shrinks to dressing the world. You stop asking it to invent choreography and start asking it to paint over choreography you already fixed. Fewer decisions in its hands, fewer ways to be wrong.

This is not Higgsfield-specific. Any 3D tool that exports video and any model that accepts a video input does the same job. The plugin removes friction; it is not the technique.

TAKEAWAY

Prompting decides what things look like. Blocking decides where they are and when they move. Trying to do the second with words is the single most expensive habit in AI video.

Step two: lock the people before the scenes

04

Reference sheets first, always

The same pattern, applied to identity. If a character appears in twelve shots, generating them fresh each time gives you twelve near-relatives. Build a character sheet first — three views on a plain background, evenly lit — and attach it to every scene.

Their love-story tutorial, published 21 August 2026, front-loads all of it: character sheets, location plates and a scale reference built before any scene is generated.2

Write the lock as anatomy, not beauty. Their prompt specifies bone structure, eye shape, brow line, jaw, hairline — and explicitly says do not reshape, slim, symmetrize or beautify. That negative is doing real work: models default to making faces prettier, which is precisely how identity slips.3

05

The hybrid portrait

The one technique here that is not obvious. Generate the sheet, keep the costume and body the model produced, then paste the person's real photograph into the portrait panel of the sheet.

Reference sheets get you most of the way and stall at "convincing lookalike". Feeding a real face back in for the close panel is what closes the gap. It costs nothing and it is the difference between a character who resembles someone and one who reads as them.

Everything else on identity — seeds, reference images, per-model behaviour — is on character consistency.

Step three: let it plan before it renders

06

What a connector actually gives you

The third workflow connects a generation service to an assistant over MCP, so the assistant writes the prompt, sends it, and the result comes back in the conversation. Higgsfield publishes an endpoint at higgsfield.ai/mcp and describes it as working with any agent — their own page lists ChatGPT, Claude, Grok and Cursor, and names more besides.4

That discrepancy is the useful lesson. The model named in any of these tutorials is the least durable fact in them. The connector is the thing; which assistant sits on top changes every few months, and a workflow built on one model version dates at that model's pace.

07

Where planning-first stops helping

It does not fix aesthetics. Blocking settles geometry and timing. Whether the light is beautiful, the grade is right, or the performance lands is still generation, still iterative, still where your remaining attempts go.

It adds a step. For a single short shot, previz is slower than just generating twice. The break-even arrives with length, with camera movement, and with anything that has to be consistent across shots.

And it is not free of the usual limits — hands, text, crowds, physics. See why your generation came out wrong for which failures respond to prompting at all.

DECIDE IT CHEAP · THEN GENERATE
The structural decisions come first, where changing them costs nothing.
Block it in 3DGrey boxes: camera position and path, subject placement, the timing of each beat, where cuts fall.
Lock the people and placesCharacter sheets and location plates, built before any scene is generated.
Generate over the renderVideo-to-video with the references attached. The previz drives the camera; the model dresses the world.
Spend what is left on aestheticsLight, grade and performance are still generation, and still iterative.
Reasoning — summarises this page’s three steps and its limits section. The workflows are Higgsfield’s vendor tutorials, page checked 6 Sep 2026.
A NOTE ON THE NUMBERS IN TUTORIALS LIKE THESE

The marketing-studio piece behind part of this page is framed around income — a figure per day, a figure per month, a price band per fifteen seconds of video. None of it is measured, and none of it is repeated here.

What people advertise charging is not what you will earn, and this site removed its own versions of exactly those figures in August. If you want the reasoning, it is on how to make money with AI content, which now separates asking prices from results and says which is which.

Step four: storyboard it, one frame per shot

08

A storyboard is the cheap version of the film

A storyboard is a row of still pictures, one per shot, that shows what each shot looks like before anything moves. It is the 2D cousin of the 3D blocking above: quicker to make, less exact about camera paths, and good enough for most short pieces.

With AI it has a second job. A storyboard frame you are happy with can become the first frame of the shot, and the next panel can become its last frame, so the frames you planned are the frames the model must hit. How that works, and which models take a start and an end image, is on first frame and last frame in AI video.

ByteDance’s own Seedance guide asks for exactly this: attach the planned frames and write the prompt to match them. 5 Google’s Veo 3.1 guide says reference images help “maintain a consistent aesthetic across multiple shots.” 6

How to make one quickly: write the shot list first (below), then generate one image per row with the same character sheet attached every time. Keep the frames rough. You are checking the order, the framing and the continuity, not the final look.

USE WHEN: turning one shot-list row into a storyboard frame (character sheet attached)
Storyboard frame for shot [NUMBER] of a short film. [SHOT SIZE, e.g. medium close-up], [CAMERA ANGLE, e.g. eye level]. [SUBJECT from the attached character sheet] [ACTION at the START of the shot]. Location: [PLACE], [TIME OF DAY]. Light: [LIGHT, e.g. soft window light from the left]. Aspect ratio [16:9 / 9:16]. Keep the character’s face, hair and costume exactly as on the character sheet. Clean, readable composition; plain cinematic still, no text, no captions, no panel borders.

The shot-list template

09

One row per shot, nine columns

A shot list is the written version of the storyboard: one row per shot, with everything a generation needs decided in advance. Fill it in before you spend anything. Each row then becomes a prompt almost word for word.

  • Shot: the number and a two-word name (“3 · the reveal”).
  • Size: wide, medium, close-up, extreme close-up. The words are on camera grammar.
  • Subject: who or what is in frame, named the same way in every row.
  • Action: the one thing that happens, from the first frame to the last.
  • Camera move: one move, or “locked”. ByteDance’s guide gives the same rule. 5
  • Lens: wide, normal or long is enough detail for most prompts; add a focal length (24mm, 35mm, 85mm) if you want to be exact.
  • Light: where it comes from, what colour, how hard. Keep it the same across shots in one scene.
  • Duration: in seconds, within what your model offers. ByteDance suggests a default for Seedance. 5
  • Reference: which character sheet, location plate or storyboard frame goes with the shot, and whether it is a first frame, last frame or style reference.

Copy the blank template into a spreadsheet, a document or a chatbot. The filled-in example under it shows the level of detail that is enough.

USE WHEN: you want a blank shot list to fill in (paste into a spreadsheet or document)
| Shot | Size | Subject | Action | Camera move | Lens | Light | Duration | Reference | |------|------|---------|--------|-------------|------|-------|----------|-----------| | 1 · | | | | | | | s | | | 2 · | | | | | | | s | | | 3 · | | | | | | | s | | | 4 · | | | | | | | s | | | 5 · | | | | | | | s | |
USE WHEN: you want to see a filled-in example (a four-shot scene)
| Shot | Size | Subject | Action | Camera move | Lens | Light | Duration | Reference | |------|------|---------|--------|-------------|------|-------|----------|-----------| | 1 · arrival | Wide | MAYA (sheet A) at an empty bus stop | Maya walks into frame and stops at the shelter | Locked | 24mm | Overcast dusk, cool blue; warm street lamp on the right | 6s | Location plate L1; storyboard 1 as first frame | | 2 · waiting | Medium | MAYA | She checks her phone, then looks up the road | Slow push-in | 35mm | Same as shot 1 | 6s | Sheet A; storyboard 2 as first frame | | 3 · the reveal | Close-up | MAYA | Her expression changes from bored to surprised | Locked | 85mm | Same, plus headlights sweeping across her face from the left | 4s | Sheet A; storyboard 3 as first frame, storyboard 4 as last frame | | 4 · the bus | Wide | Old green bus, MAYA | The bus pulls in and the doors open | Locked | 24mm | Same as shot 1, plus warm light from inside the bus | 8s | Location plate L1; storyboard 4 as first frame |
USE WHEN: you have a script or idea and want a chatbot to draft the shot list for you
Turn the scene below into a shot list for an AI video model. Give one row per shot in a table with exactly these columns: Shot, Size, Subject, Action, Camera move, Lens, Light, Duration, Reference. Rules: one action and one camera move per shot (or "Locked"); give each recurring character the same name every time; keep the light consistent within the scene; durations between [MIN] and [MAX] seconds; in Reference, say which character sheet or location image each shot needs, and where one shot’s last frame should be the next shot’s first frame. Do not add shots that are not in the scene. Scene: [PASTE YOUR SCENE OR IDEA]
TAKEAWAY

If a column is empty, the model fills it with its own average. A full shot list is the cheapest way to decide everything before anything is billed.

A storyboard character-sheet prompt

10

Make the sheet once, attach it to every frame

Step two above explains why the character sheet comes first. This is a copyable version to start from. It uses the same idea as the vendor prompt quoted there: describe the face as anatomy and forbid beautifying, because models drift toward a prettier, more average face when nothing stops them.

For locks on outfits, objects and camera as well as faces, use the fuller set on identity lock prompt templates. The wider method is on character consistency.

USE WHEN: making the character sheet that every storyboard frame and shot will use
Character reference sheet for [NAME], for use in a storyboard. Plain light-grey background, soft even studio light, no shadows, no props, no text. Three full-body views side by side (front, three-quarter, back) plus one head-and-shoulders close-up. Same person in every view. Identity lock: [AGE], [BUILD], [FACE SHAPE], [EYE SHAPE AND COLOUR], [BROWS], [NOSE], [LIPS], [JAWLINE], [HAIRLINE AND HAIR: colour, length, texture, parting], [SKIN TONE], [ANY MARKS OR FEATURES]. Costume: [EACH GARMENT with colour and material], [SHOES], [ACCESSORIES]. Do not reshape, slim, symmetrize or beautify the face. Keep natural asymmetry. Neutral expression. Natural proportions.

The pre-generation checklist

1 — Do I know where the camera goes, or am I hoping the model picks well?□
2 — Is the timing of each beat decided, or will I find out on playback?□
3 — Does every recurring character have a locked reference sheet?□
4 — Does the identity prompt describe bone structure, and forbid beautifying?□
5 — Am I driving this with a video input, or asking for choreography in words?□
6 — For a short single shot: is previz actually cheaper here, or am I adding a step?□
7 — Does every shot have a row in the shot list, with one action and one camera move?□
8 — Do I have a storyboard frame for where each shot starts, and where it must end?□

Related on this site: What Is a Keyframe?

SOURCES AND HONESTY ABOUT THEM
  1. 1Higgsfield, "This Blender + Higgsfield AI Workflow Changes How You Make AI Video", by @adilinthewild, dated Aug 28, 2026 on higgsfield.ai. Read at source 11 Sep 2026, verbatim: "Builds the final 30-second one-take previz in Blender through Higgsfield Bridge: a textured corridor, a proxy character, and a continuously animated camera — one smooth take, no cuts," and "Five use cases that will completely change your AI video creation process." Vendor tutorial, described here rather than verified by this site: the workflow, the render settings and the five worked examples are theirs. The reasoning about why previz reduces attempts is this page's, not a measured result. Checked 6 Sep 2026, re-read 11 Sep 2026↑
  2. 2Higgsfield, "How to Create AI Love Stories (Full Tutorial)", by @adilinthewild, dated Aug 21, 2026 on higgsfield.ai. Read at source 11 Sep 2026, verbatim: "Build your assets first — the leads, the giant and the locations repeat in every scene, and locked reference sheets are what keep them consistent," and on the hybrid portrait, "the sheets get you 90% of the way, and the last 10% is a hybrid — keep the generated costume and body, then paste the person's real photo onto the portrait panel." Vendor tutorial. The identity-lock language and the hybrid-portrait technique below are quoted in substance from it and are not independently tested here. Checked 6 Sep 2026, re-read 11 Sep 2026↑
  3. 3Verbatim from that tutorial's character-sheet prompt, read at source 11 Sep 2026"Identity lock, highest priority: his exact face geometry, bone structure and natural slight asymmetry, dark brown almond-shaped eyes, straight dark brows, his nose shape, lips, defined jawline, ears, light warm skin tone, his hairline… Do not reshape, slim, symmetrize or beautify." ↑
  4. 4Higgsfield MCP page, read at source 11 Sep 2026, titled "Higgsfield MCP | AI Image & Video Generation for Any Agent". Its FAQ, verbatim: "Higgsfield MCP works with Claude (web, Cowork, and Claude Code), OpenClaw, Hermes Agent, and NemoClaw. Any agent or client that supports MCP can connect to Higgsfield." Corrected 11 Sep 2026: this citation previously said "one video names an OpenAI model, the companion article names an Anthropic one." Not verified as written — what is verifiable at source is that the MCP page's own headline is "Add Higgsfield plugin to ChatGPT" while both tutorials behind this page drive the work with Claude ("Claude builds the blocking for you"; "The Claude Skill below writes the video prompts for you"). Same endpoint, different assistant named in each place.↑
  5. 5Dreamina (ByteDance), Seedance 2.5 prompt guide, 18 Aug 2026, read at source 12 Oct 2026“Storyboard parity: attach beat-by-beat frames and mirror them in the prompt so you can check how closely the output follows the plan.” ↑a“Put shot size and camera move before lighting and mood; use one move per shot”. ↑b“Beat math: 6–8 seconds per beat is usually a solid default, with 3–4 seconds saved for the resolution.” ↑c
  6. 6Google Cloud Blog, The ultimate prompting guide for Veo 3.1, 16 Oct 2025, read at source 12 Oct 2026“Provide reference images of a scene, character, object, or style to maintain a consistent aesthetic across multiple shots.” ↑

Higgsfield published tutorials: the Blender workflow (28 Aug 2026), the love-story tutorial (21 Aug 2026), and the MCP endpoint page. All read on 6 September 2026.

These are vendor tutorials and this page treats them as such. The workflows, settings and prompt techniques are described as theirs, with dates. There is no timing, no measured credit saving, and no claim that the method produces a particular quality of output. The argument that previz reduces attempts is reasoning from how the billing works, not an experiment.

No income figures, view counts or affiliate links appear on this page. The tutorials carry all three; none of them is verifiable, and the last is a disclosure this site is not currently in a position to make.

The through-line: you are not buying video, you are buying attempts. Everything above is a way of deciding something for free that you would otherwise pay a model to guess at.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE