---
name: ai-film-shot-board
description: Build a canonical shot board for a short AI-generated film — shot list, per-shot image and motion prompts, character reference sets, and the production rules that stop generations being wasted. Use when planning, prompting, or reviewing an AI short film, demo film, ad, or music video.
---

# AI Film Shot Board — template and method

A working method for planning and generating a short film with AI video tools,
and a blank board to fill in. Built from a real 60-second production; every rule
in here exists because breaking it cost a generation.

---

## How to use this

**Three ways, same file.**

1. **Paste it into any AI** — Claude, ChatGPT, Gemini, whatever you use — and say
   *"use this as a template for my film."* The AI will interview you and hand
   back a filled board.
2. **Install it as a Claude skill** — it is already formatted as one. Then just
   ask for a shot board.
3. **Fill it in by hand.** Everything below the interview is a blank form.

---

## For the AI reading this

You are helping someone plan a short AI-generated film. Follow these rules.

**Interview first. Do not invent.** Ask the questions in Part 1 before writing a
single shot. If the person gives you a one-line brief, that is the start of a
conversation, not a spec. Ask in small batches — three or four questions at a
time, not twenty.

**Never invent a fact about their product, company, customers or story.** If you
need a detail they have not given you, ask. A board full of plausible invented
specifics is worse than a board with gaps, because they will generate from it.

**Enforce the rules in Parts 4 and 5 on every prompt you write.** They are not
style preferences. Each one exists because a generation was lost to breaking it.
If the person asks for something that breaks one, say so and explain the cost —
then do it their way if they still want it. They are the director.

**Fill the board completely, or mark the gap.** Every field gets real content or
an explicit `[TBD — need X from you]`. Never leave a placeholder looking like an
answer.

**Output the board as one document** they can keep, in the structure of Part 2.
If they want it as a web page, a doc, or a file, build it.

**Flag contradictions.** If shot 3 says he holds the phone in his right hand and
shot 4 says left, say so. Continuity errors are the main thing this board exists
to prevent.

---

## PART 1 — The interview

Ask these before writing anything. Skip what they have already told you.

### The film

1. **What is it for, and who watches it?** A sales demo, a brand film, a music
   video, a proof of concept, a pitch. Who is in the room when it plays.
2. **How long?** Total runtime in seconds. Be exact — it drives everything below.
3. **What has to be true by the end?** The one thing the viewer should believe or
   feel. Not a list. One thing.
4. **What's the story in three beats?** Setup, turn, payoff. If they can't say it
   in three, the film is too big for its runtime.
5. **Is there dialogue or voice-over?** And is anyone speaking *on camera* — a
   visible moving mouth — or does every line run over a cutaway?
6. **What's real and what's generated?** Screen recordings, stock, live footage,
   fully generated. Mark each shot.

### The look

7. **Frame and format.** Aspect ratio, and whether it's a wide cinematic crop
   inside a standard frame.
8. **One sentence of visual grammar.** Lens, light, grade, grain. This becomes
   the style bible and goes into every prompt.
9. **How are places or groups told apart?** If the film cuts between locations or
   characters, colour and light should do that work, not dialogue. Assign a
   palette per location.

### The people

10. **Who appears on camera?** For each: name, role, age range, build, and the
    wardrobe they wear in *every* shot.
11. **Whose face carries the film?** That person needs the strongest reference
    set and gets generated first.
12. **Does anyone need a specific expression or action** the model will struggle
    with — a held look, speech, a precise hand movement?

### The tools

13. **Which image tool and which video tool?** And is the image tool cheaper to
    iterate in than the video tool? (It usually is.)
14. **What does a generation cost, and is there a free or cheap tier?** See Part 6
    — this decides where iteration happens.
15. **What's the reference limit per generation** on the video tool? It varies a
    lot by model and it changes what's possible.

---

## PART 2 — The board

> Fill every field. This document is canonical: **nothing gets generated from a
> prompt that isn't on it.** If a prompt lives anywhere else and disagrees, this
> board gets fixed — not the other way round.

### Vitals

| | |
|---|---|
| **Title** | `[working title]` |
| **Runtime** | `[  ]s` |
| **Shots** | `[  ]` |
| **Generated clips** | `[  ]` |
| **Real / captured** | `[  ]` |
| **Frame** | `[e.g. 2.39:1 inside 16:9]` |
| **Faces on camera** | `[yes / no / which shots]` |
| **Iterate in** | `[cheap model]` |
| **Finish in** | `[quality model + resolution]` |

**One-paragraph premise.** `[What happens, in plain language, as if telling a
colleague. No jargon, no feature names.]`

---

### Shot list

Every shot, with its time, in order. Fill this before writing a single prompt —
it is where you find out the film doesn't fit.

| # | Time | Dur | Kind | Who | Location | Beat | References |
|---|---|---|---|---|---|---|---|
| 01 | `0:00–0:0_` | `_s` | `AI / Screen / Card` | `[who]` | `[where]` | `[what happens]` | `[which refs]` |
| 02 | | | | | | | |
| 03 | | | | | | | |
| 04 | | | | | | | |
| 05 | | | | | | | |

*Add rows as needed. Durations must sum to the runtime — check it.*

**Act structure**

| Act | Time | Beat |
|---|---|---|
| One — `[name]` | `0:00–0:__` | `[what this act does]` |
| Two — `[name]` | | |
| Three — `[name]` | | |

---

### Dialogue

> **Record every line separately and cut the picture to the audio**, never the
> other way round. You want the performance in the voice, not whatever a
> generated mouth happens to imply.

If a line runs across a cut, split it on an em-dash — the sentence continuing
over the edit is what sells it as one take rather than several line readings.

| Shot | On camera? | Line |
|---|---|---|
| `__` | `yes / no` | `"[line]"` |
| | | |

---

### Cast

Repeat per character.

**`[Name]`** — `[role, and what they do in the film]`

- **Age and build:** `[  ]`
- **Wardrobe, identical in every shot:** `[full description — every garment, every accessory]`
- **Palette:** `[which location palette they live in]`
- **Why this wardrobe:** `[how it separates them from their background and from the other characters]`
- **Props that are NOT wardrobe:** `[things they carry sometimes — keep these out of reference stills]`

**Palettes**

| Palette | Who / where | Look |
|---|---|---|
| `[name]` | `[  ]` | `[light, colour, texture]` |

---

### Style bible

Pasted into every image prompt. Write it once, never vary it.

```
[One block: format and framing, camera and lens, depth of field, light,
grain, grade, and the negatives — no lens flare, no text, no logos, no
watermark, no on-screen UI.]
```

**Do not prepend this to character reference stills.** References want flat,
even, boring light. A reference generated in your cinematic grade bakes that
grade into every scene you pull from it.

---

### Shot panels

One per generated shot. Copy the block.

---

#### Shot `__` — `[name]`
**`__`s · references: `[list]`**

**What happens.** `[Plain description of the beat — what the audience sees and
why it's in the film.]`

**Line, if any.** `"[dialogue]"`

**Upload:** `[exact reference filenames]`

**Image prompt — run as-is**

```
IDENTITY — [style bible, verbatim]

CAST — [who is in frame, and the instruction to take identity from the
uploaded references and not redesign them]

WARDROBE — [verbatim from the cast sheet]

LOCATION — [where, what light, what's behind them, which palette]

FRAME — [shot size, camera position, what each hand is doing, what is
entering or leaving frame, crop points]

EXPRESSION — [stated separately so it can't get buried]

NEGATIVE — [what must not appear: other people, text, signage, props,
accessories, wrong wardrobe]
```

**Motion prompt**

```
ACTION — [what moves, described relative to the camera, with timings if
the shot has beats]

HELD — [what must NOT move. Naming the still thing is as useful as naming
the moving one.]

SFX — [direction, not the final mix — it sharpens the motion the model
produces]

NEGATIVE — [no cuts, no zoom, no orbit, nobody else in frame — whatever
applies]
```

**Notes.** `[Anything the generator needs to know: what failed last time, what
to check before spending a credit, what to do if it doesn't work.]`

---

## PART 3 — Character references

The single biggest cause of a film falling apart is a face that drifts. A
character who changes between two shots reads as two different people and breaks
the story worse than any rejected generation.

### The three habits that actually drive consistency

The reference sheet is a document. These are what keep a character consistent:

1. **One canonical hero portrait per character**, uploaded on every generation.
   **Never chain off the last image you made** — drift compounds generation over
   generation. Always go back to the original.
2. **A fixed identity block**, pasted verbatim into every prompt. Not paraphrased.
   A reworded description is a different person to the model.
3. **Person and setting described separately.** The identity block never changes;
   everything below it does.

### The reference set

Generate a still per view. Neutral light, plain background, no scene.

| # | View | Why you need it |
|---|---|---|
| 01 | **Front, neutral** | The casting. Run it 4–5 times, pick one, discard the rest. Everything else is generated *from* this file. |
| 02 | Three-quarter | Most scene angles |
| 03 | Profile | Any side-on shot |
| 04 | Full body | Anything showing the whole silhouette, and the source for hands and back views |
| 05 | Back | Shots from behind |
| 06 | Seated | If they sit in the film |
| 07 | Hands | Hands-only shots — generate from the full body, not the portrait, so sleeves carry over |
| 08 | **Expression** | Any specific look the film needs |
| 09 | **Speaking** | Only if a mouth moves on camera. Models animate a mouth far better having seen it open. |

**Expression and speaking references leak.** Load them only for the shots that
need them. Leave one in a default set and that look appears in every frame the
character shows up in — *unless* a shot deliberately travels between two
expressions, which is the one case for loading both.

### Wardrobe discipline

- Every accessory has to earn its place. Cut anything doing no story work — it is
  one more thing to drift, duplicate or vanish.
- **Props are not wardrobe.** A hat someone carries but never wears goes in the
  scene prompt, never in a reference. Put it in a reference and it appears in
  every shot.
- **New wardrobe = a new named variant** with its own hero portrait and its own
  frozen identity block. Ad-hoc wardrobe edits are how one character quietly
  becomes two.

### QA before any reference is used

Open the front, three-quarter and full body side by side at full size:

- [ ] Same face — eye spacing, nose, jaw, hairline
- [ ] Same hair, same length, same way it sits
- [ ] Facial hair at the same length (models quietly shave and beard people)
- [ ] Every garment the same colour — colour drift is the second most common
      failure after face drift
- [ ] Accessories present, correct, and not duplicated
- [ ] Age reads the same in all three (models quietly de-age people)

Then check the expression and speaking stills against the front separately —
identity slips most when the model is changing a mouth.

**Regenerate the one still that fails, not the set. And never build a reference
off a generated scene image** — always go back to the original files.

---

## PART 4 — Writing a prompt that doesn't waste a generation

### Labelled blocks, not prose

Use the block structure in the panel template — IDENTITY, CAST, WARDROBE,
LOCATION, FRAME, EXPRESSION, ACTION, HELD, SFX, NEGATIVE. Scannable, diffable,
and much harder to leave a coin flip in than a paragraph.

Two earn their place beyond the obvious:

- **HELD** names what must not move. *"The phone never lifts"* is why the phone
  survives the shot.
- **EXPRESSION** is stated separately so it can't get buried. A
  walking-toward-camera shot is a product reveal or a character beat depending on
  that one block alone.

### The rule that costs the most when broken

**Describe motion relative to the camera, never relative to the world.**

*"Outward"* is meaningless to a model that has not decided where inside and
outside are. *"Away from camera"* cannot be misread. Never write "forward",
"back", "out" or "in".

### The six coin flips to check before any prompt goes out

1. **Direction** — toward camera, away from camera, or across frame left to
   right. Never a world-relative word.
2. **Hinge and swing** — doors, laptops, lids, car doors. Name the hinge edge and
   which way it opens *relative to camera*. Say what it does **not** do.
3. **Which hand** — and does it match the shot before? Remember the mirror:
   facing camera, someone's left hand reads at frame right.
4. **Camera side** — which shoulder you're behind, which side of the vehicle
   you're on. Get it wrong between two shots of one scene and the geography flips
   on the cut.
5. **What must NOT move.**
6. **What is not in frame** — crop points, and anything that must stay out: a
   second person, a head entering from above, text, signage, a watch.

### Continuity locks

Write these down before generating anything, and never vary them:

- **Handedness.** Which hand holds the thing, across every shot it appears in.
  Change it in one shot and you change it in all of them.
- **Direction of travel.** If they exit frame right, they enter the next shot
  frame left.
- **Time of day and light direction**, per location.

### Beats inside a shot

If a shot has an emotional arc rather than one state, write it as timed beats in
the ACTION block:

```
0.0–1.0s — [beat one]
1.0–2.0s — [beat two, and say which one is the turn]
2.0–3.0s — [beat three]
```

**Anchor the still on beat one, not the payoff.** Image-to-video starts from the
first frame. Anchor on the emotional destination and the shot has nowhere left to
travel — you get a flat, held expression, which is the most common failure in
generated performance.

### Patch, don't regenerate

When a still is wrong in one detail, upload it on its own and ask for the single
change:

> *"Keep this image exactly as it is — same person, same wardrobe, same location,
> same lighting, same composition, same grain and grade. Change one thing only:
> [the fix, in camera-relative language]. Nothing else about the image changes."*

You keep the frame you already approved. If it drifts the face or the light while
fixing the detail, fall back to a clean generation.

### Durations do not go in prompts

Duration is a setting, not text. Putting it in the prompt just gives the model
something to argue with. **Generate long, never short**, and trim in the edit so
you have handles to cut on.

---

## PART 5 — Before you spend anything

### The thirty seconds that saves the credit

Stills are cheap or free. Video is not. Open the anchor still and check it
against its panel **before** it goes to the video tool:

- [ ] Right person, and they match the reference side by side at full size
- [ ] Wardrobe correct, every item, right colours
- [ ] No accessories that shouldn't be there
- [ ] Props right
- [ ] Direction and handedness right, and consistent with the shot before
- [ ] Nothing cropped that shouldn't be
- [ ] No letterboxing baked into the file
- [ ] Nobody in the background who shouldn't be there

Every credit lost from here should be lost to motion going wrong — never to
something already sitting in a frame you approved.

### Prove the hard shots first

Identify the one or two shots most likely to fail — usually a lit, unobstructed
face; sustained on-camera speech; a precise hand action; or any shot with an
emotional arc. **Take those from reference to still to finished clip before
generating anything else.** Half an hour there tells you whether the whole
approach holds. The alternative is finding out after you've built everything else.

### Budget

Expect **8–12 generations per usable clip.** Plan for it. If that number is
frightening at your tool's prices, that is the signal to move iteration somewhere
cheaper — see Part 6.

---

## PART 6 — Where iteration happens

The single biggest cost lever on an AI film is **not generating the expensive
thing until you already know it's right.**

Fill this in for your own tools before you start:

| Stage | Tool / model | Max res | Cost per generation | Free or cheap tier? |
|---|---|---|---|---|
| Stills | `[  ]` | `[  ]` | `[  ]` | `[  ]` |
| Video — iterate | `[  ]` | `[  ]` | `[  ]` | `[  ]` |
| Video — finish | `[  ]` | `[  ]` | `[  ]` | `[  ]` |

**Three things to check that people usually get wrong:**

1. **A platform is not a model.** Most video platforms host several models —
   their own and third-party ones — priced very differently. Find out which one
   you are actually running.
2. **Cheap or unlimited tiers often cover only the platform's own models.**
   Third-party models usually burn credits no matter what your plan is.
3. **Resolution caps differ by model.** The cheap model may not reach your
   delivery resolution — which is fine, as long as you iterate there and finish
   somewhere else.

**The pattern that falls out of this:** get composition, expression, motion and
hands right in the cheapest thing that renders, then spend one expensive
generation on the take you have already approved. **Generating iterations at low
resolution in an expensive model is the worst of both** — you pay a premium for a
resolution the cheap model gives away, and cap your master anyway.

**Reference limits are per model, not universal.** One model takes three
references, another takes thirty. If yours takes thirty, load the whole character
set — under-referencing is a leading cause of face drift.

---

## PART 7 — Reviewing what comes back

You cannot judge a generated clip by watching it once. Pull frames.

**Per clip, check:**

- [ ] **Identity** — put a frame beside the reference at full size. Not "close
      enough." The same person.
- [ ] **Wardrobe** — every item, every colour, nothing added (watches and rings
      appear out of nowhere)
- [ ] **The beat** — does the thing the shot exists for actually happen, and does
      it happen *inside the part you'll use*? A gesture that lands at 6.8s of a
      7s clip is gone the moment you trim to 4s.
- [ ] **Direction and handedness** against the shot before and after
- [ ] **Background** — anyone or anything that shouldn't be there
- [ ] **Framing** against what the panel specified
- [ ] **Motion quality** — warping, morphing, jitter, extra fingers. This lives
      *between* frames, so sampled stills can hide it. Watch for it.

**Keep a render log.** One row per generation: shot, version, tool, resolution,
duration, verdict, and what specifically was wrong. Without it you regenerate the
same mistake.

| Shot | Ver | Date | Tool / res | Dur | Verdict | Notes |
|---|---|---|---|---|---|---|
| | | | | | | |

---

## The short version

If you remember nothing else:

1. **One canonical board.** Every prompt lives there. If something elsewhere
   disagrees, fix the board.
2. **Camera-relative language, always.**
3. **Faces come from a fixed reference, never a fresh text prompt.**
4. **Check the still before you spend on the video.**
5. **Iterate cheap, finish expensive.**
6. **Generate long, trim in the edit.**
7. **Record the voice separately and cut picture to it.**
8. **Prove the hardest shot first.**

---

*Template v1 · adapt freely · the rules are the useful part*