# THE CINEMATIC PROMPT STARTER KIT

*Free from insertfunhere.com*

Most AI video looks like AI video for one reason: people write a **description**
when they should be writing a **shot**. A description gets you something that
moves. A shot gets you something that cuts.

This kit is the structure, the vocabulary, and the failure fixes. Works with
Seedance, Kling, Veo, Nano Banana, Seedream, Sora, Runway — anything that takes
a text prompt.

---

## PART 1 — THE ANATOMY OF A CINEMATIC PROMPT

Every good prompt has four parts, in this order:

```
1. STYLE LOCK      — the look. Same text every shot in the scene.
2. THE SHOT        — what physically happens, and when.
3. CONTINUITY      — what must not change.
4. NEGATIVES       — what must not appear.
```

Most people write only part 2. That's why their shots don't match each other.

**The style lock is the single highest-leverage thing in this document.** It's a
long paragraph describing camera, lighting, palette, grain, and rendering. You
write it once and paste it at the top of every prompt in the scene, unchanged.
Not similar — identical. The moment you paraphrase it, the look drifts.

---

## PART 2 — BUILD YOUR STYLE LOCK

Fill in each line. Keep every one — the specificity is the point.

```
[GENRE AND MEDIUM] — e.g. "Photoreal contemporary thriller" /
    "Premium 2D hand-drawn anime feature" / "1970s documentary"

[CAMERA AND LENS] — sensor, focal length, aperture feel, movement style.
    e.g. "Full-frame sensor, 35mm lens, shallow but not showy stop.
    Handheld or locked-off only."

[TIME AND LIGHT] — when it is, and what the ONE motivated source is.
    e.g. "Night interior. A single warm tungsten practical overhead is
    the only motivated source."

[SECONDARY LIGHT] — if any, and how it behaves.
    e.g. "Cold blue spill from an adjacent room that never mixes with
    the warm."

[BLACKS AND CONTRAST] — e.g. "Deep blacks that hold detail. No lifted
    matte, no crushed silhouette."

[TEXTURE] — e.g. "Fine natural film grain. Real skin texture with pores
    and shine."

[PALETTE] — name 4-6 actual colors. Not "moody." Not "cinematic."
    e.g. "Heather grey, faded olive, warm wood, black window glass."

[WHAT IT IS NOT] — e.g. "Nothing saturated, nothing teal-and-orange,
    nothing glossy. No crane moves, no lens flare, no haze."
```

**The last line is doing more work than you think.** Models default hard toward
commercial-looking output — teal and orange, haze, flares, glossy skin. You have
to name those and refuse them or you get them.

---

## PART 3 — THREE STARTER LOCKS

Use these as-is, or as templates for your own.

### NIGHT EXTERIOR — GROUNDED

> Photoreal contemporary drama, full-frame sensor, 40mm lens, shallow depth of
> field. Night exterior, wet asphalt. Practical sources only — sodium
> streetlight, headlights, a lit shop window — with hard falloff and no fill
> light. Deep blacks that hold detail. Desaturated and cold, leaning slate and
> cyan, with sodium amber as the only warm accent. Real skin texture with
> moisture and shine. Fine natural film grain. Handheld with restrained
> movement. Slightly loose headroom; the frame feels observed rather than
> composed. AVOID: teal-and-orange grade, lens flare, haze or fog machines,
> glossy commercial lighting, plastic skin, drone or crane moves, saturated
> color.

### GOLDEN HOUR — WARM AND WIDE

> Photoreal narrative feature, full-frame sensor, 28mm lens, deep focus. Late
> golden hour, sun low and directly behind the subject creating a strong rim and
> long shadows across the ground. Natural bounce fill only. Warm amber and honey
> highlights falling into cool blue shadow. Palette of dry grass, sun-bleached
> wood, faded denim, pale sky. Visible atmospheric depth from dust in the air,
> but no artificial haze. Fine film grain. Locked-off or slow steady dolly.
> AVOID: lens flare, overexposed blowout, orange-teal grade, HDR look, drone
> shots, saturated color, glossy skin.

### HANDHELD DOCUMENTARY

> Photoreal observational documentary, Super 35 sensor, 24mm lens, deep focus,
> available light only. Overcast daylight or fluorescent interior — flat, even,
> unflattering. Slightly cool and desaturated. Real unlit skin with visible
> texture and imperfection. Noticeable grain. Handheld with small constant
> correction, occasional soft focus hunt, imperfect framing. Subject sometimes
> cut off at the edge. AVOID: dramatic lighting, shallow depth of field,
> color grading, smooth gimbal movement, composed symmetry, cinematic bloom.

---

## PART 4 — CAMERA AND LIGHT VOCABULARY

Swap vague words for these.

**Instead of "close-up":** medium close-up (chest to top of head) · close-up
(shoulders up) · extreme close-up (eyes only, or one detail) · over-the-shoulder
· two-shot · single · profile · three-quarter

**Instead of "camera moves":** slow push in · slow pull out · lateral dolly left
· arc around the subject · locked-off, no movement · handheld with small drift ·
tilt down to reveal · whip pan · rack focus from foreground to background

**Instead of "well lit":** single source from above and slightly ahead ·
side-lit with the far side in shadow · backlit rim with no fill · underlit from
a screen · motivated by a practical in frame · soft wrap from a window

**Instead of "moody":** deep blacks holding detail · hard falloff · high contrast
with no fill · muted desaturated palette · one warm source against one cold
source, unmixed

**Instead of "cinematic":** name the lens, the grain, and the palette. "Cinematic"
means nothing to a model. "35mm, fine film grain, muted olive and warm wood" does.

---

## PART 5 — THE DIALOGUE TIMING RULE

The most common reason a clip's last line gets clipped off:

> **2.5 words per second, plus one full second of silence between lines, plus
> any silent action beat at the head and tail.**

Do the arithmetic before you generate. If the total runs past your clip length,
**split the shot** — don't compress. Compressing makes the model cram everything
into one breath and cut the ending.

Working ceiling: **about one short line per three seconds of runtime.**

Also state the count explicitly:

> EXACTLY 3 SPOKEN LINES EXIST IN THIS CLIP. No extra line, no added words. Each
> line separated by at least one full second of silence. No line begins until
> the previous has finished. Every line finished and audible before the 8-second
> mark.

Without that, models improvise extra dialogue.

---

## PART 6 — CONTINUITY THAT ACTUALLY HOLDS

**Lock your spatial layout as a left-to-right order, and repeat it verbatim in
every prompt of the scene.**

> FRAME FAR LEFT — window. THEN counter. THEN CHARACTER A seated left of centre
> facing right. THEN the table. THEN CHARACTER B seated right of centre facing
> left. THEN a doorway at FRAME FAR RIGHT.

Then state what's forbidden: *never side by side, never the same side of the
table, never a round table.*

**Screen direction is a promise.** If A faces right and B faces left, the camera
only ever sees A's right side and B's left side. Keep it. Breaking it once, on
purpose, is a dramatic event. Breaking it by accident reads as a mistake.

**To keep someone out of frame, describe the framing that excludes them — never
their absence.** Writing "the chair opposite is empty" gets you a shot built
around a visibly empty chair. Write what's behind the subject, say the camera
never turns that way, and add: *no empty chair in any frame, no unoccupied seat.*

**Freeze small objects, move the camera.** Fine articulated motion — fingers on a
clasp, seating an earpiece — morphs and looks fake. Cut from the hand rising to
the result already done and still.

**Describe objects as fixed properties, never as behaviour.** Write *"a small
steady amber LED, identical in every frame"* — not *"the LED brightens as she
speaks."* Behaviour language makes the model perform it, and it will invent
gestures you didn't ask for.

---

## PART 7 — THE NEGATIVE LIBRARY

Paste the block that matches your project.

**Universal:**
> No text or captions in frame. No watermarks. No visible logos or branding. No
> extra people or background crowd beyond those named. No mirrored or flipped
> compositions. No establishing wide shot at the head of the clip.

**Photoreal:**
> No teal-and-orange grade. No lens flare. No haze or fog. No glamour or
> glossy commercial lighting. No plastic or waxy skin. No CGI-looking surfaces.
> No drone or crane moves. No second unmotivated light source.

**Animated:**
> No cartoonish or chibi proportions. No simplified faces. No flat shading. No
> western comic-book rendering. No 3D CGI-looking characters. No inconsistent
> line weight. No pastel or washed-out coloring. No character redesigns. No
> switching art styles mid-sequence.

**Originality guard — use this if you plan to publish or sell:**
> No branded insignia of any kind. No franchise emblems, crests, or symbols the
> prompt did not specify. No metal forehead plates. No spiral or circular clan
> marks.

Models autocomplete famous franchise marks onto original characters constantly.
Scan every generation for symbols you didn't design, and remove them before you
publish. This one line catches most of them at generation time.

---

## PART 8 — FIVE FAILURES AND THEIR FIXES

**"Every shot looks like a different movie."**
You're rewriting the style lock each time. Paste it identically, every shot.

**"The same character looks like a different person each shot."**
A style lock holds the *world*, not the *people*. Any character in more than one
shot needs a reference image or turnaround sheet carried into every single shot
they appear in. No exceptions.

**"The fight looks like flailing."**
You wrote camera coverage instead of choreography. Name each strike, the specific
counter, and the outcome — in order, as physical action. Attach camera direction
to individual moves, never lead with it.

**"The last line gets cut off."**
Do the timing math in Part 5. Then split the shot.

**"It added a character I never mentioned."**
Two fixes: name your cast explicitly in the prompt, and add *no extra people or
background crowd beyond those named* to your negatives.

---

## PART 9 — THE FILL-IN TEMPLATE

```
[STYLE LOCK — paste verbatim, unchanged from the last shot]

SHOT: <n> seconds. <aspect ratio>. <resolution>.

LOCKED LAYOUT
<left-to-right spatial order, verbatim from the scene>
<what is forbidden about that layout>

IN FRAME
<named characters, with reference images attached>
<named props, as fixed physical properties>

DIALOGUE
EXACTLY <n> SPOKEN LINES EXIST IN THIS CLIP. No extra line, no added words.
<one full second of silence between lines>
<n>: "<exact line>"

TIMELINE
0:00-0:0X — <BEAT NAME>
<what physically happens, then camera>
HARD CUT AT X.Xs
<repeat>

CONTINUITY
<screen direction, axis rules, what must not change>

FINAL FRAME
<exact end composition — stable enough to open the next shot>

[NEGATIVES — paste verbatim]
```

---

## WHERE TO GO NEXT

This kit gives you the structure. The hard part is writing a style lock that
actually holds across thirty shots — that takes a finished film's worth of
trial and error to get right.

The paid packs at **insertfunhere.com** are those locks, already solved:
character plates with the exact descriptions that keep them consistent, scene
layouts that don't drift, and one-click style presets pulled from finished
films.

Start free. Generate something. Then come get the presets.

---

*Free to use, free to share. Please don't resell it.*
*insertfunhere.com*
