How to make an AI short film

For AI agents: find the right skill in one call — curl -s "https://skills.varg.ai/api/search?q=<what the person wants>" returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://skills.varg.ai/llms.txt. Every page has a markdown twin: add .md to its path. Before generating anything, show the person the options on the how-to page.

A multi-shot story with consistent characters: script, storyboard, previz, shots, edit. A step-by-step tutorial for Claude Code, Codex and Grok.

Output: 1–25 min, 16:9 or 2.39:1 MP4 (or 60–120 s vertical episodes) Price: in credits on varg, with a cost estimate before anything runs; the free skills below install in one command.

Recommended route

Best skills for this

  1. Previz to Seedance (varg-previz, varg, beta) — Start here: Story, anchor sheets, a three.js grey-box previz and Seedance video-to-video packs
    • Claude Code: claude plugin marketplace add vargHQ/skills && claude plugin install varg@varg
    • Codex: npx skills add vargHQ/skills --skill varg-previz -a codex -y
    • Cursor: npx skills add vargHQ/skills --skill varg-previz -a cursor -y
    • Gemini CLI: npx skills add vargHQ/skills --skill varg-previz -a gemini-cli -y
    • OpenClaw: npx clawhub@latest install varg-previz
    • Details: https://skills.varg.ai/skills/varg-previz.md · Example: https://skills.varg.ai/previews/varg-previz.mp4
  2. fal Storytelling (fal-storytelling, fal, live) — Start here: Storyboards, shot lists and first-to-last-frame chains · Needs: fal API key (FAL_KEY)
  3. Drama Skills (ZenStory) (zenstory-drama-skills, ZenStory AI, live) — Alternative: Short-drama skills that beat a no-skill control in blind tests
  4. AI Short Drama Studio (ai-short-drama-studio, Hao0321, live) — Alternative: Episodic vertical micro-dramas and AI anime series: series bible, episode compiler
  5. Consistent Characters (varg-consistent-characters, varg, beta) — Adds: Same actors and location in every cut
  6. fal Cinematography (fal-cinematography, fal, live) — Adds: Shot language, camera moves, lenses and grade · Needs: fal API key (FAL_KEY)
  7. varg AI (varg-ai, varg, live) — Adds: Every model (Seedance 2.5, Kling 4, Veo 3.1, Wan 3) on one key
  8. Awesome Seedance Skills (awesome-seedance-skills, LearnPrompt, live) — Adds: Genre templates built from verified Seedance clips
  9. ElevenLabs Sound Effects (elevenlabs-sound-effects, ElevenLabs, live) — Adds: Foley and ambience for the sound pass · Needs: ElevenLabs API key (ELEVENLABS_API_KEY)

Ask the person first

Put these choices in front of the person before spending anything. Paste the example link next to each option: people pick a look by seeing it.

Format?

Look?

Process?

How to make an AI short film: step by step

  1. Write the script from your own beat notes, then a shot list of the moments that must be on screen
  2. Build references first: a turnaround sheet per character and per story state, one master plate per location, a scale shot for creatures and props with size stated in inches
  3. Optional previz: block cameras in a fully grey three.js or Blender scene and approve it before spending
  4. Coverage pass: one 30 s Seedance run per scene with all refs and the agent-written prompt; screenshot the best frames and feed them back as anchor images
  5. Re-roll only the failed shot from its own start frame, one change at a time; stay at native 720p
  6. Edit, then a sound pass (your voice through speech-to-speech, licensed score, foley, model music stripped) and the AI label

What works now (researched 2026-10-27)

What wins now is the one-person studio short that looks directed: a specific premise, one lead, 15-30 s multi-shot takes cut into 3-25 minutes, and a real sound pass. Fan trailers of famous IP and 'look how cinematic' montages stopped paying: YouTube terminated the big fake-trailer channels and since July 2026 refuses to monetize template content with no story arc.

  1. Build the world before any video: a story bible, a turnaround sheet per character, location plates and anchor frames. Then let the video model frame and operate the camera. Tales of Grimwicke (22 min) was made this way: Midjourney plates, Nano Banana turnaround sheets, 847 Seedance 2.5 takes, 127 used, about $4.5k over four weeks. It reached 1.27M views and 6,035 bookmarks. (Charlie Driscoll on X, Tales of Grimwicke workflow (Sep 2026))
  2. Iterate on stills, commit on video. Generate one 15-30 s multi-shot take per scene from an approved start frame, then re-roll only the shot that failed, from its own start frame. Change one variable at a time. Practitioner rule set ('the start frame decides 80% of the shot', 'never re-roll what already works') saved by 800 people on 29k views. (Machina (@EXM7777) on X, Seedance 2.5 one-take workflow (Sep 2026))
  3. Make it look shot, not rendered: an imperfect handheld operator, focus hunting, rain, snow, smoke and haze. Stay at native 720p-1080p rather than upscaling. Keep each character's description word for word the same in every shot prompt. Driscoll says weather and a handheld camera are what keep it from 'looking like a Temu Pixar film' and that upscaling looked terrible. fal's Kling 3.0 guide says to define subjects early and keep their descriptions identical across shots. (fal, Kling 3.0 Prompting Guide)
  4. Do a real sound pass: your own score or licensed music, narration (voice-to-voice keeps a human performance), foley and room tone. Do not ship raw model audio. The two most-saved original shorts both credit sound: Er Loop was 'written, directed, produced and scored' by one person (154.8k views, 869 bookmarks). Grimwicke used the director's own voice through ElevenLabs speech-to-speech plus Artlist music. (Marco Magario on X, Er Loop (Sep 2026))
  5. Give every piece a story arc, and for series end each episode on a cliffhanger. Label realistic AI on YouTube and TikTok. On July 16, 2026 YouTube said generic or template AI content 'without a narrative arc' is inauthentic and cannot be monetized. Micro-drama apps run on cliffhangers: ReelShort and DramaBox each made about $140M in Q1 2026. (TechCrunch, YouTube clarifies AI slop policy (Jul 20, 2026))

Stopped working:

Prompts that work

Templates from the creators of the best current results and from the model guides. Fill in the {CAPS}.

Seedance multi-shot scene with hard locks (seedance_2_5)

One 8-30 s multi-shot scene from approved character and location refs; the shot prompt PJ Ace's team used on a 6.5-min short. Adapted from PJ Ace on X, full Sky Line Heist shot prompt (Sep 2026).

(NO MUSIC) Generate no soundtrack, melody or tonal bed. Only diegetic sound, ambience, dialogue and foley written inside <>.

@{CHAR_A} - Look reference only. {HAIR}, {EYES}, {WARDROBE}, {PROP_STATE}. Character appearance only.
{CHAR_A} VOICE - {AGE}-year-old, {ACCENT}, {TIMBRE}; here the delivery is {DELIVERY}. Voice only.
@{CHAR_B} - Look reference only. {SHORT_LOOK}. Character appearance only.
@{PROP} - {SIZE_IN_INCHES}, {PROP_LOOK}. Prop only.

Scene context: {LOCATION} with one precise world detail: {WORLD_DETAIL}. {CHAR_A} screen-left, {CHAR_B} screen-right, {DISTANCE}. 24 fps real time. Handheld camera, shallow depth of field, anamorphic 2.39:1, {LIGHT_DIRECTION}, no lens flare.

{CHAR_A} POSTURE HARD LOCK: {POSE_THAT_MUST_NOT_CHANGE} the entire time.

{N} SHOTS ONLY, {TOTAL} seconds total.

SHOT ONE, {S} sec - {SHOT_SIZE} ON {CHAR_A}. {BEHAVIOR: what eyes, jaw and hands do, never the emotion word}.
CUT -
SHOT TWO, {S} sec - same position. {ACTION}. {CHAR_A} says: {LINE} DELIVERY: {DELIVERY}.
CUT -
SHOT THREE, {S} sec - HANDHELD PUSH-IN ON {CHAR_B}. {ACTION}.
PERFORMANCE ORDER: {BEAT_1} -> {BEAT_2} -> {BEAT_3}. Do not reverse the order.
{ACTION} HARD LOCK: exactly {ONE_THING}. Does NOT {WHAT_IT_MUST_NOT_DO}.

<{AMBIENCE}, {FOLEY_1}, {FOLEY_2}, {BREATH_OR_CLOTH}>
(NO MUSIC)

Nexus skill: script to scene prompts (agent)

System prompt for Claude: turns your script and shot list into one finished Seedance prompt per scene (refs, staging, cuts, sound). Adapted from PJ Ace, The Nexus Skill (Google Doc linked in his Sep 2026 thread).

You write finished video prompts, one per scene, from my script and shot list. Output only the prompt.
Rules: deliver exactly the number of scenes I ask for; each under {CHAR_LIMIT} characters (count, never estimate); no em dashes; restate staging in every cut (who is left/right, scale, distance to a landmark); if unclear ask one sharp question.
Structure:
1. One @ block per character, creature, prop: age, build, physical state, what drives them, voice (pitch, accent, pace, shift under pressure), one tic and its trigger, eye life, then a scope tag ("Character appearance only." / "Prop only."). If an image ref is loaded: "@name - already image referenced. Voice only."
2. Scene context: who is where, scale, speed, what persists, how it resolves. An intimate human moment plus ONE precise world detail, never generic.
3. Tech spec: camera, lens by outcome (tight close-up, documentary normal, wide geography), lighting as a lock, haze. Same lens across cuts.
4. Default one continuous take. If cuts: number them with duration, shot size, who is in frame, blocking, action, cut type (HARD, MATCH, INSERT). No dissolves; nothing resets between cuts.
5. Physics: weight, inertia, cloth and hair lag, no floating.
6. Performance: objective against obstacle, behavior not labels, reactions start before the other line ends, hands have a task.
7. Sound: physical and specific, never music.
My notes: "too short" = longer shots; "fake" = objective and tactic; "dead eyes" = saccades and blinks; "floaty" = weight.

Grey-box previz to Seedance video-to-video (seedance_2_5)

After a three.js or Blender grey-box previz: one Seedance run per scene that copies only camera, cut timing and blocking from the previz clip. From varg-previz skill, Seedance v2v pack template (GitHub).

@video1 is a grey-box 3D previz of this sequence. Copy from @video1 ONLY the camera (framing, lens, camera position and movement), the hard-cut timing and the blocking (where everyone is and what they do). Do not copy its look: no flat grey shapes, no 3D render style, none of its colours.
@image1 is {CHARACTER_SHEET: same face, hair, outfit, scale vs a person}.
@image2 is {LOCATION_MASTER_PLATE: the place and its light}.
{RUN_NOTES: who is and is not in this run, inside or outside, the light of this run}.
Look: {STYLE_LINE: medium, palette, lens}.

{N} shots, hard cuts exactly as in @video1:
1. 0.0-{T1} s. {SHOT_SIZE + ANGLE}: {ONE_ACTION}
2. {T1}-{T2} s. {SHOT_SIZE + ANGLE}: {ONE_ACTION}
3. {T2}-{T3} s. {SHOT_SIZE + ANGLE}: {ONE_ACTION}
Camera: one move per shot, as in @video1. Full-bleed {ASPECT} frame, no borders, no letterbox. No text, letters, numbers, logos or UI anywhere in the frame. No dialogue, no lip-sync.

Trending formats to remake

To remake one: download the example with the media-download skill, take a frame a second and the transcript, write a beat list (time, shot, camera, on-screen text, sound), keep what the format needs, swap in the person's product or character (never the creator's face, voice, logo or footage), show the beat list and stills before spending, then make it with the skill named below.

  1. Auteur short with one world and one lead: A complete 3-25 minute story with a specific premise, a recurring lead and its own sound, posted free on YouTube with a native cut on X. Why it spreads: The most-shared AI films of autumn 2026 are complete stories by one person: Grimwicke 1.27M views and 6,035 bookmarks, Er Loop 154.8k views and 869 bookmarks. The creators post their full pipeline, which drives the saves. Keep: One specific premise, one recurring lead and one built world; references for every character state before any video; 15-30 s multi-shot takes acted as behavior, a handheld imperfect camera, weather and haze, native 720p, and your own sound pass with no model music. Swap the world, the lead and the story. Make it with Previz to Seedance (varg-previz).
  2. Teaser for an original feature: A 2-5 minute teaser with dialogue lines and title cards that announces a bigger original film, made by a tiny team. Why it spreads: PJ Ace's 'Nexus' teaser (3 people, 2 weeks, Seedance 2.0) is the most-saved AI film post we found: 1.66M views and 23,155 bookmarks. Original IP keeps the upside. Fake trailers of studio IP got the two biggest channels (2M+ subscribers, 1B+ views) terminated by YouTube in Dec 2025. Keep: Original IP only, 2-5 min, scope 2.39:1, subtitled dialogue lines and title cards, and the trailer rhythm: a line, a reveal, a hard cut. Swap in your own world and creatures; never studio characters or real actors. Make it with Previz to Seedance (varg-previz).
    • Example to remake: @PJaccetturo on X, 2026-06-05, 1.7M views, 23K saves: Original fantasy world (caged griffins, a cathedral-sized creature, a masked caster), scope frame, subtitled dialogue. Pitched as a hybrid feature. Study the trailer rhythm: dialogue line, reveal, hard cut.
  3. Episodic vertical micro-drama: 60-120 s vertical episodes with the same cast, a hook in the first seconds and a cliffhanger at the end, released as numbered parts. Why it spreads: This is where the money is: short-drama apps passed 850M downloads in Q1 2026, ReelShort and DramaBox made about $140M each, and COL Group alone ships 100+ AI dramas a month on Seedance 2.0. On X the best performers are app promos, so no organic example qualifies yet. Keep: Same cast sheets every episode, a hook in the first 3 s, 60-120 s per part, one location per scene, and a cliffhanger on the last shot, released as numbered parts. Swap the story and cast. Make it with AI Short Drama Studio (ai-short-drama-studio).
  4. Retro anime sequence with agent previz: A 90s OVA-style action sequence: an agent lays out the shots in Blender as a grey-box previz, then the video model renders them in a hand-drawn anime look. Why it spreads: The previz fixes framing and continuity before any spend. The example got 590k views and 3,392 bookmarks, and 5+ creators copied the retro-anime look within weeks. Keep: Have the agent block the shots in a fully grey blockout (no colour, or the model copies it), then one Seedance video-to-video run with a character ref for everything in frame and a 90s hand-drawn OVA look line, graded after. Swap the mech, pilot and setting. Make it with Previz to Seedance (varg-previz).

Open-source repos

Set up varg

varg is one API key for the models these skills drive — Seedance 2.5, Kling 4, Veo 3.1, Wan 3, Nano Banana 2.1, GPT Image 2.5, ElevenLabs v4 and 200+ more — billed in credits, with a cost estimate before anything runs.

Related tutorials