YouMind
Увійти

Solo AI Short Drama Production with Claude and H3: Full Workflow Guide

@TuyetloanP81840
СПРОЩЕНА КИТАЙСЬКА08 лист. 2026 р.
2.8M
14.2K
2.1K
368
22.5K

Коротко

A comprehensive tutorial on creating AI short dramas solo using Claude for scripting and MiniMax H3 for video generation. It details a low-cost workflow, providing specific prompts for script, assets, and video generation, along with editing techniques in CapCut.

For a 5-minute episode of an AI short drama where characters speak, my calculation for the video generation step comes to just over 27 RMB.

The same volume using Seedance 2.5 at Volcano Engine's official price would cost around 1,360 RMB.

Here are the results first 👇

  • One person, no GPU purchase needed, creates short dramas with speaking characters
  • Video segments cost about 0.2 RMB each; a 5-minute episode costs less than 30 RMB. For higher quality, refined segments cost between 0.3 and 1.5 RMB
  • From script to editing, every step has copy-paste prompts (8 total)
  • My tuned H3 workflow is released directly, ready to use
  • Works for anime-style comics and realistic live-action styles; choose your style in the prompts

The division of labor in this article is as follows:

  • Scriptwriting and storyboarding: Claude (DeepSeek works too if you don't have it)
  • Art direction, character/scene/prop images: GPT Image in ChatGPT, Jimeng, or Nano Banana
  • Cinematography and acting: MiniMax H3, run on RunningHub
  • Editing: CapCut
  • You: Producer, defining the story, selecting shots, making final decisions

One person is one film crew. The order is simple: Cast first, then shoot, finally edit.

Miller - inline image

View image description

ALT

How to read this article:

  • First time doing this: Look at the "Beginner Route" below to make a 10-second clip in one hour
  • Want to make a full episode: Follow Sections 1, 2, and 3
  • Want better visuals: Check the "Advanced" section later
  • Want tools directly: There is a lighting quick-reference page and a short drama skill pack for Claude at the end

First Time? Start Here: Make a 10-Second Clip in One Hour

The next three sections cover a full episode. But for your first time, don't start with 5 minutes.

Make a ~10-second clip first: one character, one scene, one line of dialogue. Once you complete this, you'll understand what each step looks like before tackling a full episode.

1. Write one line of dialogue (5 minutes)

Use the script prompt from Section 1.1, but only for one scene, one character, and one or two lines. Example: Yang Yang stands in the center of the academy courtyard and says calmly, "Actually, I am the Poet Immortal."

2. Generate two images (15 minutes)

Use the prompts from Section 1.3 to generate only two images: one character sheet (full body, front view, white background, remember to specify shoes) and one scene image (no people in the frame). Skip identity boards, props, and storyboards for now.

3. Write the video prompt (5 minutes)

Give the dialogue and the purpose of the two images to Claude. Use the prompt from Section 2.2 to write just one segment.

4. Run on RunningHub (20 minutes)

Follow the 6 steps in Section 2.3. Run a 5-second preview first; if satisfied, generate 10 seconds. One segment costs about 100 RH coins. You can try with free daily check-in coins; even 3 attempts cost less than 1 RMB.

5. Add subtitles in CapCut and export (10 minutes)

Import the clip, recognize subtitles, verify against the script, and export in 9:16 aspect ratio.

Miller - inline image

View image description

ALT

Once you finish this segment, you've completed the entire workflow.

Afterward, if you want to make a full episode, follow the three sections: storyboards, identity boards, reaction shots, and editing lists are things you need once you're making longer pieces.

Section 1: Pre-production Assets, Define the Script and "Actors" First

Many people jump straight into video generation, resulting in characters looking different in every segment and plots that don't connect.

The problem isn't the video model; it's pre-production. This stage involves three things: script, storyboard, and asset images.

This step only costs money for text and images. Changing a word is free. Once you reach the video stage, every change costs money for another attempt. Solve everything possible here.

1.1 Script: Only Write What the Camera Can See

AI video models don't understand "he felt sad inside," they only understand "he lowered his head and gripped his sleeve tightly."

So scripts for AI are different from novels for humans. Remember three rules:

1. Convert psychological activity into actions, expressions, or inner monologues.

Don't write "the judge was dissatisfied." Write "the judge narrowed his eyes and slowly stroked his beard."

2. Fixed format, clearly labeled sound types.

Scenes, actions, dialogue, and sound effects have specific formats. Spoken aloud, thought internally, or voice-over (character off-screen) must be distinguished in the script so AI knows which parts are visual and which are audio.

3. A 5-minute episode revolves around one main conflict.

Introduce conflict in the first 3 seconds, end on suspense. At roughly 2 seconds per 10 characters of dialogue, 750–1,000 characters of dialogue are enough for one episode.

Yang Yang's scene looks like this (excerpt):

text
1Characters in this episode: Yang Yang, Judge Teacher, Scholar Li, Student Zhao, Crowd
2
31-1 Ancient Town Academy Courtyard [Ext] [Day]
4△ Yang Yang wears plain cloth clothes, standing in the center of the yard, expression calm.
5【SFX】Whispers and low laughter from the crowd.
6Yang Yang: (calmly) Actually, I am the Poet Immortal.
7△ The crowd freezes, then bursts into laughter.
8Scholar Li: (looking up and down, face full of mockery) The Poet Immortal? Truly ignorant and fearless.
9△ The Judge Teacher narrows his eyes, slowly stroking his beard.
10Judge Teacher (inner monologue): No matter what poem this person recites, as long as I am present, he cannot possibly win.

Copy directly 👇 (Send to Claude or DeepSeek)

text
1You are a short drama screenwriter familiar with the pacing of Douyin and WeChat Channels. Adapt the following story into a short drama script for a ~5-minute episode.
2
3Requirements:
41. List characters appearing in this episode at the beginning.
52. Start each scene with a header: Scene Number Scene Name [Int/Ext] [Day/Night], e.g., 1-1 Ancient Town Academy Courtyard [Ext] [Day].
63. Start visual/action lines with "△", writing only what the camera can see. Do not write "he thought inside." Convert psychological activity into actions, expressions, or inner monologues.
74. Dialogue format: Character Name: (Emotion/Action) Dialogue. Inner thoughts written as "Character Name (Inner Monologue):". Voice-overs (character off-screen) written as "Character Name (Voice-over):".
85. Important sounds written as 【SFX】...
96. Total dialogue 750–1,000 characters. Conflict appears within the first 3 seconds, ending on suspense without premature resolution.
107. Each sentence on a new line. Output only the script, no explanations or analysis.
11
12Story:
13(Paste novel excerpt or your story synopsis)

Only have an idea, no story yet? Have it write 3 versions of a 300-character synopsis, pick one, then use the above to turn it into a script.

1.2 Storyboard: Advanced, Beginners Can Skip

Scripts describe "what happens," storyboards describe "what the audience sees": shot size, duration, who is in frame.

Beginners can skip this and go directly to Section 2 to write prompts from the script; you can still produce video. Use storyboards when there is heavy dialogue or you want to control pacing.

Remember four rules for cutting:

1. Cut when speakers change.

Don't shoot Q&A in one take. Cut between speaker and listener. Yang Yang's line can be split like this:

  1. Yang Yang speaks "Actually, I am the Poet Immortal": Close-up, 2s
  2. Crowd freezes, then laughs: Wide shot, 2s
  3. Scholar Li "The Poet Immortal? Truly ignorant...": Close-up, 2s
  4. Yang Yang remains unmoved: Close-up, 1s

2. Cut when emotion changes. Calm to shock, shock to anger—each change gets a shot.

3. Break big actions into three steps. Preparation, execution, result. Cramming a whole action sequence into one shot confuses AI.

4. Must cut for scene/time jumps. Leave 1–2 seconds for transition.

Shot durations are specified in the prompts below. Remember two numbers: dialogue ~2s per 10 characters; listener reactions ~1s.

After breaking down the storyboard, combine consecutive 2–4 shots into a "material segment" (5–15s). Section 2 generates by "segment."

Copy directly 👇 (Send previous script to Claude)

text
1You are a storyboard director with 20 years of experience. Break the following script into a storyboard table.
2
3Cutting rules:
41. Cut when speakers change: Speaker → Listener reaction → Speaker. Long dialogues must cut to listeners mid-sentence.
52. Give an independent shot for each emotional change.
63. Break big actions into Preparation → Execution → Result; complex actions into 3–5 shots.
74. Must cut for scene/time jumps; add 1–2 second transition shots between scenes.
85. Do not use the same shot size for 3 consecutive shots.
9
10Duration: Extreme/Wide 3–4s, Medium 2–3s, Close-up/Extreme Close-up 2–3s; Dialogue ~2s per 10 characters; Reaction shots ~1s; Climax 3–5s; Single shot max 10s.
11
12Output a table with columns: Shot # | Duration | Shot Size | Camera Position/Movement (Static/Push/Pull/Pan/Follow) | Visual Content | Dialogue/SFX | Characters in Frame
13Write "Visual Content" as a paragraph including: style, environment, character positions, actions, expressions, lighting, composition.
14
15After the table, combine consecutive 2–4 shots into material segments, 5–15s each, labeling correspondence like "Material Segment 1: Shots 1–3".
16
17Style: (3D Chinese Anime / 2D Anime / Realistic Live Action, choose one)
18Script:
19(Paste here)

1.3 Asset Images: "Cast" Before Shooting

Asset images are your actors, sets, and props. H3 accepts up to 9 reference images at once. The cleaner the input images, the more consistent the characters look across segments.

Character Images, Five Rules:

  1. Full body, front view, white background, hands down, holding nothing, neutral expression
  2. Always specify shoes; otherwise AI often generates half-body shots
  3. Clearly state height, build, age, face shape, hairstyle, clothing, matching the era
  4. Adult characters ~8 heads tall; villains shouldn't be excessively ugly
  5. Extras don't need individual designs

After defining the look, create a "Character Identity Board": front, side, back full body, plus several expressions and poses. Video models use this to recognize the person from any angle.

Use AI-designed characters, not real photos, especially celebrity photos: platforms block them, and there are portrait rights risks.

Scene Images, Three Rules:

  1. No people in the frame
  2. Same art style as characters
  3. For dialogue scenes, generate three: Reverse Shot A (from A's perspective), Reverse Shot B (from B's perspective), and a Wide Shot showing all positions. Modify the Reverse Shot A prompt to get B; prompts below.

Prop Images, Two Rules:

  1. Pure white background, no context, no people
  2. Specify actual dimensions, e.g., "Teacup diameter ~8cm, fits in one hand," otherwise sizes fluctuate in video

Copy directly 👇 (Step 1: Send script to Claude to list assets and image prompts)

text
1You are an AI image designer. Based on the script below, list all required characters, scenes, and props, and write image prompts for each. Prompts only, no images.
2
3Characters: Full body front view, white background, hands naturally down, holding nothing, no expression. Specify height, build, age, face shape, eyes, hairstyle, hair color, clothing, and shoes, matching the story era. Adults ~8 heads tall; villains not excessively ugly. Skip extras.
4Scenes: No people in frame. For dialogue scenes, write one reverse shot for each character's position, plus one wide shot showing all positions.
5Props: Pure white background, no context/people, specify actual dimensions (e.g., "diameter ~8cm, fits in one hand").
6
7Style: (3D Chinese Anime UE5 render / 2D Anime / Realistic Live Action, choose one). Keep all characters, scenes, props in the same style.
8
9Script:
10(Paste here)

Copy directly 👇 (Step 2: After character image is set, send image + this prompt to image model for Identity Board)

text
1Reference this character image to create a 16:9 Character Identity Board. White or off-white background, no environment, props, or watermarks.
2
3- Place a large front-view full-body image slightly left of center; leave spacing around it for back, side full-body views, sitting pose, crouching pose, and 3 facial expression close-ups.
4- All images non-overlapping; faces, hands, feet not cropped.
5- Maintain identical face, hairstyle, outfit, and body proportions across all angles.
6- Add a simple character card in the corner: Name, Role, Personality Keywords.
7- Style clean and minimal, like an animation studio character sheet, not a poster.

Copy directly 👇 (Step 3: After generating Reverse Shot A scene, generate Reverse Shot B)

text
1Keep the art style, lighting, and decor of this scene unchanged. Generate the view from the opposite direction: swap left/right content, swap foreground/background. Replace the original distant door/outdoor background with a wall matching the building's style. No people in the frame.
Miller - inline image

View image description

ALT

At this point, pre-production is complete: one script, one storyboard (optional), one identity board per main character, 2–3 images per scene, one per key prop.

Before shooting, spend two minutes checking these four questions:

  • Does every main character have an identity board with consistent faces across angles?
  • Are there any people mixed into scene images?
  • Did you specify dimensions for props?
  • Is total dialogue under 1,000 characters?

Next, feed them to H3 to make characters move and speak.

Section 2: Video Generation, Making Characters Move and Speak

With asset images ready, the next step is turning images into moving, speaking clips.

I use MiniMax H3, running on RunningHub. First, why not use the currently popular Seedance 2.5?

2.1 Why Choose H3: Cheap and Generates Audio Simultaneously

Brief intro to both:

  • H3: MiniMax video model released July 31 this year. Max 15s per segment, generates visuals and audio (dialogue, SFX) simultaneously. Open-sourced August 3.
  • RunningHub: A website for running ComfyUI workflows online. Use others' built workflows by uploading images and filling prompts; no need to buy GPUs or install software.

Honestly, Seedance 2.5 is stronger in visual quality and complex motion.

But for short dramas, a 5-minute episode needs dozens of segments, requiring multiple retries. Small unit price differences multiply into huge costs. My calculation:

  • RunningHub Pro is currently 99 RMB/month (originally 159). The "Coin Heavy" package gives 54,000 RH coins. 1 RH coin ≈ 0.0018 RMB
  • Generating a 5–10s video segment costs ~100 RH coins, i.e., ~0.2 RMB per segment
  • For a 5-minute episode (~50 segments avg 6s, 3 retries each = 150 generations, 15,000 RH coins), video step costs ~27 RMB
  • Same volume (5 min × 3 retries = 900s) with Seedance 2.5 720p at Volcano Engine official price (no ref video, 5s @ 7.56 RMB, ~1.51 RMB/s) costs ~1,360 RMB. On RunningHub (Aug launch promo 720P ~0.49 RMB/s), it costs ~440 RMB
  • Monthly 99 RMB membership coins allow ~3.5 episodes (~18 mins finished video)
Miller - inline image

View image description

ALT

Of course, 0.2 RMB per segment is the "usable" tier, outputting 640P. For higher quality, refine the first generation. For the same 10s shot, I tested three tiers:

  • Budget Tier: First pass only, 640P, ~100 RH coins, ~0.2 RMB/segment. For practice/story testing
  • Refined Tier: First pass + super-resolution upscale (blended 50% with original), 1920 width, ~185 RH coins, ~0.3+ RMB/segment. For publishing
  • Ultimate Tier: Second pass, full re-refinement, 1920 width, ~795 RH coins, ~1.5 RMB/segment. For delivery/key shots

[Confirm: Are 185 and 795 total prices including first pass?]

In my view, refined results rival Seedance 2.0. Even if all 50 selected shots per episode use Ultimate Tier, adding ~1+ RMB per segment means ~60-70 RMB extra per episode, still less than 1/10th of Seedance's official price.

H3 official API charges per second: 768P ~0.5 RMB/s, 2K ~0.8 RMB/s. A 10s 768P segment costs ~5 RMB. Why is mine 0.2 RMB? Three reasons: H3 is open-source; my workflow uses RunningHub's INT8 quantized version with acceleration nodes for speed; RunningHub charges by machine runtime, faster = cheaper; I usually output 640P, smaller than 768P.

Miller - inline image

Two other things I value about H3:

  1. Character Consistency: Accepts up to 9 reference images. Identity boards, scene images, prop images can all be included.
  2. One Workflow, Three Uses: Reference-to-video, Image-to-video/Start-End frames, Digital Human. Covers daily production, shot patching, lip-syncing (explained in 2.3, 2.4).

2.2 Video Prompts: One Segment, One Thing

The biggest difference between short dramas and generic AI shorts is: characters speak, and plots continue scene by scene.

Prompts must be much more detailed than "a person walking." Course rules simplified to six:

1. One segment, one thing. 5–15s, handling only one action, emotional shift, or key line.

2. Write positions first, then actions. Who is left/right, where is the door, who holds the prop. Without this, characters teleport in the next segment.

3. Actions continue from previous segment's end. If they stood up last time, start this time standing.

4. Dialogue verbatim, three sound types:

  • Spoken aloud: Yang Yang says: "Actually, I am the Poet Immortal."
  • Inner thought: Judge Teacher inner monologue: "No matter what poem..."
  • Voice-over (off-screen): Yang Yang voice-over: "..."

Without distinction, AI makes characters mouth their inner thoughts.

5. Don't keep camera on speaker constantly. Long dialogue freezes the visual. Cut to listener reactions mid-line, keeping audio continuous (change to voice-over for those seconds). If hard to do in one segment, fix in editing (3.2).

6. Label each reference image's purpose. In my workflow, type @ in the prompt box to select materials: "@Image1 is Yang Yang's face, @Image2 is the academy courtyard." Better than dumping them all.

Miller - inline image

View image description

ALT

Copy directly 👇 (Send to Claude to rewrite script/storyboard into H3 prompts)

text
1You are a short drama director and AI video prompt engineer. Rewrite the following script or storyboard into video prompts usable directly by MiniMax H3.
2
3Rules:
41. Each segment 5–15s, handling one action, emotional shift, or key line. Different scenes/times must be separate segments.
52. Start each segment with 【Positions】: Who is left/right, door location, camera side, key prop holder/direction.
63. Actions continue from previous segment's end; don't repeat completed actions.
74. Dialogue verbatim. Three sound types: X says: "" / X inner monologue: "" / X voice-over: "". Specify tone for each line.
85. Reference materials by purpose: @Image1 = X's face, @Image2 = Scene, @Image3 = Prop... If audio ref: @Audio1 = X's voice.
96. Specify concrete lighting and camera: Shot size, position, static/push/pull. Camera static during dialogue.
107. Sound only includes dialogue and visible-cause SFX (footsteps, door closing). No background music, no subtitles.
118. Write a fixed voice description for each speaking character first (age, timbre, speed), then copy exactly in every segment. Max 1-2 lines of dialogue for 5s segments.
12
13Output format per segment:
14【Material Segment X】(X seconds)
15【References】@Image1: ... @Image2: ... (If audio: @Audio1 = X's voice)
16【Positions】...
17【Visuals】0-Xs: Shot size, camera, events, expressions/actions.
18【Sound】X says: "..." (tone); SFX: ...
19【Prohibitions】No unscripted characters, no modern items (for period dramas), props don't switch hands/disappear, no teleportation, no dialogue changes, no subtitles/music.
20
21Script or Storyboard:
22(Paste here)

2.3 Running H3 on RunningHub: From Registration to Download

I released my tuned H3 workflow. First time, follow these 6 steps:

Step 1: Register. Go to runninghub.cn, register/login with phone number. Daily check-ins give free RH coins; test with those first.

Step 2: Don't buy membership immediately. Use free check-in coins to complete the "Beginner Route." Buy the 99 RMB membership if you like it. Membership auto-renews monthly; cancel if unwanted. Member coins expire; don't hoard.

Step 3: Open my H3 workflow. [Workflow Public Link]. Left side: 9 image boxes, 3 audio boxes, one "Disable Box". Middle: Prompt box and parameters. Right side: Run area (don't touch). Uses open-source node ComfyUI-MiniMaxH3-Easy; thanks to author.

Miller - inline image

View image description

ALT

Step 4: Upload reference images and audio. Image box numbers match @Image1, @Image2 in prompts:

  • Image 1: Main character's identity board (from Section 1; Beginner Route: just the character image)
  • Image 2: Scene image for this segment (only one)
  • Image 3+: Other characters, key props, or wide positioning shot
  • Audio 1: Character voice reference (see 2.4); leave empty if none
  • Enable corresponding boxes in "Disable Box"; default is only Image 1 enabled

Step 5: Paste prompt, adjust parameters.

  • Mode: "Reference-to-Video" for daily production; "Image-to-Video/Start-End Frames" for patching (see 2.5); "Digital Human" for lip-syncing pre-recorded audio (see 2.4)
  • Prompt: Paste, then use @ to re-select materials to ensure mapping
  • Resolution: 640P (Budget Tier; use Refined for clarity)
  • Aspect Ratio: Change to 9:16 vertical (default is 16:9)
  • Duration: Set to 5s for first run

Common pitfall: Workflow seed is fixed. Fixed means same images/prompts yield nearly identical results. To "retry" for variation, change the seed number. To tweak prompt only, keep seed constant.

Step 6: Click Run, Download. Top right has three run buttons for different machine speeds/costs [Add: Which button do you usually click, cost per segment?]. Check lip-sync and face first; download if satisfied. Name files "Segment#-Attempt#" (e.g., 03-2) for easy editing.

Two money-saving habits:

  • Run 5s preview first, then 15s if good.
  • Test characters individually first. Run each character separately for 5s to confirm consistency, then combine for long shots. Combining multiple characters for 15s immediately wastes money on failures.

2.4 Character Voices: Same Person Speaking Across Segments

Short dramas rely on speaking characters. If Yang Yang has a clear young male voice in one segment and a deep uncle voice in the next, viewers break immersion.

Three methods, simple to advanced:

1. Hardcode voice descriptions, copy same sentence every segment.

Write a fixed voice description for each main character, e.g., "Yang Yang: Early 20s male voice, clear, steady pace, slightly indifferent." Copy exactly in every prompt; don't vary wording.

2. Extract audio from first good segment as reference.

H3 accepts up to 3 audio references. Docs state this ensures voice consistency across segments. After generating a satisfactory first segment, extract the character's speech (clean, no background noise). Upload this as reference in subsequent segments, prompting "@Audio1 is Yang Yang's voice."

3. Dub first, then use "Digital Human" mode for lip-sync.

Use dubbing tools to record the line (same voice each time). Switch mode to "Digital Human," upload the audio track. H3 generates visuals/lip-sync based on this audio, ensuring voice stability. This mode accepts only one driving audio track.

Two reminders:

  • Don't cram too much dialogue per segment. Docs say 1-2 sentences fit 5s clips. Too long, unclear tone, or short clips cause lip-sync errors.
  • Use only your own or authorized voices. Don't clone celebrities or others' voices.

2.5 Glitches? Don't Re-roll Immediately

AI videos always have glitches: extra fingers, resizing props, face swaps.

Handling order:

  1. Edit out if possible. Issues in first/last few frames? Just trim them. This is "Edit if possible" from 2.5.
  2. Patch if possible. Issue in middle segment? Capture the frame before the error, fix it with an image model (e.g., resize cup), use as start frame for H3 to regenerate just that small part. My workflow includes H3 Start-End Frame model: Switch mode to "Image-to-Video/Start-End Frames," enable one image box with the fixed image to generate from that frame; two images define start/end.
  3. Re-roll last. Before re-rolling, modify prompt or references. Re-rolling identically likely yields same error.
Miller - inline image

View image description

ALT

Three prevention tips:

  • Feed only one scene image per video segment. Multiple scene images confuse AI, causing inconsistent backgrounds.
  • For dialogue scenes, include the "Wide Positioning Shot" from Section 1. It clarifies positions better than text descriptions.
  • Specify prop dimensions in prompts, matching Section 1 prop images.

Section 3: Post-production, Editing Clips into an Episode with CapCut

With materials generated, the final step is editing dozens of clips into one episode.

I use CapCut, much simpler than DaVinci Resolve. Free version covers most needs. Desktop Pro version is more convenient.

Post-production involves four tasks: Ordering, Dialogue Fixing, Audio Mixing, Subtitles.

3.1 Rough Cut: Order First, Trim Ends

  1. Drag all materials to timeline in script order. If storyboarded, follow material segment numbers.
  2. Review each segment start-to-end. Trim problematic head/tail frames. This is "Edit if possible" from 2.5.
  3. Insert patched segments (from 2.5) back into their original positions.

Pacing for short dramas is faster than films:

  • Conflict in first 3 seconds. First frame should be the hook. Yang Yang's episode starts with "Actually, I am the Poet Immortal," followed by crowd laughter.
  • Don't linger on screens with no new info. Tighten gaps between dialogue.
  • End on suspense. Student Zhao shouts "Maybe he stole the poem!", crowd gasps, cut to black. Don't drag further.

3.2 Dialogue Scenes: Continuous Audio, Visuals Cut to Listeners

Section 1 mentioned cutting between speaker/listener. What if generated segment keeps camera on speaker throughout?

Fix in editing:

  1. Separate audio from speaking segment (Right-click "Separate Audio" in CapCut). Keep audio in place, uninterrupted.
  2. On upper track, overlay a "listener" reaction shot (e.g., Scholar Li widening eyes, crowd exchanging glances) for 1–2 seconds.
  3. Viewer sees: First half on Yang Yang, second half on reaction, while Yang Yang's voice continues.

Where do reaction shots come from? During generation, run extra 3–5s silent reaction shots for each main character. Per Section 2 math, ~0.2 RMB/segment. 10 extra segments cost ~2 RMB, saving significant editing effort.

3.3 Audio: Dialogue, SFX, Music in Three Layers

H3 generates dialogue and SFX together. Add music uniformly in CapCut. Letting AI add music per segment causes tonal inconsistencies, hence "No background music" in Section 2 prompts.

Volume hierarchy:

  • Dialogue loudest, must be clear
  • SFX next, footsteps/doors follow visuals
  • Music lowest, duck when someone speaks (add keyframes to lower volume during speech)

Common issues:

  • Inconsistent dialogue volume. Select all clips, use CapCut's "Loudness Normalize" to balance.
  • Same character, varying voice in some segments. Prevention in 2.4. If missed, easiest fix is re-generating just those lines or re-dubbing.

Use CapCut library or royalty-free music; check licenses for commercial use.

3.4 Subtitles: Auto-Recognize, Then Verify Against Script

AI-generated text is often garbled/wrong. Hence Section 2 instructed H3 to avoid subtitles. Add them in CapCut.

  1. Use CapCut's "Recognize Subtitles" to auto-transcribe dialogue.
  2. Use "Script Match," pasting script dialogue. It aligns timing automatically, reducing errors.
  3. Verify line-by-line against script. AI sometimes mispronounces words; subtitles must match audio: either re-generate segment or correct subtitle to actual pronunciation.

Styling: Bottom third of screen, don't cover faces. Keep lines short. Distinguish inner monologues with color/brackets.

3.5 Color Grading, Quality, Export

Color Grading: Dozens of segments may have slight color variations. Select all clips, apply same filter or adjust brightness/color temp uniformly for cohesive look.

Quality: My workflow outputs 640P first pass. For clarity, use Refined/Ultimate Tiers (1920 width). CapCut also has enhancement features (usually paid).

Export: Vertical 9:16, 1080P, 30fps. Suitable for Douyin, WeChat Channels, X.

Before Publishing, Two Final Checks:

  1. Watch full screen on mobile. Check glitch list from 2.5: fingers, props, face swaps, door directions.
  2. Mark as "AI Generated." China's "Measures for Labeling AI-Generated Synthetic Content" (effective Sept 1, 2025) Article 10 requires active declaration and platform labeling. Check the AI generation declaration box on Douyin/WeChat Channels.

Copy directly 👇 (After materials are ready, send script and material list to Claude for an editing list)

text
1You are a short drama editor. Below is this episode's script and my generated material list (segment #, duration, content, issues). Create an editing list.
2
3Requirements:
41. Order materials by plot, specifying keep-trim ranges for each segment.
52. For long dialogue stuck on speaker, mark where to overlay which reaction shot, duration, keeping audio continuous.
63. List missing shots needing generation: what visual, length, whose reaction.
74. Music segmentation: Emotion (tense/light/reversal) and start/end times per segment. Mark "Duck Music" where dialogue occurs.
85. List all subtitles separately, verbatim from script, labeling "Spoken", "Inner Monologue", or "Voice-over".
96. Ensure conflict in first 3s, suspense at end. Adjust order if current doesn't achieve this.
10
11Script:
12(Paste here)
13
14Material List:
15(e.g., Segment 1, 8s, Yang Yang says "Actually, I am the Poet Immortal", finger distortion at 7s)
Miller - inline image

View image description

ALT

And there you have it, one episode done.

Advanced: Elevating Visuals from "Watchable" to "Beautiful"

Once the workflow runs, the next question is: How to make visuals more cinematic?

This section shares my lighting and camera techniques. Top three rules here; full quick-reference attached at end.

1. Specify light concretely, don't just write "cinematic."

AI doesn't understand "premium feel." It understands four things: Direction (front/side/back), Hard/Soft, Cool/Warm, Bright/Dark. Simplest trick: Change front light to side light; faces instantly gain dimension.

2. Lock one lighting setup per scene.

Within a scene, every segment must have identical light specs: direction, temperature, time of day. If previous segment was "warm afternoon light," don't switch to "cool white light" next, or edits will look disjointed.

Yang Yang's drama uses distinct setups: Daytime academy scenes use bright, warm, transparent afternoon hard light. Night candlelight scenes use red candles outlining silhouettes from side-back, preserving shadow detail.

3. Include six elements per shot.

Camera movement, shot size, angle, subject, motion, lighting. Missing one leads AI to guess, costing retries.

Copy directly 👇 (Template first, example second: Yang Yang's climax close-up)

text
1Template: (Camera Movement), (Shot Size), (Angle), (Subject: Who, Expression), (Motion: Amount); (Lighting: Direction, Hard/Soft, Warm/Cool, Bright/Dark); (Focal Length), cinematic texture, realistic skin texture, layered shadows, clear undeformed face.
2
3Example: Static camera, extreme close-up, eye-level, Yang Yang passionate gaze forward, barely moving except brightening eyes and lifting chin; Side light, warm shifting to bright white, light flickering in eyes, blurred background; 50mm, cinematic texture, realistic skin texture, layered shadows, clear undeformed face.

Finally: 9 Most Common AI Mistakes

Short drama pitfalls mostly fall into these 9 categories, with solutions referenced:

  1. Characters look different in each segment: Generate identity board, feed every segment (1.3)
  2. Scenes change randomly: Feed only one scene image per segment (2.5)
  3. Teleportation/left-right flips: Write positions before actions (2.2)
  4. Characters mouthing inner thoughts: Distinguish "Spoken", "Inner Monologue", "Voice-over" (2.2)
  5. Same character, varying voice: Hardcode voice description or use first segment audio as reference (2.4)
  6. Static visuals during long dialogue: Overlay listener reaction shots (3.2)
  7. Retries yield similar results: Workflow seed is fixed; change number to vary (2.3)
  8. Props resize randomly: Specify dimensions in images and prompts (1.3)
  9. Lack of cinematic feel: Specify light concretely, include all six shot elements (Advanced)
Miller - inline image

View image description

ALT

Finally, here is a finished clip made with this workflow 👇

[Embed Finished Video Post]

Appendix: Lighting Quick-Reference and Claude Skill Pack

Lighting and Camera Quick-Reference. When stuck on shot prompts, I consult these 25 pages: Light direction, shot size, movement. Available online or as PDF download 👇

miliagent.com/light

AI Short Drama Skill Pack. I turned this workflow into 3 Claude skills: Scriptwriting, Prompt Engineering, H3 Production. Install in Claude, say "I want to make an AI short drama episode, guide me step by step," and it walks you through. Downloads and installation instructions in the resource library 👇

miliagent.com/library

RunningHub membership and RH coin prices cited are from November 2026. Seedance 2.5 price on RunningHub is the August launch promo. Volcano Engine and MiniMax prices are standard official rates. Platforms adjust prices; verify current rates.

This article is long; bookmark it and follow step-by-step during production.

What story do you want to tell with AI? Share a one-sentence synopsis in comments. I'll pick 3 and help write the first episode script.

Переробити в YouMind

Перетворіть одну віральну статтю на повноцінний робочий процес

Збирайте джерела, розшифровуйте патерни, створюйте матеріали, пишіть чернетки та поширюйте контент в одному AI-робочому просторі.

Дослідити YouMind
Для авторів

Перетворіть свій Markdown на охайну статтю для 𝕏

Коли ви публікуєте власні лонгріди, зображення, таблиці та блоки коду роблять форматування в 𝕏 складним. YouMind перетворює повну чернетку в Markdown на чисту статтю для 𝕏, готову до публікації.

Спробувати Markdown для 𝕏

Більше патернів для аналізу

Останні віральні статті

Переглянути більше віральних статей