Phase 4 · A Reel

A Reel

A 5–8 second video sequence with a character voice. Pick a camera style — vlog or documentary — anchor the character, build one scene by hand, then scale.

Duration · ~4 hoursSteps · 4Output · Reel + Creator's Note + AI Audit

1. Anchor — character reference sheet

Your character must look the same in every scene. The fix is a reference sheet: one image showing the same person from many angles. You'll attach it to every future scene.

AOpen your project in your AI workspace

Start a new thread named Untold Stories Video — [your character's name].
Paste in your character description (3–4 sentences from Phase 1 + Phase 2).
Upload your Phase 1 character image.

BGenerate the reference sheet

Send this prompt to your image tool with your Phase 1 image attached as the input:

Reference sheet prompt

Create a professional character reference sheet of the exact same person from the reference image, on a plain white background. LAYOUT Top row: four close-up head shots — front-facing, left profile, right profile, back of head. Bottom row: three full-body shots — front, side, back. CONSISTENCY Replicate facial structure, skin tone, hair colour and styling, eye colour, eye shape from the reference image. PERIOD-ACCURATE CLOTHING (same in every view) [describe — fabric, cut, period, condition] LIGHTING Soft, neutral studio lighting. Flat and even. No shadows. STYLE Photorealistic, sharp micro-detail, RAW photograph. NOT ALLOWED Modern clothing, modern objects, fantasy elements, sunglasses, masks, anything anachronistic for [TIME / PLACE].

CSave it

Download the image. Name it [character]-reference-sheet.png.
Pin it inside your thread.
You'll attach this to every future scene.

CHECK

Same person across all 7 views?
Period-accurate clothing matches across views?
No text, captions, or bubbles drawn into the image?
Style consistent — no random changes in lighting or style?
If unsure, write: "We are not sure if our chosen style fits because…"

2. Build Scene 1's still

You're not making the video yet. You're making one still picture — the first frame of the video clip you'll generate next. The look you settle on becomes the recipe for everything else.

APick your camera style first

Sensitivity check — read first

If your character is someone whose voice was historically silenced — an enslaved person, a child labourer, a prisoner — the "selfie influencer" framing can feel disrespectful. Documentary often serves them better.

	Vlog / Selfie	Documentary / Observed
What it looks like	The character holds the camera and talks directly to the viewer.	The camera watches the character from outside, like a film scene.
Best for	Personality-driven stories, characters who feel like guides, modern-feeling reels.	Quiet observed moments, characters who should not break the fourth wall.
Sensitivity check	Would a vlog framing make this character look like a tourist in their own life?	Does the story need the character speaking to you, or to someone in the scene?
Examples that fit	A merchant showing their goods. A scholar describing their work.	A child slave in a kitchen. A prisoner waiting. A widow at a grave.

BPull Scene 1 from your storyboard

Note for Scene 1: who is in it / where / what is happening / why this scene matters.

CFill in the image prompt

Vlog version:

Scene 1 still — vlog prompt

A still image - the first frame of a short video. [CHARACTER NAME], a [age] [role] from [TIME / PLACE], holds a small hand-held camera in selfie style with one arm outstretched. The camera is not visible. [CHARACTER] is in [SETTING - physical location, time of day, what is around them]. The location must feel alive. [CHARACTER] looks directly at the camera with [EXPRESSION]. PERIOD DETAILS (must match the reference sheet) - Clothing: [details] - Hairstyle / headwear: [details] - Setting only contains period objects: [list 2-3] LIGHTING [natural, period-appropriate] STYLE Photorealistic, cinematic, hyperrealistic micro-detail. NOT ALLOWED Modern clothing, modern objects, modern buildings, plastic, electrical wires, anachronisms.

Documentary version · same template but the camera observes from outside; character does not look at camera, holds nothing in selfie style.

DGenerate, check, iterate

Run with your image tool, attach your character reference sheet as the input image.

Regenerate ONLY if: different person / clearly modern detail / wrong framing / object missing.

Don't regenerate for: slight expression / pose variations / cosmetics.

CHECK

Same character as the reference sheet?
No AI-rendered text in the image?
Period-appropriate clothing, setting, lighting?
One clear action that matches the brief?
No modern slips?

3. Bring Scene 1 to life

Now the still moves and speaks. The clip is 5–8 seconds. Your character says one short sentence. Video gen is the most expensive step — get the prompt right before generating.

AWrite ONE short sentence of dialogue

Length — 12–25 words for a 5–8 second clip.
Style — how this character would actually talk. Not modern. Not over-dramatic.
One thought, not a paragraph.
Read it out loud with a stopwatch. Lands in 5–8 seconds at a natural pace? If not, cut.

BFill in the image-to-video prompt

Image-to-video prompt — vlog example

[CHARACTER NAME] films [himself / herself / themselves] vlogging style, looking at the camera. VOICE [tone - e.g. "calm and thoughtful"] [accent - e.g. "regional accent for [TIME/PLACE]"] Natural human delivery — slight pauses and breath. No modern slang. DIALOGUE [CHARACTER] says: "[ONE SHORT SENTENCE - 12 to 25 words]" ACTION WHILE SPEAKING [one small gesture - e.g. "the wind catches their hair"] BACKGROUND [same as the still - minimal change]. AMBIENT SOUND [period-appropriate - e.g. "distant market chatter"] DURATION 5-second clip. One clear continuous action. NOT ALLOWED Modern slang, modern accents, dramatic over-acting, character holding a phone, music swells.

CGenerate the clip

In your video tool, use:

First-frame image — your Scene 1 still
Reference image — your character reference sheet
Prompt — the one above

Set duration to 5 seconds (up to 8 if dialogue needs it). The model creates the video and the voice — no separate voice tool needed.

If the call is rejected

"Unsupported video generation request" → drop the reference image, keep only the first-frame image. Your still already contains the character.
"durationSeconds out of bound" → try 8 seconds instead of 5. Veo sometimes rejects 5s for image-to-video; 8s always works.

CHECK

Dialogue fits the clip? Was the character cut off, or were there long awkward silences? Adjust length and regenerate.
Same person? Does the character look like the same person as in the reference sheet?
Stays in period? No modern slip that crept in during animation?
Audio matches setting? The ambient sound is period-appropriate, not random?
Tone right? Voice not too modern, not over-acted, not silly when the story is serious?
Regenerate only for major character changes, unclear dialogue, modern objects, or clearly wrong tone. Ignore tiny mouth, blink, or cosmetic issues.

4. Recipe → Scale → Edit → Publish

One scene works. That pattern is your recipe. Save it once, scale to the rest, edit in CapCut, attach Creator's Note + AI Audit. Ship.

ASave the recipe as a skill

Reflect first (5 min, no AI): what 3 details did I repeat? What kept the character / dialogue right? What did AI keep getting wrong?

Save scene recipe as a skill

Save my scene recipe as a skill called "[Character] Scene Generator". The skill should know: my character, evidence, style, image-prompt template, video-prompt template, and "NOT ALLOWED" list. When I describe a new scene from my storyboard, the skill should produce: an image prompt, an image-to-video prompt, one short dialogue sentence (12-25 words), and an evidence note marking fact / smart guess / uncertain.

Group running short on time? Save the recipe as a project document instead — paste it manually for each scene.

Test before scaling — ask the skill for Scene 2's plan only (no image yet). If it drifts, fix the recipe now.

BScale to the rest

For each remaining scene: plan → audit → still → clip → save → next.

A scene plan should give you: title, what happens, image prompt, video prompt, dialogue, evidence ID(s) (F1, SG2), things to verify.

AUDIT every scene plan BEFORE generating

Perspective — character's knowledge only.
Evidence by ID — every claim cites F1 / SG2 / U3.
No modern slips — voice, prompt, or visuals.
Dialogue length — 12–25 words, sayable in 5–8s.
Storyboard match + dignity.
Third-reference trick — once 1–2 scenes are approved, attach one approved still as a third reference for the rest. Locks consistency harder.

CEdit in CapCut

Import all clips, place in storyboard order, trim awkward starts/ends.
Add a title card at start and a closing card with the Creator's Note.
Captions: auto-caption, lower third, simple sans-serif.
Soft transitions only. No music — the native audio is enough.
Export: 9:16 (reels) or 16:9 (classroom), 1080p, 30fps, MP4.

Pro tip · generate your title and closing card images with the same AI workspace — keeps the visual world consistent to the last frame.

When you finish early

Share your skill with the class — make the description and "when to use" clear so another group can swap in their character.
Project webpage — host the video, Creator's Note, AI Audit, museum object photo, evidence list at one shareable URL.
Alternative perspective companion — generate one short clip from the opposite viewpoint on the same object; place side-by-side in CapCut.
Quality review — trade videos with another group; review using their CHECK panels and AI Audit template.

CHECKFinal review

Same character across all scenes so far? Evidence respected?
Nothing the character couldn't know about?
Dialogue length right (5–8s, 1 sentence)? No modern slips?
Matches storyboard description?
Recipe still serving you, or does the recipe need an update?
If two scenes fail in the same way, stop generating and fix the recipe first. Don't waste time or credits on the same mistake.