FULL AI YOUTUBE SHORTS WORKFLOW (ChatGPT 6 Astra / Seedance / NanoBanana)

Made by Alexandrus — full breakdown of how I go from a blank idea to a finished animated short, start to finish. Text-only, no editing software tutorial fluff, just the actual steps + prompt structure. Feel free to copy this process for your own channel.

============================================================================= STEP 1 — GENERATING CHANNEL NAME IDEAS

Before touching any visuals, lock the channel identity. I ask an AI chat model (ChatGPT / a general LLM) for a batch of names, giving it constraints up front rather than a vague "give me channel names":

PROMPT USED: "Give me 20 YouTube channel name ideas for a channel that posts short, emotional animated stories in a simple, hand-drawn stick-figure art style with cinematic storytelling (think: quiet, bittersweet, relatable human moments — not comedy, not kids content). The name should sound like an actual studio/brand, not a generic word mashup. Avoid anything that sounds like a stock footage channel. Give a one-line reason why each name works."

I run the list through a gut-check filter afterward:

Is it easy to say out loud and spell from memory?

Is the .com / handle actually available? (quick manual check)

Does it hint at the TONE (emotional, quiet, cinematic) not just the visual style (stick figures)?

Pick one, lock it, move on — don't keep re-rolling this step forever.

============================================================================= STEP 2 — GENERATING CHARACTER IDEAS

Next, brainstorm the cast BEFORE any story exists. I ask for character concepts that fit the channel's tone and art style constraints:

PROMPT USED (chat model, text only, no image yet): "I'm building a YouTube channel of short emotional stories in a simple hand-drawn stick-figure style — thick ink lines, minimal flat color, cream paper background, no shading. Give me 5 possible main character concepts. For each: a name, one signature visual trait that makes them instantly recognizable in silhouette (a hair shape, a prop, an accessory — something readable even in a blurry thumbnail), a simple personality in one sentence, and why that trait supports emotional storytelling rather than just looking cute."

Why the "signature visual trait" constraint matters: with a minimal stick-figure style, you don't have facial detail to rely on — the character has to be identifiable by silhouette alone (a cowlick, a scarf, a prop like a pocket watch). Bake that into the brainstorm, not as an afterthought later.

Pick the concept with the strongest single silhouette hook, not the most "designed" one.

============================================================================= STEP 3 — LOCKING IN THE CHARACTER AND CONSISTENCY

This is the step people skip and regret. Before generating a single story scene, generate a dedicated character reference chart with an image model (I use NanoBanana / Gemini-family image models for this — they hold linework style well across generations):

PROMPT STRUCTURE USED: "SUBJECT: character reference chart of [NAME], a simple hand-drawn stick-figure character for an emotional-story illustration series. Cream/off-white paper texture background, soft even light, no harsh shadows. LAYOUT — TOP ROW: 4 views at consistent scale, labelled: FRONT / SIDE / BACK / FRONT CLOSE-UP. LAYOUT — BOTTOM ROW: labelled macro crops of [SIGNATURE TRAIT] close up / [KEY PROP] close up / [WARDROBE COLOR] swatch with hex code / LINE STYLE close-up. APPEARANCE: round, slightly wobbly hand-drawn head, thick soft black ink outline (imperfect, slightly hand-wobbled, not vector-clean), small dot eyes, one simple curved mouth line, [SIGNATURE SILHOUETTE DETAIL]. WARDROBE: [one plain solid accent color], no patterns, identical across all views. STYLE: thick soft black ink lines on cream paper texture, minimal flat color — black line art, cream background, ONE accent color only, no gradients, no shading, no photoreal rendering, no watermark."

Then generate an EXPRESSION SHEET off that same chart (attach the chart image as a reference so the model matches it exactly):

"SUBJECT: expression reference sheet of [NAME] (from the attached reference image). 100% matches the reference — same head shape, same signature detail, same ink style, same wardrobe. LAYOUT: 6 labelled head-and-shoulders panels: HAPPY / SAD / SURPRISED / ANGRY / SCARED / CALM. Only eyes/mouth/eyebrows change between panels — head shape, signature detail, and wardrobe stay identical in every panel."

THE CONSISTENCY TRICK THAT MATTERS MOST: from this point on, every single future image prompt gets the character chart attached as a reference image, AND restates the identifying details in text (exact hair/prop/wardrobe description) even though the reference is attached. Don't rely on the model "remembering" the character between separate generations — restate it every time, every prompt, no exceptions. This is the single biggest driver of whether a 20-scene video looks consistent or like 20 different fan-art drawings of the same character.

If a second character appears in the story, generate their chart the same way, attaching the FIRST character's chart as a style reference so both share construction rules (same head shape, same line weight) while having one clearly different trait so they're distinguishable in silhouette.

============================================================================= STEP 4 — COMING UP WITH STORY IDEAS

With character(s) locked, brainstorm story concepts that specifically use their trait/prop as the emotional engine — not stories the character happens to be dropped into.

PROMPT USED (chat model): "I have a character named [NAME] whose defining trait/prop is [TRAIT/PROP]. Give me 6 short story concepts (60-90 seconds when narrated) for an emotional animated short that makes THIS specific trait/prop the emotional core of the story — not just decoration. Each pitch: one paragraph, a clear emotional arc (setup → escalation → turn → resolution), and the one image that should be the thumbnail/hook."

Pick the pitch where the prop/trait's payoff lands hardest emotionally at the end — that's usually the strongest one, and it's the one people share.

============================================================================= STEP 5 — LOCKING IN THE STORY AND CREATING SCENES

Once a story pitch is picked, break it into a numbered scene list — not a paragraph script. Each scene block needs:

Narration line (the exact VO text for that beat)

Visual description (camera framing + action)

Characters present

Location

Emotional intent

Example scene block: Narration: "Not even for her." Visual: She sits on a park bench, waving warmly as he passes. Characters: [Main], [Friend] Location: Park bench Emotion: Warm (from her), oblivious (from him)

Aim for ~15-20 scene blocks for a 60-90s short — that's roughly one beat per 3-4 seconds of narration once padded with breathing room.

Then generate the actual scene images, one per block, always:

attaching the locked character chart(s) as references

restating the character's identifying details in the text prompt

using a consistent reused STYLE BLOCK across every single scene so the linework/color treatment never drifts, e.g.:

"A cinematic hand-drawn stick-figure illustration in the exact same style as the reference: thick soft black ink lines on cream paper texture, simple rounded stick-figure body with circle joints, minimal flat color with subtle color accents, clean sketch lines, emotional atmosphere, no text or lettering anywhere in the image."

...with only the scene-specific action appended after that fixed block.

For flashback/memory beats specifically, describe the treatment explicitly (e.g. "a soft ghostly translucent memory-vision, faded and semi-transparent" and specify a dim grey — not solid black — silhouette so linework detail stays readable) rather than leaving it to the model's interpretation.

Also generate one contact-sheet grid image of all scenes side by side at this point — cheap sanity check on pacing and consistency before moving forward.

============================================================================= STEP 6 — REVIEWING THE GENERATIONS AND FIXING IF NEEDED

Go scene by scene against a short checklist before approving anything:

Does the character match the locked reference chart (silhouette, prop, color) in THIS scene, not just close-enough?

Did any stray text/lettering sneak into the image?

Does the composition read clearly at small size (thumbnail-scale)?

Does the emotional beat match what the scene block called for?

For anything that fails the checklist, regenerate with a MORE specific prompt, not a softer one — call out exactly what went wrong and correct it explicitly. Example correction pattern that works well:

"[CHARACTER] MUST MATCH REFERENCE EXACTLY: plain round head with [exact detail], NOT [the thing that went wrong last time]. [OTHER CHARACTER] MUST MATCH REFERENCE EXACTLY: [exact detail], with NO [specific wrong color/shape that appeared]."

Naming the specific failure in the correction prompt fixes it far more reliably than just re-rolling the same prompt again.

============================================================================= STEP 7 — VOICE-OVER + VIDEO STRUCTURE APPROVAL

Before generating any audio or video, lock the full structure as a plan and get it approved (by yourself, a collaborator, or whoever signs off) — this is the cheapest point in the whole pipeline to change something.

The structure to approve covers:

Full narration script, scene by scene, final wording

Voice choice (tone/gender/pacing — pick one voice and stay consistent for the whole video)

Total runtime target

Which scenes get a camera move vs. a static hold, and which motion type (push in, pull back, pan, slow zoom) matches each emotional beat

Where music/SFX should sit versus where silence should be left to breathe

The hook/thumbnail moment and the final closing beat

Only once this is signed off do you move to actual generation — treat this step like a storyboard-approval gate, not a formality.

============================================================================= STEP 8 — GENERATING THE VIDEO

With structure approved, generate in this order:

NARRATION AUDIO — generate one audio clip per scene line (not one long track cut up afterward), using the approved voice consistently throughout. Measure each clip's exact duration afterward — don't estimate it, that duration drives every downstream timing decision.

SCENE IMAGES — generate (or finalize, if step 5/6 already produced them) every scene still, same recipe as Step 5/6.

MOTION — for most scenes, apply a programmatic camera move (push-in, pull-back, pan, slow zoom) over the still image rather than generating AI video for every beat — it's far cheaper, keeps the linework perfectly crisp, and stays 100% consistent since nothing is being redrawn frame by frame. Pick the motion type per scene to match its emotional beat from the approved structure.

AI VIDEO GENERATION (used selectively, not for every scene) — for beats that need actual character motion (a wave, a running cycle, a hand typing) rather than just a camera move, generate short clips from a keyframe still using an image-to-video model (Seedance-class models work well here), with a motion-only prompt describing ONLY what should move and explicitly stating the art style must stay static otherwise, e.g.:

"The hand-drawn 2D ink illustration style stays completely static in linework and color — only [the specific motion] happens. No new elements, no distortion of the character or background."

Keep native audio off on these generations and add music/SFX separately — it gives far more control in the final mix than baked-in generated audio.

ASSEMBLY — normalize every clip to the same fps/format first (this avoids silent failures later), then fold them together with short crossfades between most scenes, using a deliberate hard cut (no crossfade) at any spot where the story has a tonal break rather than a smooth transition. Add a short fade-to-black + fade-out audio at the very end.

THUMBNAIL — generate this as its own dedicated image, not a frame pulled from the final render, so it can use tighter framing and higher contrast than any single in-video frame would have.

FINAL QA + EXPORT CHECK — pull a few frame extractions at key timestamps to check crossfade timing and motion smoothness. Then check the FINAL export's pixel format before uploading anywhere — some encoders default to a 4:4:4 chroma profile that plays fine locally but silently fails or gets rejected on YouTube/other platforms. Re-encode to standard yuv420p if so:

ffmpeg -y -i final_export.mp4 -c:v libx264 -profile:v high -pix_fmt yuv420p -crf 18 -preset slow -movflags +faststart -c:a aac -b:a 192k final_export_fixed.mp4

============================================================================= RECAP

Channel name → character brainstorm → locked character chart + restated consistency on every prompt → story pitch built around that character's trait → scene-by-scene breakdown and generation → strict review/fix pass with specific corrections → full structure approved before any audio/video → narration recorded per line → stills animated mostly via camera moves, AI video generation used only for beats needing real character motion → crossfade assembly → dedicated thumbnail → QA → codec-safe export.

No manual frame-by-frame animation anywhere in this pipeline — consistency comes from re-attaching the character reference every time and restating its details in text, and motion comes primarily from programmatic camera moves over stills, with AI video generation reserved for the handful of beats that actually need it.

— Alexandrus