Sections 4-5: AI Clones and AI Video — the 5-Step Workflow, Image-to-Video Prompting, First/Last Frame, HeyGen, Runway Aleph, ElevenLabs Voice, and Wan 2.2 Replace

'The fundamental five step workflow to create a video doesn't change whether you use the traditional tools or AI tools': step one an idea, 'the seed'; two the script; three the storyboard - the first visual step, done as stills because 'the best strategy for making detailed AI videos is... image to video'; four video creation; five editing. Framed as a 'hybrid human plus AI workflow' against one-click tools that 'usually create terrible looking videos' and undifferentiated slop. Step four has three routes: human-AI hybrid (you on camera plus AI assets), image-to-video (Kling, Veo), and long-form talking-head tools (HeyGen, Hedra, Lip Sync on fal.ai). A film scholar's analysis of 15,000 movies puts the average shot at 2.5-4 s, so 5-10 s clip caps are the norm, not a constraint; editing tips follow. Demo: a horror short built end to end.

Nano Banana Pro accepts up to 14 reference images; describing your appearance in words loses to uploading photos. Photo rules: one person per image, clear facial detail, good light, the phone's rear 1x lens not the selfie camera. Baseline: a close-up plus a full-body shot. Best: a character sheet - one composited PNG with two full-body angles (front, side) and four head close-ups (neutral, smiling, two profiles), the format character designers have always used; a free Canva template is linked. Make a sheet per look (beard vs clean-shaven, outfits). Prompt six elements every time: subject and action, environment, outfit, expression, lighting, style - or substitute a reference image for any of them ('this man at the spooky castle from the reference image'). Strong example: 'create a cinematic ultra realistic image of the man from the reference image standing on top of a snowy mountain at sunrise... dark winter jacket... confident expression. Soft golden sunlight... dramatic clouds.' Weak: 'make me look cool on a mountain.'

Image-to-video (still first, then animate) is the default: the image 'gives the AI video tool the context on the subject, the time of day, the style' so output lands closer to intent, and stills are cheaper to iterate than 30-second-plus paid generations. Tools: Kling as default (cheaper; versions 2.1/2.6/3.0 at recording - 'the newest version is always going to be the best'), Veo as fallback. Kling settings: model, aspect ratio, 5 vs 10 s (he stays at 5), Generate Sound toggle. Prompt only what the still cannot show: (1) subject action; (2) camera movement - say 'no camera movement' or 'tripod shot', or name pan left/right, tilt up/down, tracking, push in/out, jib up/down, 'the main camera movements that Hollywood directors use for 90% of the shots'; (3) emotion/mood/tone - running to train or running from a threat. Static shots read as more professional. Case study: ChatGPT idea and script -> ElevenLabs voiceover -> Nano Banana character and shots -> Kling ('The dinosaur looks at the lightning and then runs away') -> Premiere; shoot coverage; the same method supplies B-roll for long-form.

Required: subject; shot composition/framing (close-up, wide, medium waist-up, long, side profile, macro, drone, over-the-shoulder); environment/setting; visual style. Bonus: lighting style (direction and intensity in plain words), colour of light (daylight white vs neon or orange), time-of-day light (sunset, sunrise, blue hour), atmosphere (fog, rain, snow, wind, smoke, dust, embers), colour palette (saturated, noir, or 'like Blade Runner'). Tip: paste the worksheet into ChatGPT with a one-line scene description and let it draft the prompt, then edit rather than accept.

Two Nano Banana Pro stills plus a prompt; Veo 3.1 or Kling 2.5 on PromptEdit connect them. Veo 3.1: slightly better realism and auto sound effects, but 'can struggle... with complex transitions' - it cuts or fades; Kling 2.5 'is really, really good at creatively transitioning' (car into Transformer, bridging unrelated honeymoon clips) at slightly lower quality. Durations 4/6/8 s; 16:9 used. Examples: barn and horse, a hat transition, a logo animation. Shot sequencing: chain pairs so each clip's first frame is the previous clip's last - the camera commercial floats, deconstructs, reforms in a hand, then the reviewer smiles; each pair made in Nano Banana ('keeping the camera about the same size in the frame, but now it's being held by a blonde girl... field of flowers at sunset'), animated in Kling 2.5, cut in Premiere.

Video-to-video: an existing clip plus a text change. Examples verbatim: 'change the scene to take place at night with a large full moon' (mixed), 'engulf the dinosaur in flame', 'create a new wider angle version of this clip', 'change the weather outside to snow instead of rain', 'outside of the window a large fire rages and burns', 'create a new angle from directly behind the person... symmetrical position, subtle push in.' Best on additive, localized effects and reframing; full day/night reversals were weaker. 5-second cap. Subscription on Runway directly (not PromptEdit), affordable entry tier plus ~60 credits per generation, 1,000 credits for $10. The hand-on-fire finish: real clip + Aleph 'arm and hand is on fire' clip cut together in Premiere with sound effects.

Input one image (a Nano Banana clone) plus a typed script with a preset voice, or an uploaded audio file - a real recording or an ElevenLabs clone - that HeyGen lip-syncs. Settings: talking style (stable/default/expressive), 720p vs 1080p (cost scales), 16:9 / 9:16 / 1:1. Weakness: background motion drifts ('the smoke will almost start going backwards') - keep backgrounds simple. Practices: split long pieces into per-segment generations so one flub does not re-run five minutes; the cost-saving hybrid - HeyGen for the first 30 s, a middle minute and the last 30 s of a ten-minute video, ElevenLabs voiceover plus B-roll between; for dynamic scenes with a consistent voice, generate in Kling/Veo and run a voice-swap pass. Also on PromptEdit now, where it 'used to be' website-only at a high price.

All via PromptEdit's Audio tab. Voiceover: paste the script, pick a voice, set stability (low = expressive, high = monotone; he likes 0.1-0.4), language, generate, download - commercials, faceless channels, audiobooks, course narration. Clone: name it, upload a long clean sample (his: 11 minutes), leave 'remove background noise' unchecked unless the source is noisy, tune stability to your natural delivery; feed the clone to HeyGen or to Seedance 2.0, which - unlike Kling/Veo that 'use whatever voice it wants' - accepts your audio to drive a cinematic clip. Swap: export audio-only MP3 from the Premiere timeline, upload to Voice Changer, choose the target voice (preset or clone), stability, generate one consistent track.

Screenshot a frame of your video; in Nano Banana Pro redraw the subject ('replace the man in the image with a humanoid robot with glowing blue eyes. Everything else... the same'); feed the original video and the new image into Wan 2.2 Replace. Settings: guidance scale, inference steps (raise if results look weird, at time cost), video quality (max), fast mode. Movement, facial motion and lip-sync carry over. Large or 4K files fail - compress, lower resolution, or split. Uses: low-budget shorts, social hooks, two-character skits filmed solo; the Gollum motion-capture pipeline without the suit. Wan 2.2 Move (background swap) is 'the next video' - not captioned here.