Style guide vs. per-shot prompting for AI video: which produces more consistent results?

Loading a full directorial treatment document into the invideo agent before any generation produces dramatically more consistent results than per-shot prompting. The agent holds camera, lighting, palette, composition, and mood directives across every frame without re-explaining — eliminating style drift and the cognitive overhead of rebuilding context each clip.

Use per-shot prompting only when you're generating one-off clips with no shared visual language. For anything multi-scene — a short film, an episode, a brand promo — load a complete visual language document into the invideo agent first, then direct conversationally.

Here's how the two approaches compare across the dimensions that actually matter in production:

Dimension	Treatment-loaded direction	Per-shot prompting
Style consistency across scenes	Held by persistent context — every shot gated against the locked doc	Drifts shot-to-shot as phrasings vary
Character identity across scenes	Holds with locked character sheets (no LoRA needed)	Breaks past ~5 scenes without reference assets
Cognitive load on the director	Conversational, on-set register — "hold on him till he lunges"	High — rebuilding style language every prompt
Cost per finished minute	$315–$750 across four documented productions	Not reliably measurable — iteration cost spikes with drift
Production days (short film)	2–5 days end-to-end	Manual-prompting equivalent estimated at 1+ week for a 2-minute film
Where the model choice lives	invideo agent routes each shot to Seedance 2.0, Kling, Veo, Recraft, Nano Banana, or GPT-Image-2	You pick and re-specify per prompt
Best fit	Multi-scene films, episodes, promos, anything with shared visual language	One-off hero clips, isolated experiments

The rest of the answer fills in the how:

Treatment-loaded direction (the high-consistency workflow). Write a structured visual language document — 14 sections covering camera, angles, colour tone, atmosphere, mood, lighting, composition, movement, film palettes, prompt templates, negative prompts, and a quick-reference card — and upload it once to a creative producer agent at project start along with your full script. The invideo agent locks it into persistent context and gates every generation against it. In one 70-second short film, a 25-page treatment doc produced a complete production at $750 (3,000 credits) over two days, with the invideo agent autonomously sequencing a six-shot ending and applying a slow-shutter motion smear effect pulled from page 17 without being prompted. In a 90-second horror short built on a Fincher-style protocol (9-step shot design, 8-step color grading, 85:15 dark-to-light ratio, 2.40:1 hard matte), the invideo agent flagged a shadow drift to blue-green during Scene 1 generation and pulled the locked Stage A rule to correct it — total cost $870, 4,100 credits, two days.

Per-shot prompting (where it actually fits). Writing a fresh prompt for each clip — even with a per-shot formula like subject / action / camera / mood — forces you to rebuild style language every time, and small phrasings drift across shots. It works for isolated experiments or single hero clips. For anything longer, it leaves you mentally wrecked and can't sustain character identity across 20+ scenes without reference assets locked upfront.

What makes the treatment approach hold. Six things, all stackable in the invideo agent: (1) the 9-element prompt assembly order — camera spec, lens & aspect ratio, lighting source, palette, composition, atmosphere, mood register, film/DP attribution, negative prompt — applied identically to every generation; (2) named tonal modes with hex values (e.g. "Mode A — split-toned amber and emerald") so palette is reproducible, not described; (3) explicit negative constraints in the style block ("this MUST look and feel like the reference — not live action, not photorealistic") to stop model drift; (4) a section isolating exceptions and adaptations so the invideo agent doesn't misapply generalised rules; (5) a sound architecture module — half of what makes a horror film land is what you hear before what you see; (6) per-stage "what never to do" lists so the invideo agent makes autonomous calls in your absence.

Solve character consistency separately — it's the harder problem. Style holds easily from the treatment; character identity does not. Generate four reference options per character and lock the best one, build multi-angle character sheets with four angles plus face and mid-angle closeups, remove objects from characters' hands before generating turnarounds, and include close-up panels for scars and accessories. Across a 70-second short, this approach held two characters consistent across every scene with no LoRA. When characters evolve mid-film (a trinket added per beat, a costume change), generate a distinct character sheet per beat rather than one master sheet.

Validate the doc before you generate at scale. Ask the invideo agent to apply your director's style to a genre that director never worked in — a courtroom thriller through a horror lens, for example. If the invideo agent asks clarifying questions and returns stylistically coherent output, the doc has been internalized as grammar. If it mirrors surface aesthetics only, rewrite. Once validated, a three-word continuation prompt like "everything should match" is enough to carry character, lighting, lens grammar, and spatial logic across multi-shot sequences.

Multi-modal references complement the doc — they don't replace it. Where text under-specifies, hand off physical inputs: shoot a mock POV on your phone and upload it for a difficult camera move, hand-sketch a complex physical configuration and pass the drawing through to image generation, batch references thematically (spatial logic in one batch, color theory in another) and tell the invideo agent explicitly what to take and what to leave out from each. The treatment doc carries the persistent grammar; multi-modal references handle the specific shot the doc doesn't anticipate.

Route models per shot inside the same agent. Different models hold different things well: Seedance 2.0 reference-to-video carries character and location context across continuous segments where start/end-frame methods can't; Kling handles multi-shot sequences natively; Veo has its own lane for cinematic motion. Every roster model runs inside invideo, so you don't pick a platform per shot — the invideo agent routes your treatment-loaded context to the right one. For image work, Recraft produces facial portraits with pores, lines, and stubble; Nano Banana and GPT-Image-2 handle character sheets and grid generation.

The cost case for treatment-loaded direction. Across documented productions, treatment-loaded workflows ran $315–$750 per finished minute — a 70-second short at $750 ($750/min), a 90-second horror short at $870 (~$580/min), a 3-minute animated episode at $950 ($315/min), and a 2-minute brand promo at $1,500 ($750/min). The brand promo replaced an estimated $100,000–$500,000 traditional shoot at roughly 20x time reduction (3 days vs ~2 months). The savings aren't from cheaper generations — they're from not regenerating, because the doc-locked invideo agent gets shots right earlier in the iteration curve.

As Hridaye, invideo's creative director, puts it: "One agent that reads your treatment once and holds every directive across every shot, every scene. No re-prompting. No drift. So now, you direct, and the Agent remembers."

These are the workflow ranges that work in production — what holds for your film depends on length, complexity, and whether character identity or stylistic continuity is your harder problem.

Watch some of these to see what works for you:

One agent that reads your treatment once and holds every directive across every shot, every scene. No re-prompting. No drift. So now, you direct, and the Agent remembers.

— Hridaye, invideo's creative director