How to Brief an AI Video Project

The most expensive mistake in an AI video project happens before the first generation — a brief that's thinner than it feels. This is the discovery pass worth running first: what to ask, which narrative framework to reach for, and the review habit that catches a weak prompt before it costs a generation credit instead of after.

Discovery, before anything else

Infer what you can from context; ask only what's genuinely missing. Group questions so it feels like being understood, not interrogated.

Question	Why it matters
Outcome	What should the viewer do or feel after watching? (Remember the brand, click, sign up, trust the product, laugh then feel relief — these need different films.)
Audience	Who is this for — one primary persona, not a demographic range?
Placement	Where will this actually live? Aspect ratio and length follow from this.
Source material	What already exists — real footage, brand assets, a character idea only, or fully generated from scratch?
Who has to stay the same person	Anyone recurring or visually important needs a locked identity reference before video generation starts, not stills improvised per scene.
Brand/product	What must appear, and what must never appear?
Tone	Pick 2–3 anchors, not a mood-board essay.
Hard constraints	What's absolutely off the table?

Reflect it back before drafting anything

Summarize in 3–5 sentences: "You're making a [format] for [audience] on [platform]. The viewer should [outcome]. Tone is [tone]. We'll [source approach] and avoid [constraints]." Then ask: does this match what you're trying to make? Catching a misread here costs a sentence; catching it after a draft costs the draft.

Pick a storyline framework — don't invent structure from scratch

Framework	Shape	Best for
3-act (Hook → Message → CTA)	~5s hook, ~17s single idea, ~8s clear next step	30-second TV or premium digital spots — one idea per spot, roughly 65–75 spoken words total if voiceover-led
5-shot performance	Hook → Problem → Solution → Proof → CTA	Fast-paced social/performance ads — trim each generated clip to its 1–3 highest-signal seconds in the edit
Hook–Story–Offer	Attention → compressed emotion or comedy → offer	Anything from a 6-second bumper to a 30-second spot

One shot per generation call, assembled afterward — don't try to prompt an entire ad in a single generation. This is broad industry consensus, not an Ekly-specific quirk.

A creative-brief template

Copy, fill, and confirm with whoever's requesting the video before generating anything

Title — Objective (one KPI) — Audience (one persona: who, where, language) — Platform (ratio, length, sound-on vs. sound-off default) — Source (existing footage / full AI / hybrid) — Hook (exact first-2-seconds visual or line) — Key message (one sentence) — Proof/demo beat — CTA — Tone (2-3 words) — Visual style — Characters (names, roles, relationship) — Must show — Must not — Dialogue table (beat / speaker / exact line) — Deliverables (hero length, cutdowns, aspect ratios) — Constraints (broadcast/family-safe, legal/claims).

Review every prompt before you generate — a rubric, not a vibe check

Score each dimension Pass / Fix / Block — Block means don't generate until it's fixed.

Check	Pass looks like	Fix if...
Hook specified	The opening action/visual is first in the prompt	Move the hook to the top; name the first 2 seconds explicitly
One idea	A single gag, or a single line, or a single product moment	Split it into separate beats
Mute-readable	The story is clear with the sound off (for social)	Add visual contrast or a physical gag instead of relying on audio
CTA separated	Any call-to-action lives in the edit, not the generation prompt	Remove URLs/on-screen text from the generation prompt itself
Brand kindness	The people in the ad aren't mocked; the product reads as relief	Reframe the copy

Helping someone who doesn't know their storyline yet

Anchor on the real situation, not a feature list. What actually started this — a real frustration, a piece of footage, an event?
Offer two spine options, don't leave it open-ended. A concrete choice between two named directions moves things forward faster than one more open question.
Steal a proven framework rather than inventing structure from nothing.
Write dialogue last. Lock the visual beats first — lines are easy to swap once the shots exist.
Promote important "background" people to real characters. If someone recurs across scenes, lock their identity reference before video generation, rather than letting each shot re-cast them.

"Make it premium" with nothing more specific? Propose a named menu, don't guess blind

A vague upgrade request answered with one guessed redo just restarts the same guessing loop when that guess gets rejected too. The fix that actually works: propose several distinctly named directions — each a one-line visual philosophy, not just "try again but nicer" — and let the person choose a lane, then push harder inside that one choice. This is the same principle used for a scene that isn't landing (see comedy timing) — a diagnostic ladder with named options, not one more guess.

The three-layer model for a recurring campaign

If a project has an identity (a character, a spokesperson) that has to hold across multiple videos, vary one layer at a time when iterating:

Layer	What it covers	Vary it...
Identity	Face, outfit, voice — shouldn't change per variant	Rarely, and only deliberately
Message	Hook, script, CTA, pacing	Per variant — this is what A/B testing usually means
Rendering	Scene, camera, platform export	Per shot

If identity is drifting across shots, reduce how much the scene/camera is changing before rewriting the whole prompt — don't fix an identity problem by changing everything at once.