What real projects taught us (case-study lessons)

Every rule above was earned on shipped work. A tight selection of the lessons:

• OWAMBE (painterly fashion film): the piece needed to be small to feel large. Every failed iteration had too much in it — era references, party montages, dual art registers. The version that worked had less than half the content of the original brief, one unified painted register, and a culturally specific anchor (aso ebi, money spraying, the celebrant being honoured). Its ending never fades out; the sisters freeze mid-arrangement posing for a photo and the piece becomes the painting.
• LANIASO (cosmic trailer): the tension between personal and colossal scale is the trailer. A generic astronaut became the robotic consciousness of a Black woman scientist, her holographic face visible inside her visor, her Igbo-lineage voice whispering "Is anyone there?" against binary black holes. Either scale alone would have been generic.
• Type II Dyson: cosmic scale needs intimate detail. A robotic termite-drone repairing a hexagonal panel pauses for one frame to look at the lens. Without that moment, the planet-sized sphere would have stayed an abstract CGI object.

▶ Type II — the motion experiment behind that lesson: planetary-scale structure kept human by one intimate beat

• Lagos Lagoon Impact: disaster works through specificity, not scale. One white van the audience learns to care about, one droplet on the lens, one car spinning on its roof in brown water, one Yoruba curse under the operator's breath. Each specific detail outperforms the kilometre-high water column.
• Captain Morgan packshot: describe the brand, never name it. Content filters scan prompt text, not uploaded reference images. "Captain Morgan" became "premium dark spiced rum bottle, red-and-gold label"; the reference image carried 100% of the brand identity, the prompt carried 0%. This pattern is reusable for any branded packshot work — essential knowledge for anyone making AI ads.
• Ceremonial Lounge: turn model limitations into directorial choices. Models smudge distant faces, so the figures were specified with faces turned away. A mural kept trying to animate, so it was locked: "flat painted surface… NOT 3D, NOT animated, NOT hologram. Painting stays still."

The keyword clusters that control light, lens and motion

You do not need a thousand keywords. You need eight clusters, used deliberately. These are the strongest, with when to reach for each:

1. Camera and lens — use when a beat feels visually flat. Assign a motivated lens per beat: "14mm ultra-wide," "27mm cathedral wide," "50mm portrait," "100mm macro"; moves like "locked tripod," "slow forward push," "orbital arc," "rack focus."
2. Lighting sources and qualities — use when renders look like flat CGI. "Single key from upper-left," "volumetric god-rays through architectural gaps," "warm molten-gold firelight from hidden practicals off-frame," "Rembrandt lighting," "single key, no fill," "museum-lighting logic."
3. Texture and material — use when the world feels mass-produced. "Terracotta composite," "oxidized bronze," "obsidian flooring," "gold leaf accents," "visible canvas texture," plus imperfections: "imperfect, handmade, lived-in," "compressed cushions," "lifted blacks," "warm halation."
4. Motion (camera, character, environmental) — use when clips feel static or chaotic. "Eyes find the lens at [time] and HOLD," "knowing breath, micro-smile barely registering," "fine sand drifts," "naira notes suspended mid-fall," "earrings swing once as head turns."
5. Pacing and transitions — use to set tempo before the model invents its own. "Slow contemplative pacing," "held three seconds," "one unbroken camera breath, no cuts" for reverence; "frame jerks," "camera makes mistakes," "cut on motion" for urgency.
6. Realism register — use to stop style drift between photoreal, painterly and found-footage. "Photographic, not CGI," "cinema look, warm film grain, halation, lifted blacks," "one unified painterly register only," "handheld, raw, no stabilization."
7. Cinematic mood — use to set the emotional register as global state at the top of the prompt. "Sacred social sanctuary that already exists," "reverent observation," "documentary stellar physics," "recovered visual memory from another era."
8. Continuity and global state — use on any multi-beat piece. "Carries forward: [element]," "same key light across all beats," "wind direction frame-right to frame-left throughout," plus global blocks: LIGHTING, COLOR HIERARCHY, ATMOSPHERE, CAMERA, CONSTRAINTS.

Marketing teams: these eight clusters are your internal prompt library's table of contents. Put each in a shared doc, add your brand's specifics under each heading, and every prompt anyone on the team writes starts from the same vocabulary.

A master prompt template you can copy today

This is the fashion film template, verbatim from the working method — the same architecture behind the OWAMBE piece. Drop your specifics into the brackets:

[TITLE] — [DURATION]s painterly fashion film, [ASPECT RATIO]. Single continuous shot, slow forward push. [NUMBER OF BEATS] sustained held moments. Hypnotic, observational. Recovered visual memory.

THE SUBJECT: [identity block]. [Skin/feature treatment with "NOT" specifications]. [Wardrobe identity locked]. [Posture or gestural defaults]. Identity locked.

[OTHER FIGURES IF ANY]: [Description with same painterly register]. All in same [style/cloth/world]. Distinct individuals, not clones.

STYLE: [primary aesthetic — e.g., "Hyperreal painterly portraiture, 2D hand-painted, visible-but-controlled brushwork on background, smoother on skin"]. [Most refined treatment on subject's face — focal painted object]. No CGI gloss. No photorealism. One unified register.

COLOR HIERARCHY: [percentages and qualitative locks]. Color is hierarchy, not decoration.

CAMERA: Single continuous slow forward push across [duration]. Subtle anamorphic breathing. No cuts.

[BEAT 1 — PRESENCE] [held moment of subject alone in negative space]

[BEAT 2 — DEVELOPMENT] [environmental detail or supporting figures enter at frame edges]

[BEAT 3 — RESOLUTION] [either an emotional landing or a snapshot freeze]

CONSTRAINTS: One register only. Color as hierarchy. [60%+ negative space per frame]. Identity locked. No CGI gloss. No saturation drift.

▶ Nextberries Fashion Film — exactly this register of prompt in action: the painterly fashion-film architecture above, filled in and shipped

Anatomy: what each clause does

• The header line sets duration, aspect ratio and register as global state before anything else — every beat inherits from it. "Recovered visual memory" is an atmosphere thesis, not decoration.
• THE SUBJECT block with "Identity locked" fights the number-one AI video failure: identity drift. The "NOT" specifications ("NOT flat black") pre-empt known failure modes.
• The OTHER FIGURES clause ("distinct individuals, not clones") stops the model from copy-pasting your subject into every supporting figure.
• STYLE and COLOR HIERARCHY enforce the single-register law and percentage discipline — the model renders value structure instead of random saturation.
• CAMERA as its own block locks one continuous move so the model cannot invent cuts.
• Three named beats give the model structure without overloading it — presence, development, resolution. The OWAMBE lesson lives here: less content, more weight.
• CONSTRAINTS at the end are your negative prompts stated as law, closing the loopholes the model would otherwise exploit.

For teams making product ads, the same architecture adapts directly: the product cinematic variant swaps the subject block for @image_[N] = PRODUCT. [Generic description of the product without brand name]. Match exactly. Identity locked, zero drift. and ends its packshot beat with "Label readable, never deforming" — the Captain Morgan filter-safe pattern built straight into the template.

Automating content creation: from single prompts to scene systems

If you are automating content production — batch-generating campaign variants, localising one concept across markets — the unit of work is no longer the prompt. It is the scene system: a structured set of prompts that share character identity locks, world specifications, colour hierarchies, atmospheric continuities and camera grammars. You write the system once; each shot inherits from it. Build a world bible per project, reuse it verbatim across every prompt, and validate each generation against it before shipping.

Two production disciplines make this reliable at scale. First, validate length before you paste: target 3450–3499 characters for Runway and 3950–3999 for Seedance, and if you are over cap, cut decoration words, then redundant adjectives, then the weakest beat. Second, rate every generation across seven categories out of 100 — composition, motion, light, material, continuity, emotion, discipline. Anything under 80 in a category tells you exactly where to revise. Producer-grade work is 95+ across the board.

Where this is going: prompting becomes directing

The first generation of AI video was about prompting: write what you want and hope. The second generation is about directing: write what you want with such specificity that the model has no degrees of freedom. In five years, "prompt engineering" as a separate discipline will probably not exist. What will exist is the role that has existed since the silent era — the person who chooses what the audience sees.

The skills that will transfer: cinematographic vocabulary, motion hierarchy, colour discipline, continuity logic, emotional pacing, material specification, audience attention management. The skills that will become obsolete: prompt tricks, keyword stuffing, model-specific workarounds. Invest in the durable skills; treat the platform hacks as scaffolding.

Every limitation will eventually be solved — identity drift, continuity, long durations. When that happens, the bottleneck shifts entirely to taste. And the question that outlives every platform change: what does the audience need to feel at the end of this 15 seconds? If you can answer that in one sentence, the prompt writes itself.

Cinema is the art of making time feel like something. AI video is cinema, eventually — we are still in its silent era. You have inherited a working method. Now do better than we did with it.

Next in the series

Part 4 puts this method under commercial pressure: the brand-campaign case studies — how these templates and keyword clusters perform on real ads for real clients, from filter-safe packshots to heritage campaigns, with the numbers and the revisions shown. Read it next: How the method performs on real brand campaigns.

Learn this live

Reading a template is one thing; directing a model in the room with someone who ships this work is another. At PxLabs in Lagos we teach exactly this method, live:

• Generative Cinematography — the full directing discipline: camera grammar, lighting vocabulary, beat structure and scene systems.
• Synthography 101 — the foundations, if you are starting from zero.
• Join the PxLabs community — Nigeria's AI creative community, where members share prompt libraries, critiques and shipped work.
• Building AI ads for your brand, or want a custom team workshop? Talk to us.