| name | Dave Clark: Cinematic AI Direction |
| description | The taste layer above tool craft — why AI video reads flat vs cinematic, hybrid pipeline architecture, and shot-list-before-prompt direction discipline from the CCO of Promise |
| version | 1.0 |
| format | completion-engine |
| workflows | 3 |
Dave Clark: Cinematic AI Direction
Co-Founder & Chief Creative Officer, Promise. Two decades directing film and commercials
(Coca-Cola, Snapchat, HP, Warner Bros., Intel) before generative tools existed. Borrowing Time,
Dismal Swamp, Battalion, Another (Cannes Next 2024), NinjaPunk, My Friend Zeph, Hardcore 94.
What he adds that nobody else in the house does: the judgment layer above prompt craft. Everyone else
teaches syntax. Clark is the one who can name — mechanically, in a DP's vocabulary — why the image you just
generated is dead, and the one shipping AI work through platform QC and studio chain-of-title.
The one-line thesis
"To me, the future of AI is just filmmaking… the goal is for AI to become an afterthought."
— Marché du Film, 2025-06-24
If the AI is the interesting part, the film has failed.
The core claim
Flatness is almost never a prompting failure. It is a directing failure wearing a prompting costume.
Six of the eight causes of flat AI video are decided before generation or after it — in selection, coverage,
cadence, lighting specification, capture layer, and whether anything is at stake. Only two are prompt problems,
and none of the eight is a model-choice problem. Swapping generators moves two of them slightly and the other
six not at all. That is why prompt-craft skills plateau at "impressive clip" and never reach "watchable film."
Available Workflows
| # | Workflow | Produces | Use When |
|---|
| 01 | Flat-to-Cinematic Audit | Shot-level diagnosis: why it reads flat, the fix per defect, and what's not worth fixing | Something "looks AI", a client says "make it more cinematic", or a reel is technically clean and dead |
| 02 | Hybrid Pipeline Plan | Layer map + persistence register + delivery spec + provenance plan + risk register | The piece is longer than a clip, mixes live action with generated material, or has to clear QC |
| 03 | Shot List → Generation Brief | Directed shot list as camera reports, with look card, cadence plan and generation protocol | Starting anything longer than one image, or previous attempts came out as a slideshow |
Quick Reference
- Genius Context: genius.md — load before any workflow. Carries the flat-vs-cinematic axis,
18 patterns, 11 signature moves, the quality rubric, and the observed visual signature.
- Source notes & fidelity ledger: references/source-notes.md — every source dated,
every unverified claim flagged.
Doctrine: model-independence
Everything in the core method is model-independent. It survives Runway → Veo → Sora → whatever ships next
quarter, because none of it is about a model. It is about direction, selection, coverage, cadence, light,
colour space and provenance.
Tool-specific mechanics are quarantined in genius.md Appendix A — "Era-Bound Mechanics (2023–24), verify
before use." No workflow and no execution prompt in this skill depends on a single line of that appendix.
If swapping the generator invalidates an output, the output was wrong.
The eight causes of flat, in likelihood order
- One generation deep — a selection problem, not a prompting one
- Metronome cutting — uniform clip length is the loudest AI tell there is
- Unmotivated light — you can't point at where it's coming from; shadows lifted grey
- Clean air — nothing physically between camera and subject, so the planes collapse
- No capture layer — digitally immaculate, therefore never photographed
- No coverage — every shot a new setup, so nothing reads as a scene
- Adjective prompting — "cinematic, moody, epic" is a wish, not a direction
- Nothing at stake — you can't say what it's about in a sentence
Escalations for real delivery: drift · edit survivability (colour space, bit depth) · provenance.
Signature moves (short list)
Rule of Five · Composite the take (don't pick it) · Mask the fidelity back in · Retime rather than accept the
clip length · The coverage pair · The global capture layer · Phenomenon over category · Extend, don't crop ·
Externalise what must persist · Spec the delivery before you generate · Direct the performance, synthesise
the colour.
Routing
- Diagnosing bad AI video / "make it more cinematic" → workflow 01
- Longer-form, hybrid, or QC-bound work → workflow 02
- Anything before generation starts → workflow 03
- Prompt syntax and style codes → NOT this skill.
skills/cinema-worldbuilder-pro/, /art-direct,
/creative-prompt. Clark is deliberately upstream of prompt syntax.
- Ad-shaped AI video →
skills/pj-accetturo-ai-video/ (PJ optimises for the platform; Clark optimises for
the release print — when they disagree, the deliverable decides)
Execution Prompts (structure-pure v2)
3 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
- Dave Clark — Flat-to-Cinematic Audit —
skills/dave-clark/references/prompts-v2/flat-to-cinematic-audit.md
- Dave Clark — Hybrid Pipeline Plan —
skills/dave-clark/references/prompts-v2/hybrid-pipeline-plan.md
- Dave Clark — Shot List & Generation Brief —
skills/dave-clark/references/prompts-v2/shot-list-generation-brief.md
Fidelity note
Three prompts, not twelve. One per deliverable the corpus can honestly carry. Clark never published a
flat-vs-cinematic framework, never gives a lighting ratio, f-stop, colour temperature or grading value, and
never publishes a lens table — so none appear here. The eight-cause diagnostic is Clark-derived, assembled
by this extraction from his stated practice plus the observed signature of his 2025 director reel, with
per-check source tags. Full ledger: references/source-notes.md.