Kling TIFF Panel: The Script Is Still the Blueprint — Write Shots Before You Prompt

Kling TIFF Panel: The Script Is Still the Blueprint — Write Shots Before You Prompt

The script is still the blueprint. That is the operator rule from Kling AI’s 12 September 2026 panel at TIFF: The Market, not a slogan for a prompt box. Evolutionary Films producer Diane Shorthouse told the room she is not thinking about prompts first. She is thinking about shots: where the camera goes, what the performance is, what it cuts to, which assets exist. ClipCanva is not affiliated with Kling, Canva, Google, or TIFF. For a campaign clip, lock the story on the AI script generator, turn each beat into a shot card, lock a still, then generate a batch take on the AI video generator or image to video. Do not ask one box to invent the plot, the eyeline, and the pack shot.

Launch facts (12 September 2026)

Source	Public claim	Workflow implication
Variety: Kling AI TIFF Market panel	Dated 12 September 2026. Panel title: “A New Creative Paradigm: Consistent, Cinematic AI Video Powered by Kling AI.” Opening morning of the inaugural TIFF: The Market. Moderator: Tony Pu, Kling AI.	This is a production conversation, not a model-card drop. Treat quotes as process, not as a SKU you can click.
Same report	Shorthouse: “The script is still the blueprint, so I’m not thinking about prompts, I’m thinking about shots.” Camera, performance, cut, assets, and who has the skill — including where AI artists sit.	Write a shot list before a motion prompt. A prompt without a shot is a vibe.
Same report	Teams generate about 20 takes of a beat; 19 are wrong. “Those 19 are important. It’s human judgment that you need to know what’s wrong.”	Budget fails. Do not ship take one because the motion looked expensive.
Same report	Control request: keep character, costume, and environment; change only the eyeline. “I need them to obey me more.”	One variable per retry. If the model restyles the set when you asked for an eyeline, it is a fail.
Same report	Issa Sissoko (Amissa Studios): every ~90 days the floor moves — from three-second clips toward 15 seconds with more consistency. Pitch-deck trailers can get a story picked up. Hybrid live-action still struggles to keep an actor’s performance intact while adding AI.	Duration is improving. Performance lock is not free. Do not treat a 15-second ceiling as a feature you have until the live control says so.
Same report	Advice to new filmmakers: learn cameras, lighting, and performance, then put that knowledge into prompts. Sissoko: learn to write stories first; writing helps the prompts.	Script desk before generator desk.
Canva Help: Create a video clip	Canva AI Create a video clip: 8 seconds with audio, 6 seconds without. English prompts only. Optional reference image. Generation may take up to 2 minutes. Separate from Magic Media text-to-video and from Magic Video assembly.	A clip desk is short and batch. It will not hold a feature previz. Split the script into clip-length beats.
Google Flow creative controls	27 August 2026: Gemini Omni 1.1 Flash in Flow adds start and end frames, 1080p/4K export, and faster lower-credit 360p drafts before a higher-resolution render.	Once the shot is locked, first/last frames are how you specify the camera path. Draft cheap, then keep.

Confirm the live generator before you quote duration, audio, or credit cost. The Variety piece is a panel report. It does not publish Kling’s current second cap as a public spec sheet.

Pick the desk before you prompt

The fail is treating “write a cinematic prompt” as the same job as “direct the scene.” Shorthouse’s split is the map: script and shots first, generation second.

Job	Where it actually runs	What you must supply	What the system may invent	Fail condition
Story, claim, spoken line	AI script generator	Audience, one claim, duration, forbidden words	Extra scenes, unearned superlatives	Prompting video before the line is signed
Shot list / previz packet	Script output + a scene card you keep	Shot size, lens feel, camera move, cut point, duration	Decorative adjectives that fight the still	One paragraph that tries to be five setups
Still that must stay on-model	Stills desk, then image to video	Approved photo as first frame	Camera path, secondary motion, warped type	Animating a doodle that still has the wrong SKU
Finished campaign clip	Batch generator: AI video generator or a named model on ClipCanva Compare	Locked still and/or locked caption, crop, seconds	Extra props, extra logos, a second spoken line	Shipping the first of 20 because it moved
Cheap motion test	Flow 360p draft (documented Google path) or a low-res batch pass	Same first frame and caption as the keeper	Texture, lighting drift	Spending 4K credits on a caption you have not locked
Existing footage you must match	AI video summarizer then script	Source cut, what must stay from the plate	New blocking the plate does not support	Asking the model to “add AI” on top of a performance you cannot break
Reusable camera language	Prompt ideas	Shot type, lighting, camera, duration	A chat you will lose	Changing five variables on every retry
Templated social cut from clips you already have	Canva Magic Video (assembly, not a generator)	Clips or photos you already approved	Template, music, a longer cut	Treating assembly as text-to-video

If the brief is “this bottle already exists; make it move 8 seconds,” that is image-to-video from a locked still. If the brief is “we do not know the shots yet,” that is script. If the brief is “keep the actor, change only the eyeline,” that is one-variable control, and most current clip desks will fail it. Name the desk before you open a tab.

Script vs shot card vs clip vs assembly

Dimension	Script / blueprint	Shot card	Batch AI clip	Assembly (Magic Video-class)
Official surface	Writer / AI script generator	Your packet, not a model	Canva clip, Flow, ClipCanva generators, Kling-class models	Canva Magic Video and similar
Input	Audience, claim, scenes	One beat: size, move, duration, still	Prompt and/or still; wait for a file	clips or photos you already have
Output	Lines, beats, forbidden claims	One generate request	6–8s class clip on Canva’s clip desk; other desks vary — confirm live	A sequenced cut, not a generated world
Human gate	Legal line, story logic	“Would I call this shot on set?”	Identity, label, extra limbs	Continuity of the files you fed it
Do not use it for	Camera path	Shipping the ad	Writing the plot	Inventing a pack shot that was never shot

Canva’s own split is a useful third map, not a partnership. Create a video clip waits, then hands you 6 or 8 seconds. Magic Media text-to-video sits inside a design. Magic Video sequences footage you already have. Flow’s start and end frames are the documented Google version of “where does the camera need to go?” ClipCanva’s image to video is the operator version of the same idea: a first frame you chose, plus a caption that came from a shot, not from a mood board.

Blueprint workflow: 20 takes, 19 expected fails

Write the packet once. Generate many. Promote only winners. Instant obedience is not the default.

1. Write the blueprint, not the prompt. On the AI script generator, lock audience, one claim, close, duration, and forbidden words. Shorthouse’s line is the test: if you cannot name the shot, you do not have a prompt yet.
2. Break the script into clip-length beats. Canva’s clip desk is 8 seconds with audio or 6 without. Other models may run longer — Sissoko’s “3 seconds toward 15” is a filmmaker observation, not your live cap. If the beat needs 20 seconds of story, it is two or three clips, not one overloaded prompt.
3. Fill a shot card per beat. Shot size, camera move, performance, cut, assets. Example: “MCU bottle, 35mm feel, slow push, label readable last frame, 8s, 9:16, no extra logos.” Park reusable camera language in prompt ideas.
4. Lock the still before motion. Proof type, SKU, hands, costume. Then send that still to image to video. If there is no still, generate on the AI video generator only after the caption is a shot, not a paragraph.
5. Generate a batch, then judge. Budget ~20 takes the way the TIFF panel described. Watch against the card. If the eyeline changed and the set restyled, it is a fail. Re-roll from the same first frame. Change one variable.
6. Use a cheap draft tier if you have one. Flow’s 360p draft, then 1080p/4K, is the documented Google version. Do not spend the keeper credit on a caption you have not locked.
7. Compare models when the first desk warps the label. Use ClipCanva Compare. Do not start a new story in the second model.
8. If you have plate footage, run it through the AI video summarizer so the script respects what was actually shot. Sissoko’s hybrid warning stands: keeping an actor’s performance intact while “adding AI” is a different job from generating a pack shot.

Operator card you can paste:

Output: 8s 9:16 Reel, English, one claim
Desk: script → shot card → still → batch image-to-video
Blueprint: [one sentence story]
Shot: MCU, slow push, label readable last frame
First frame: bottle_hero_v3.png (legal signed)
Caption: [shot card, not a vibe]
Forbidden: extra logos, competitor names, medical claims, extra eyeline changes
Fail if: set restyles when you asked only for eyeline; label unreadable; new VO
Retry: same first frame, change one variable
Quota: 20 takes expected; ship none of the 19 fails
Not this job: one-box “cinematic prompt” with no shot list

Why the 19 wrong takes are the work

The panel is blunt: 19 of 20 generations will be wrong, and those 19 are how you learn the model’s disobedience. That is not a bug report. It is the production math.

A prompt that says “cinematic, highly detailed, masterpiece” does not tell the camera where to go. A script that never becomes a shot card will leak into five setups in one 8-second clip. A still you have not signed will look expensive in motion and still be the wrong SKU.

Shorthouse’s eyeline example is the control test. If you cannot change one thing, you do not have a director’s desk. You have a slot machine. Treat restyles as fails. Keep the first frame. Keep the caption. Change the eyeline or the move or the duration — not all three.

Sissoko’s 90-day note is the other half: clip length and consistency are moving. Your packet should survive a model swap. The blueprint does not. If you stored the job only as a lucky prompt in a chat, the next model will not know the shot.

Creator checklist

• [ ] You wrote a script with one claim before you opened a video model.
• [ ] Each generate request maps to one shot, not the whole film.
• [ ] Duration fits the live desk (Canva clip: 8s with audio / 6s silent — confirm yours).
• [ ] The first frame is an approved still, not a thumbnail from a previous fail.
• [ ] The caption is the shot card, stored in prompt ideas if you will reuse it.
• [ ] You budgeted fails. Take one is not the delivery.
• [ ] One variable changes per retry. Eyeline requests do not restyle the set.
• [ ] Morph, extra limbs, new on-frame type, or a second spoken line is a fail.
• [ ] Plate footage, if any, was read with the AI video summarizer before you asked a model to “add AI.”
• [ ] You did not ask one box to write the legal line, invent the pack shot, and stream the take.

FAQ

What did the Kling TIFF panel actually say?

On 12 September 2026, Variety reported Kling AI’s TIFF: The Market panel. Diane Shorthouse said the script is still the blueprint and that she thinks in shots, not prompts. Her team generates about 20 takes and expects 19 to be wrong. Issa Sissoko said writing stories is the skill that improves prompts, and that hybrid jobs still struggle to keep an actor’s performance while adding AI. ClipCanva is not part of that panel or of Kling.

Should I prompt video directly from a logline?

No. A logline is the script desk. Turn it into shot cards, then into a still plus a caption, then into a batch clip on the AI video generator or image to video. If the model has to invent blocking, it will invent props you cannot legal.

How long should each AI clip be?

Depends on the desk. Canva Help documents 8 seconds with audio or 6 seconds without for Create a video clip, with waits up to two minutes. Sissoko described industry clips moving from about three seconds toward 15 with more consistency. That is not a guarantee on your button. Split the blueprint to the cap you can actually export.

How is this different from a script-to-video button?

Script-to-video tries to jump from story to motion in one request. The TIFF rule is the opposite: story → shots → assets → generate. Assembly tools such as Canva Magic Video still need clips you already have. First and last frames in Google Flow are how you specify the path once the stills exist.

When do I compare models?

When the shot card is stable and the first desk warps identity, type, or the product. Compare on ClipCanva Compare. Do not use a second model to rewrite the story.

The script is the blueprint. The shot is the prompt. The still is the approval gate. Generate 20. Ship the one that obeyed. Do not wait for the model to become the director.