I spent Thursday night writing a prompt that had the phrase "@Image1" in it four times and felt, for a second, like I was coding instead of directing. That's where Kling O3 has put us. The guide making the rounds right now is the first prompting document I've read this year that changed how I structure a brief before I even open the tool.
Nobody tells you this when you start using reference-image syntax: it's a new grammar, not a shortcut. You're not uploading a photo and hoping the model gets the vibe. You're tagging it, @Image1, @Image2, and then writing sentences that point directly at it, like "the woman in @Image1 carries the vase from @Image2 across a courtyard." That's a different cognitive mode than the moodboard-and-vibes approach most of us learned on Gen-2 and early Kling. It's closer to writing stage directions for actors who've never met. The actors happen to be pixels, so the stage direction has to be unambiguous or the model improvises badly.
A shot list beats a paragraph
The biggest mental shift from this guide, and the one I've already started applying on a Web3 summit teaser I'm cutting this week, is that Kling O3 wants a shot list, not a paragraph. Kling's own official guide backs this up with a five-layer order that I now have taped above my monitor: Scene, Characters, Action, Camera, Audio and Style. I've stopped deviating from that order.
For multi-shot clips inside a single five-second generation, you label it like an actual edit decision list. "Shot 1 (2s): wide shot, woman enters frame from left, soft morning light." "Shot 2 (3s): close-up, she sets the vase down, ambient courtyard birdsong." It reads more like an AD's call sheet than a piece of writing, and it works because the model isn't improvising structure. It's filling in detail inside a structure you already gave it.
For the action layer specifically, resist the urge to stack verbs. I did this constantly when I started: "she walks over, picks up the vase, admires it, then sets it down while laughing." Kling chokes on that. One clean verb chain, "carries the vase across a courtyard and sets it down on a low wall," reads cleaner to the model and renders cleaner on screen.
For dialogue scenes, the audio layer needs the same discipline. Label your speakers Character A and Character B, not "the man" and "the woman," and attach an actual emotional direction: angry, whispering, softly speaking. I ignored this for months and got flat, over-articulated line readings. The moment I started writing "Character A, whispering and afraid: we need to go" instead of just the dialogue line, the delivery got noticeably more human.
Kling wants negative prompts. Runway hates them
This is the part of the brief that made me sit up, because it's a direct contradiction between two tools I use the same week, sometimes the same project. Kling 3.0 defaults to a weirdly sanitized, everyone's-smiling, perfect-skin aesthetic unless you fight it. The fix is negative prompting: blur, distort, low quality, shaky camera, cartoon, anime, text, watermark, deformed face, extra limbs, stacked in every single prompt if you want anything gritty or real. I run a version of this list as a saved snippet now, because typing it out each time is tedious and I will forget half of it under deadline pressure. Runway's official position is the opposite. Their Gen-4 documentation says flat out that negative phrasing isn't supported and can produce unpredictable or backwards results. If you want a locked-off camera on Runway, you write "locked camera." You don't write "no camera movement." Telling the model what not to do apparently confuses it more than it helps.
This might be the most useful thing in the whole brief, more than the @Image syntax. It's a reminder that technique doesn't transfer. I've watched junior editors on client projects copy-paste a negative prompt block from a Kling tutorial straight into Runway and wonder why the output got worse. The two tools have different training biases and different prompt grammars, and you have to learn each one on its own terms. Treating "prompting" as one universal skill is how you waste a production day.
The reference-image trick has a dark twin
I can't write about @Image reference syntax without mentioning the Tom Cruise versus Brad Pitt fight video that got the Motion Picture Association to publicly denounce what they called massive infringement, tied to Seedance 2.0. That's the same underlying capability we're all excited about this week, just pointed at someone's likeness without consent. The tools don't distinguish between "the woman in @Image1" being your paid model on a brand shoot and "the woman in @Image1" being a public figure who never signed a release. The grammar is identical. The ethics sit entirely with the operator, and I don't think enough people in this industry are sitting with that.
On client work, this means I'm now explicit in contracts about what reference images are licensed for generation input, not just final use. A year and a half ago nobody asked for that clause. Now it's standard.
It's all converging into one browser tab
Sora, as a brand, is just gone. App shut down in April, API discontinued in September. Half the "Sora prompt guide" content still ranking on Google is instructions for a product that no longer exists, which is a good reminder that this field has a shelf life measured in months, not years.
What's replacing the single-model mindset is aggregation. Higgsfield shipped Katana in early October, basically letting you pick a preset, feed in images, and have Claude reconstruct the edit logic for you. A few days before that they added Nano Banana 2.1 with proper 1K/2K/4K resolution tiers. The platform now sits on top of Seedance 2.0, Kling 3.0, Wan 2.6, Veo 3, Sora 2, Hailuo 02, Flux Kontext, GPT Image, and Seedream 5.0, all in one interface. I used to have six browser tabs open per project. Now I have one, and it routes to whichever model handles the shot best.
Kling 4.0 is supposedly landing this month too, with the Flash tier already in limited hands since late September, pushing 4K, stereo audio, and 10-bit HDR. I haven't graded the output myself yet, so I'm withholding judgment on the quality jump, but the spec sheet alone tells you where the ceiling is heading.
The actual skill in this job isn't picking the "best" model anymore. It's knowing which model's particular grammar fits the shot you're trying to get: Kling's shot-list structure, Runway's no-negatives rule, whatever Kling 4.0 turns out to want. Each tool is its own dialect, and right now that matters more than which one is technically the best.