Make AI videos longer than the clip limit, and keep them hanging together

Find your tool's real limit

Every model has a maximum clip length and its own way of going further. As of October 2026: Kling 4.0 and Seedance 2.5 make up to 30 seconds in one go, Kling 3.0 up to 15, Luma Ray 3.2 5 or 10, Runway Gen-4.5 up to 10, and Veo 3.1 4, 6 or 8. The full list, with each tool's extend feature, is under "Put it to work".

Write two numbers down before you plan: the longest clip your tool makes, and how it continues one (an extend button, first and last frames, or reference images). Everything else is built around them.

Plan in shots, not one long take

A longer video that holds together is a sequence of shots. Each shot does one thing: an establishing view, an action, a reaction, a detail. Keep each one inside your tool's limit, and shorter is safer: the longer a single generation runs, the more room it has to wander.

Give Claude the idea, the target length and your tool's limit, and ask for a numbered shot list with a length for each shot that adds up to the total.

I want to make a [LENGTH]-second video about [IDEA] for [PLATFORM, e.g. Instagram Reels 9:16]. My video tool is [TOOL], which makes clips of up to [MAX] seconds and continues them with [EXTEND / FIRST AND LAST FRAME / REFERENCES]. Break it into a numbered shot list. For each shot give: length in seconds (never above [MAX]), what we see, the camera move, and how it joins the shot before it (hard cut, extend, last frame as the next first frame, or first and last frame). Make the lengths add up to [LENGTH]. Keep it to the shots the story needs.

Lock the look with a continuity block

Models drift between generations: a jacket changes colour, a face changes shape, the sun moves. The cure is boring repetition. Write one block that describes the character, wardrobe, location, light and colour grade, and paste it word for word into every prompt.

If your tool takes reference images (Kling 4.0 takes up to 15, Seedance 2.5 and Gemini Omni take them too), make one clean still of the character and one of the location first, and attach the same files to every shot.

Write a continuity block for this video in under 80 words. Cover: the main character (age, build, hair, face, one distinctive detail), their exact clothes with colours, the location, the time of day and where the light comes from, and the colour grade. Plain, concrete words only, no adjectives like "stunning". I will paste this block unchanged into every shot prompt, so it must not mention any action or camera move.

Pick a join for every cut

There are four ways to get from one clip to the next. A hard cut is a new generation with the same continuity block and references: use it when the camera angle changes, which is how real edits work anyway. Extend continues the same shot from where it ends: use it when the action must carry on unbroken.

Last frame to first frame: take the final frame of one clip and use it as the start image of the next, in any tool with image-to-video. First and last frame: give the model both ends and let it fill the middle, which is the most control you can have over where a shot lands. Kling, Seedance 2.5, Luma and Gemini Omni all support a start and end frame (Luma only on 5-second clips).

Make the anchor shot first

Generate the shot that defines the look, usually the first or the hero shot, and do not move on until it is right. Every later shot copies it, so a flaw here gets copied too.

Then work forwards. For extends, prompt only what happens next and keep the camera and light the same. For last-frame joins, grab the frame with the command under "Put it to work" and use it as the next clip's start image with the same continuity block.

Write the prompt for shot [N] of my shot list. Start with the continuity block, word for word. Then: what happens in this shot in [SECONDS] seconds, the camera move (one move only), and how it ends, so the next shot can pick up from that last frame. If this shot is an extend, describe only what happens next and say "same camera, same lighting, continuous action, no cut".

Check for drift after every step

Each extension builds on the clip before, so small errors add up: a face softens, a hand gains a finger, the colours slide warmer. Look at the last frame of every new clip next to your anchor shot before you continue.

When something drifts, do not keep extending a bad clip. Go back to the last good one and generate again from there. Google says Gemini Omni 1.1 Flash looks at up to 10 seconds of the previous scene when it extends, where earlier models used only the final second, which is why newer extends hold together better than older ones.

Here are two frames: the first is my anchor shot, the second is the last frame of my newest clip. Compare them for continuity: face and hair, clothes and colours, props, location details, light direction and colour grade. List every difference you can see, most noticeable first, and tell me whether I can keep going or should regenerate from the previous clip.

Join the clips and hide the seams

Put the clips in order in an editor, or join them with ffmpeg. When one clip starts from another's last frame, that frame appears twice: trim a frame or two from the start of the second clip so the motion does not stutter.

Lay one music track or ambient sound bed across the whole video. Continuous sound makes separate shots feel like one piece more than anything you can do to the pictures.

Clip limits and extend options by tool (as of October 2026)

Kling 4.0 (launched 15 October 2026): up to 30 seconds in one pass, up to 10 keyframe images and up to 15 reference assets. Kling 3.0: 3 to 15 seconds with multi-shot prompts. Kling's Extend adds 4 to 5 seconds at a time, up to 3 minutes in total, and has to use the same model and mode as the original clip. Check in the app whether Extend is offered for the model you use.

Seedance 2.5 (ByteDance): up to 30 seconds in a single generation, with the option to extend twice. Takes a first frame, a last frame and reference images.

Google Veo 3.1, Veo 3.1 Fast and Veo 3.1 Lite: 4, 6 or 8 seconds (reference images only at 8 seconds on Veo 3.1). In Google Flow, Extend works only on Veo-generated clips, and extended clips cannot then use Flow's insert, remove or camera edits.

Google Gemini Omni 1.1 Flash (Flow, the Gemini app and the API): extends a scene in 10-second steps up to 40 seconds in total, looking at up to 10 seconds of what came before. Supports first and last frames. Google's API docs say editing uploaded videos is not available in the EEA, Switzerland and the UK.

Luma Ray 3.2: 5 or 10 seconds. 10 seconds does not work with a start frame, an end frame or HDR. Extend continues a previous generation forwards, or backwards (the new clip goes before it), in standard dynamic range only.

Runway Gen-4.5: up to 10 seconds per clip, with an optional first-frame image. Runway also runs other makers' models, including Seedance 2.5 and Kling, and their own limits apply there.

Grab the last frame of a clip

This saves the final frame as a PNG, ready to use as the start image of the next shot. It reads only the last half-second of the file, so it is instant.

Use the highest-quality file you have. Every frame you take from a compressed download loses a little detail, and that loss carries into the next clip.

ffmpeg -sseof -0.5 -i clip-01.mp4 -update 1 last-01.png

Join the finished clips into one file

Make a text file called list.txt with one line per clip, in order: file 'clip-01.mp4', then file 'clip-02.mp4', and so on. Then run the command below.

This joins the clips without re-encoding, which works when they all came from the same tool at the same size and frame rate. If ffmpeg complains, or the clips came from different tools, use an editor instead.

ffmpeg -f concat -safe 0 -i list.txt -c copy full-video.mp4