Introducing ClipTap: from one shot to a whole show

Making one AI shot is easy. Write a description, pick a model, and a few minutes later the frame looks almost too good to be true.

Finishing a show is hard. Your lead needs a look locked down, and one isn't enough: at work, at home, after the injury, each with its own reference sheet. Sets need building, and a pharmacy seen from the doorway is a different image from the same pharmacy seen from behind the counter. Prompts get written, then rewritten, then re-rolled; a dozen takes for one shot is routine, and someone has to watch every one of them. Is the blocking right? Did the action finish? Does the framing cut with the previous shot? Is the thing in her hand still there? None of this is work the model does for you.

And that is one person's trouble. Once a short drama is moving, there are hundreds of reference images, thousands of prompts and tens of thousands of takes. They have to be findable, split across several people working at once, reviewed, and someone has to know which version is the right one. A prompt box can't hold that. It takes an industrial-grade way of working.

That is where the ClipTap team comes from. We have served several studios and produced a number of short dramas ourselves. The pits we fell into, the tokens we burned and the methods we settled on became the solutions inside ClipTap. Work this way and you take far fewer detours, deliver faster with steadier quality, and the money you save is usually something you only notice when you do the accounts afterwards.

Stunning frames, continuity errors everywhere

If you have made a short drama with AI, you probably know the feeling. Every shot looks fine on its own. Cut them together and the errors are everywhere. Someone is sitting at the table one second and standing at the door the next; two people face each other, then in the next shot they are walking side by side. The lead's wide shot and her close-up don't look like the same woman: a narrow face from a distance, a round one up close, and by episode three you can barely recognise the face from episode one. The living-room window has moved. The pharmacy counter is on the left, then on the right.

Space, posture, wardrobe, faces: every one of them needs watching, and one flaw is enough to pull the audience out of the story.

Models keep getting stronger. Experience matters just as much

We assumed at first that a better model would make these problems fade. The longer we worked, the more that looked half right. Models do get stronger every year, but each generation has its own temperament: what it loves to draw, what it avoids, which phrasing it understands and which it improvises on. Learn that temperament and you accumulate technique, set after set of it. Every one of those techniques was paid for in tokens, and they are worth a great deal.

Here are a few that we and our peers bought dearly.

A character isn't one card. It's a set of states. Injured, younger, in house clothes: each state gets its own sheet, not one picture with the word "injured" attached. The sheet is a three-view: full front, full back, chest-up close-up, on grey, with no divider lines. Whatever shot you make afterwards, the reference is that one sheet, and the wide shot and the close-up finally belong to the same person.

Sets are drawn per camera position. One pharmacy, looking in from the door and looking out from behind the counter, is two empty plates. Give the model the one the camera is facing.

Write the shot as a physical device. "Medium shot" is vague and the model will interpret it. Write the field of view in degrees, the distance to the subject, locked-off or handheld, who is screen-left and who is screen-right, the height difference in centimetres, and how many people there are. Count lock, height lock, position lock, all three at once, and a crowded frame holds.

Light has to come from something in the scene. A night interior is lit by a lamp, a candle or the window. Write "sunlight" and you get a sun. One lighting key per scene, not rewritten in every shot.

Write emotion as performance, not adjectives. "Cold sweat" and "oily sheen" wipe out the matte skin real people have. For tension, write pressed lips, a furrowed brow, quickening breath. Before each line, say how it is said: volume, rhythm, whether the voice breaks on the last word. Anyone not named keeps their mouth shut, or two people speak at once.

Change one thing at a time. Review against the original sheet, not against the last take, or the drift compounds take by take. Drift shows up mostly between seconds five and eight, so keep segments short. No music and no subtitles by default; those belong to the edit.

Where does all this live? In most teams we meet, in one person's head, in a few documents, or in the history of a few dozen chat windows. When that person leaves, the experience leaves with them.

Why ClipTap earns your trust

ClipTap was built for films and long-form series. Our starting point is simple: a show is not a pile of shots, it is a project. Everything that has to stay the same across ten episodes should be defined once, kept somewhere structured, and pulled from there by every generation that follows.

Script and Story Outline. One script per episode. The premise, the rules, what each episode must achieve, which setups pay off where, and who knows what from which episode on, all live in one Story Outline. The Agent reads it before writing any episode, so episode seven doesn't forget what episode two planted.

The Library. Every person, place and object in the script is an entry. A character has several looks, each with its own sheet; the voice is bound to the character, so whichever look is speaking, it is the same person's voice. Write @Lead in any prompt and her face, this episode's wardrobe and her voice come in with the name, with nothing to describe again. The experience from the previous section lives in how these cards are written and drawn.

Shot lists. One Shotlist per episode, every scene broken into shots, and every shot states who, where, holding what, what framing, how many seconds. A shot is not an isolated prompt. It knows which scene of which episode it belongs to, and what comes before and after.

Multiple takes. Generate several takes for a shot, lay them out on the canvas, and pick the one that cuts with its neighbours as the current take. It drops into the sequence at once, you play it, and if it doesn't work you swap in another without rewriting the prompt.

The Director's Desk. This exists for the rare complex shot. A film's big set pieces, a car chase, a crowd, two people facing off across a whole room, are hard to keep stable in space and camera with words alone. The Desk lets you place the set, the people and the camera in 3D first, lock everyone's position and movement, and have the model shoot that space. Ordinary dialogue scenes never need it. For the one scene that does, it saves a day of reshoots.

Agent and Skills. The ClipTap Agent ties it together. Tell it in plain words what you want, and it reads the whole project: script, Outline, Library, shot lists. Say "break scene three into five shots" and it finds scene three, works out who is in each shot, where and holding what, attaches the matching Library entries and writes the prompts. The whole section of dearly bought experience above is written down as Skills that the Agent follows. Use our short-drama Skill, or write your own team's way of shooting as one, and it remembers it on the next show.

You paste in a script. What comes out is a show

Say you have a ten-episode short drama, two minutes an episode, about a night-shift nurse who discovers someone in the hospital is swapping patients' medication.

You paste the whole script to the Agent. It asks you one card: the style and period, which video model you plan to use, and whether to touch the script where it finds a problem. Then it reads.

When it is done, the script is ten episodes. The Story Outline exists: premise, rules, what each episode has to do, how the pill bottle planted in episode two pays off in six. The nurse, the pharmacist and the inspector who only appears in episode three each become a Library character. The nurse has three looks, uniform, street clothes and how she looks after her suspension in episode nine, each with its own sheet, sharing one voice. The ward, the pharmacy and her rented flat are three sets, each with an empty plate for every camera position that appears. The swapped bottle is a prop with its own reference.

Then the storyboard. Ten episodes, a shot list each, every scene broken into shots, prompt cards already written, and every card already carrying the right people, places and things. Up to this point, not one credit has been spent. The Agent never presses generate for you. It lays out the whole show and stops before the camera rolls, waiting for you to look it over.

Roll camera. Shoot a single shot or queue a whole episode, and see the price first every time. Takes land on the canvas and you pick the one that cuts. By episode six the nurse still has episode one's face and wears what this episode calls for, because every shot references the same her in the Library.

Episode eight has the inspector behind the pharmacy counter and the nurse in the doorway, facing off across the room. For that scene you place the pharmacy, the two of them and the camera on the Director's Desk first, then let the model shoot. The dialogue scenes in the other nine episodes don't need it.

Finally the show leaves ClipTap. We don't make an editor; your Premiere Pro, Final Cut Pro, DaVinci Resolve or CapCut is already good. Each episode's chosen takes are assembled into a timeline that opens in the editor ready to cut.

Recruit Claude and ChatGPT onto the crew

You probably already have a way of working: Claude or ChatGPT writes the script, GPT Image 2 makes the references, Seedance or Kling shoots the video. The part that really costs you is carrying things between windows. The script gets copied out of a chat and pasted into a document, the character description gets typed again and again, the approved face is remembered by its file name, the prompt has three versions and nobody knows which one the current generation used.

ClipTap speaks the open MCP protocol. Connect it and Claude and ChatGPT work inside your project: they read your script and Library, continue breaking down the existing shot list, write prompt cards and build 3D scenes on the Director's Desk. Everything they write lands in the project's structure: dialogue compared line by line with the source, characters resolved to entries that really exist in the Library, seconds computed against the chosen model's limit. The moment Claude finishes, what it wrote is a storyboard ready to shoot, not a block of text for you to carry around.

The work runs on the Claude or ChatGPT plan you already pay for; script, storyboard and prompts cost no ClipTap credits. Generation, the part that spends money, is a permission that stays off until you turn it on. The Claude you pay for every month should be on the crew.

Designed from day one for people and Agents alike

Building this product, one thing became clearer and clearer: a lot of the work inside a project will be done not by people but by Agents. So we wrote down what we had learned as Skills, designed the ClipTap Agent, and opened the door to the Claude, ChatGPT or Codex you bring in from outside.

Software used to be designed as an interface for people. Now it has to be designed for Agents too, and the two want different things. People want buttons that fall to hand. Agents want clear facts: who this character is, how this line was originally written, how many seconds this model can shoot in one segment. So underneath a ClipTap project is one structured set of data, and what a person changes in the interface and what an Agent changes through tools are the same thing. Our own Agent has no privileges. It follows the same methods and passes the same checks as a Claude coming in from outside.

We also decided early that the system guards facts and never decides taste for you. AI may not quietly change a line of dialogue; to change one it has to say which line. A character a shot references must really exist in the Library. What a model can't do, the system tells it plainly. How many close-ups a scene needs, how long a shot should run, these are craft, not facts, and we don't hard-code them. Models change every few months, and a rule that makes sense today will likely be wrong next year. Craft lives in Skills, where it can be edited, swapped, or written by you.

Then there is money. An Agent can think for you and write for you, but the decision to spend stays with a person. Whether it is our Agent or one connected from outside, you see the price before every generation, and when an outside app connects for the first time, the "spend credits" permission is off until you turn it on.

In the end, we are building a place where people and Agents can make a show together. People judge. Agents remember what has to be remembered and finish what has to be finished.

Just bring a good story

A whole AI film crew, ready to roll. The methods of the writers' room, the storyboard, the art department and sound are already here. All you bring is your story.

Start using ClipTap for free