Turning a 20-Second Idea Into a Five-Shot Edit

Most creative ideas don’t arrive as scripts. They arrive as fragments. A mood. A single image. A moment you want to capture but can’t quite explain.

The 20-second piece I’m documenting here started that way. A late-afternoon street corner. A character waiting. A light shifting. Nothing else. I couldn’t tell you what it “meant.” I could tell you what it felt like.

Turning that feeling into a finished five-shot edit took three evenings, about $18 in tool spend, and more planning than I expected. This is the complete workflow — from the raw idea to the exported file — including the shot plan, the model routing, the generation log, and the failures.

Step 1: The Idea, Written as a Beat Sheet

The first mistake most creators make with a short idea is going straight to prompting. The model has nothing to work with because you have nothing to work with. A 20-second idea is not a shot list. It’s a feeling. Your job in pre-production is to translate the feeling into beats.

Here’s what my raw idea looked like when I wrote it down:

A woman waits at a street corner at dusk. She’s looking at something off-screen. The light changes. She decides to leave.

That’s it. One sentence. Not enough to generate anything.

The fix is a beat sheet — a breakdown of the emotional beats that need to happen in sequence. For a 20-second piece, five beats is the right structure. A documented five-scene template for AI video — hook, problem, mechanism, proof, CTA — uses the same logic: five beats, one character, one setting, and no new facts introduced mid-sequence.

My beat sheet:

Five beats. Four seconds each. One character. One location. One emotional arc: waiting, tension, decision, action, absence.

The beat sheet is the document the model can actually shoot. A prompt describes a world. A beat sheet describes a sequence. Once you have the sequence, each shot becomes a solvable generation problem.

Step 2: The Shot List — Writing What Each Shot Does

The beat sheet becomes a shot list by translating each beat into a concrete shot with technical parameters and an intent note. The intent note is the column that matters most — if you can’t write what the shot does in one line, the shot isn’t ready to generate.

Five shots. Five intents. Three camera behaviors: static (×2), push-in (×1), tracking (×1), and one handheld drift.

This is the document that makes generation mechanical. Each shot now has a job. The model doesn’t decide what the shot is for. I do.

A note on the five-shot structure: it’s short enough to write in a few minutes, but structured enough to keep outputs coherent. Five is the minimum number of shots that can carry a beginning-middle-end in visual form. Fewer and the piece feels like a clip. More and each shot can’t hold the weight of a beat.

Step 3: Locking the Inputs Before Generating a Single Second

This is the step that separates a finished edit from a pile of clips.

Before I generated any video, I locked three things: the character sheet, the location plate, and the style block. Every subsequent generation inherited all three.

Character sheet. I generated a multi-angle turnaround — front, side, three-quarter, plus a close-up — using an image model, then picked the strongest of four options. The character: woman, late 20s, dark hair pulled back, olive coat, no jewelry. Deliberately ordinary — no extreme features that might confuse a model. The sheet cost about $3 in image generations.

Location plate. A single reference image of the street corner: a residential intersection, low buildings, warm dusk light, a lamppost. Generated once, locked as the location anchor.

Style block. A short text description I pasted into every prompt:

Style: warm dusk light, soft contrast, 35mm feel, muted greens and ambers, gentle grain, shallow depth of field on close-ups, deep focus on wides.

The style block is the single most important piece of prompt text in the project. It tells the model what the film looks like, so the shot-specific prompts can focus on what the shot does.

Step 4: Routing Shots to Models

Not every shot goes to the same model. I routed each shot based on what it needed.

Shot 1 (wide, static, establish) → Kling 3.0. Static shots with subtle environmental life — light shifting, leaves moving — are Kling’s strength. The location plate was attached as the reference image. First generation was usable.

Shot 2 (close-up, handheld drift) → Kling 3.0. Handheld micro-jitter is where Kling wins. The character sheet was attached as the reference image, with the close-up panel emphasized. First generation was usable.

Shot 3 (medium, slow push-in) → Runway Gen-4.5. Camera movement was the point of this shot. Runway’s camera controls let me specify the push distance and speed precisely. The character sheet was attached as the reference image. Two generations — the first push-in was too fast, the second landed.

Shot 4 (medium, tracking) → Runway Gen-4.5. Tracking shots where the camera follows a subject require consistent subject lock. Runway rarely loses track of a subject mid-clip. Three generations — the first had the character walking too fast, the second had a background element drift, the third was usable.

Shot 5 (wide, static, absence) → Kling 3.0. Same static strength as Shot 1. The location plate was attached, with the character removed from the prompt. The shot needed to feel like the same corner, now empty. First generation was usable — but I generated two more to have options for the edit.

The routing logic is simple: Kling for static and handheld. Runway for camera-driven shots. For a five-shot piece, that’s all you need.

Step 5: The Generation Log

Here’s what actually happened across all five shots:

Fourteen generations for six usable clips — a 43% selection rate. That’s above the documented 25% average across AI production, and the reason is simple: the shot plan was written before generation, and the character/location references were locked before a single video credit was spent.

The tracking shot was the expensive one. Four generations. The problem wasn’t the model — it was my prompt. “Tracking shot following her as she walks” was too vague. The model didn’t know how fast she should walk or how far the camera should stay from her. When I rewrote the prompt to specify “medium tracking shot, camera at a constant 6-foot distance, subject walks at a natural pace,” the fourth generation landed.

Step 6: The Three-Shot Stitching Method

A 20-second piece with five shots doesn’t need complex continuity work. But two shots did need stitching.

Shot 3 (push-in) was assembled from two generations. The first generation gave me a strong push with a slight light drift at the end. The second gave me a strong push with a slight focus shift at the beginning. I took the first 2 seconds of the first generation and the last 2 seconds of the second, and cut them together with a 4-frame crossfade.

Shot 4 (tracking) was assembled from two generations. The first generation had the strongest opening beat — the character starting to walk. The third generation had the strongest closing beat — her reaching the edge of frame. I stitched the first 2 seconds of generation one to the last 2 seconds of generation three. The cut point was hidden under a slight camera move.

This is the “Frankenstein shot” approach that accounts for more than 40% of finished shots in documented AI productions. It’s not a failure. It’s the working reality of editorial yield — you generate clips, and you use the strongest seconds from each.

Step 7: The Edit

The six usable clips went into DaVinci Resolve. The edit took about 90 minutes.

Assembly. Laid the five shots on the timeline in order. The first assembly ran 24 seconds — four seconds too long. The tracking shot was the culprit; the model’s motion drifted after the 3-second mark.

Trimming. Cut the head and tail of each shot to the strongest beats. Shot 4 went from 5 seconds to 3.5. Shot 1 went from 5 to 4. The total landed at 20 seconds exactly.

Color matching. Kling and Runway grade differently. Kling’s default is warmer. Runway’s shadows are cooler. I used Resolve’s color match tool to pull all six clips toward a single reference frame from Shot 1, then made manual adjustments. The goal: a continuous world, not perfect consistency.

Sound. A low ambient bed (distant traffic, a bird, wind) under all five shots. A single piano note that rises as the push-in lands on Shot 3. A slight room tone shift when the corner goes empty in Shot 5. Music from Artlist, sound design in Resolve’s Fairlight.

Export. 1080p, 24fps, H.264. Delivered as a master file and a compressed web version.

The Cost

Under $16 in tool spend for a finished 20-second piece. The documented range for AI video production is $315–$750 per finished minute. My 20-second piece, normalized, is about $47 per finished minute — far below that band because it was a solo project with no team costs, a tight shot plan, and a 43% selection rate that beat the 25% average.

What I’d Do Differently

Write the tracking shot prompt more specifically from the start. The four generations on Shot 4 were avoidable. “Tracking shot following her as she walks” is not a shot description — it’s a mood. “Medium tracking shot, camera at a constant 6-foot distance, subject walks at a natural pace” is a shot description. Specificity makes generation mechanical. Vague prompts make it experimental.

Generate the empty corner shot first. Shot 5 (empty corner) is the simplest shot in the piece and the one I generated last. If I’d generated it first, I would have realized the location plate needed a slightly wider angle to accommodate the empty-frame composition. I ended up using a slightly different location plate for Shot 5 than for Shot 1, which meant the “same corner” continuity was weaker than it could have been.

Use one model for the whole piece if possible. Routing between Kling and Runway gave me the best shot execution, but it doubled the color-matching work. For a five-shot piece, the color-matching time was about 20 minutes. Not a dealbreaker, but a cost.

Plan for a six-shot piece. Five shots worked, but the tracking shot (Shot 4) carried too much narrative weight. It had to show the decision, the action, and the departure in four seconds. One more shot — a closer beat on her face as she turns — would have given the edit more room to breathe.

The Takeaway

A 20-second idea is not a shot list. It’s a feeling. The work of pre-production is translating the feeling into beats, translating the beats into shots, and translating the shots into prompts that the model can execute.

The generation is the easy part. The planning is the work. A documented production that generated 164 clips and used 41 in the final cut — a 25% selection rate — spent the most upfront time loading the script and character context before generating a single frame. My 43% selection rate came from the same discipline: lock the references, write the shot list, route the shots, and generate only when the plan is clear.

A good shot is not just generated. It is chosen. And a 20-second piece is not just five clips. It’s a sequence — five beats that carry an arc from waiting to absence.

Frequently Asked Questions

How many shots should a 20-second AI video have?

Five shots is the right structure for a 20-second piece. Each shot carries one emotional beat — hook, tension, shift, release, close — and each beat has about four seconds to land. Fewer than five shots and the piece feels like a clip, not a sequence. More than five and each shot can’t hold the weight of a beat. The five-scene template is a documented structure for keeping AI video sequences coherent across shots.

How do I turn a vague idea into a shot list?

Write a beat sheet first. Break the idea into the emotional beats that need to happen in sequence. For a 20-second piece, that’s five beats. Then translate each beat into a shot with four columns: shot size, camera behavior, duration, and intent. The intent column is the one that matters most — if you can’t write what the shot does in one line, the shot isn’t ready to generate.

How many generations does it take to get five usable AI video shots?

Expect 3–4 generations per usable shot, or about 14–20 total generations for a five-shot piece. The documented average across AI production is roughly 3 generations per usable shot and a 25% selection rate. A tight shot plan and locked character references can push your selection rate higher — my five-shot piece hit 43% because the shot list was written before generation.

Which AI video tools should I use for a short five-shot sequence?

Route by shot type. Kling 3.0 is strongest for static shots with environmental life and handheld close-ups. Runway Gen-4.5 is strongest for camera-driven shots — push-ins, tracking, and any shot where the camera movement is the point. For a five-shot piece, those two models cover everything. Luma Ray3 is the specialist for crane moves and orbit shots if your sequence needs spatial reconstruction.