AI Shot Planning: Shot Lists, Shot Order, and Blocking Before You Generate

An AI shot list is the input document for AI video generation — but the list alone doesn't prevent continuity drift, because every AI clip generates in isolation with no memory of the previous one. The fix is a sequenced shot plan held in a director's assistant agent's persistent context: lock multi-angle 4K character sheets (Recraft for portraits, Nano Banana for sheets) before any motion, then route each shot to Seedance 2.0, Kling, or Veo. Documented productions ran $315–$750 per finished minute.

A traditional shot list vs. an AI shot plan

A traditional shot list assumes physical continuity for free: the same actor, the same set, and the same light persist between setups because they physically exist. An AI shot list gets none of that. Each generated clip is a fresh render with no knowledge of the clip before it, so two adjacent shots can return different faces, different wardrobe, and different room geometry unless the plan supplies that continuity as inputs.

That changes what the document has to contain. A traditional shot list records shot number, size, lens, and movement. An AI shot plan records all of that plus the reference assets each shot must carry — character sheets, locked world images, and the style constraints — and the sequence the shots will be generated in. In one documented production, the agent evaluated every scene request against 12 parameters per shot, including shot design, lens, lighting plan, blocking, and the final prompt with its negative prompt. The shot list is the skeleton; the locked references are what keep shot 14 looking like shot 13.

Shot list vs. director's assistant agent

The difference between a list and a plan is where the list lives. invideo is an agentic video creation tool with all the current video and image models available, and the practical move is to hold your shot breakdown inside a director's assistant agent rather than in a spreadsheet: the agent keeps the breakdown in persistent context, so every generation request is checked against the same sequence, the same characters, and the same world. Re-prompting scene-by-scene from a static document is the anti-pattern; persistent context is what prevents drift across a multi-scene production.

Use the director's assistant agent specifically to sequence shots — to make sure the system knows which shot comes after which before video execution begins. In multi-agent setups, productions initialize a creative producer agent first with the full script, shot breakdown, and character details, then let the director's assistant agent tighten the edit-order logic on top of that foundation. One documented production ran six agents simultaneously this way; another ran eight across separate project pages. The shot breakdown also keeps you oriented mid-project: asking the invideo agent for a status summary surfaces what is approved, pending, or awaiting regeneration.

The agent's held context does work a paper shot list can't. In one production, the invideo agent reconstructed a spatial reverse angle using only the geography established in prior shots — no reference image — and proactively flagged narrative requirements three scenes ahead.

Plan shot order before you generate

Shot order matters in AI production for a reason it never did on set: generation order is also context order. When the invideo agent knows shot 7 follows shot 6, it carries character, lighting, and spatial logic forward; when shots are generated out of sequence with no breakdown, each one is an island. Build the order around edit flow — what cuts into what — not around generation convenience.

Two ordering tactics from documented productions:

• Chain coverage pairs immediately. After a hero shot lands, request the compositionally opposite angle in the same session. The context is hottest right after the shot you want to match, and you walk away with a matched pair that actually cuts together.
• Let the sequence expose density problems before you spend. One production's densest scene called for 18 cuts in 15 seconds; because the full sequence was loaded, the agent flagged the model limitation and recommended splitting the scene before any credits were spent — and the split version cut sharper than the original script.

Sequencing is also what makes continuous takes possible: each segment's end frame becomes the next segment's reference, which only works if the order is fixed before generation. We cover that chaining workflow in depth in our guide to extending AI shots.

Blocking and the mock-shot reference

Blocking — where bodies and the camera sit in space — is the part of a shot plan that text prompts communicate worst, and the fix is to put the blocking into a physical reference instead of more adjectives. Act the shot out yourself, film it on your phone, and upload the footage to the invideo agent as a reference video. The model gets an unambiguous spatial anchor: where the camera is, how it moves, where the subject stands relative to frame.

This came out of a documented production stuck on a POV shot. As the team put it: "It suggested that instead of prompting our way to our goal why don't we shoot like a mock video of it on our phone inside the office." The mock footage anchored the generation and the shot landed. POV shots are a documented weak point of current video models, so plan a mock reference for any POV or complex-movement shot at the shot-list stage rather than discovering the problem mid-generation.

For reverse and coverage shots, blocking planning extends to production design: instruct the invideo agent to apply art director logic rather than simple mirroring. The agent will surface undecided elements — what the reverse wall actually looks like — and present options before generating, so the reverse angle is designed, not guessed.

Plan two-character contact shots specifically

Multi-character physical contact — bodies touching, one character carrying another, ropes, shared props — breaks AI video models faster than almost any other scenario, so these shots get their own line items in the plan. In one documented short film, a two-character carry setup appeared in 75% of the film, which meant the configuration had to be locked as a reference asset before generation, not negotiated shot by shot.

Three planning rules for contact shots:

1. Build a fused character sheet for the contact configuration itself. A sheet per character isn't enough; the model needs to see the two bodies in their physical arrangement. When text prompts can't produce an accurate fused sheet, hand-sketch the configuration and upload the drawing — in the documented case, the invideo agent took the sketch, fed it to Nano Banana, and prompted its way to a usable fused sheet.
2. Remove objects from characters' hands before generating turnaround sheets. Held props drift across angles; add them back at the shot level.
3. Include close-up panels on every sheet. Small details — scars, accessories, the prop at the contact point — only stay consistent across models if the sheet shows them at close range.

Plan the chunks: act-by-act, not whole film at once

Scope the plan in acts, not in the full film. Working act-by-act — completing shot planning, generation, and edit for one act before starting the next — prevents the context loss that creeps into long-form projects and keeps you oriented inside a large shot count. One documented 7-minute animated short split its script into three acts specifically so the agent never lost narrative context, and its largest sibling project ran past scene 169 with shot variants numbered per scene.

Within an act, plan around the generation unit: clips come back in roughly 15-second chunks, and multi-shot models mean one chunk can contain several cuts. That has two planning consequences. First, you need fewer boards than the old first-frame/last-frame workflows demanded — a single planned frame can seed a 15-second multi-shot sequence, which saves both time and credits. Second, each 15-second generation typically contains 4–7 usable shot candidates, so the shot plan should mark which beat inside the chunk is the keeper rather than treating one generation as one shot.

The production math behind planning first

The generation numbers from documented productions are the argument for planning before generating. One 3-minute animated episode kept only around 25% of the clips it generated, and 17 final shots were Frankenstein shots, stitched from two or more generations of the same prompt. As the production notes put it: "MOST SHOTS AREN'T ONE SHOT. Prompt → 8 tries → Frankenstein the keepers."

Planning is what makes those ratios a budget line instead of a leak. Locking character identity up front cost roughly $9.78 per character — cheap insurance against re-rolling finished shots later. The planning-heavy animated episode came in at about $950 total. Run the invideo agent in always-ask mode so every prompt and its attached references get your approval before credits are spent — the approval gate is where the shot plan earns its keep.

A practical pre-generation checklist

Run this before the first video generation:

• Script loaded. The full screenplay is in the invideo agent's context so the shot breakdown inherits character arcs and story logic.
• Shot list generated and sequenced. A director's assistant agent holds the breakdown and knows which shot follows which.
• Character sheets locked. 4K portraits (Recraft), then multi-angle 4K character sheets with four angles plus face and mid-angle close-ups (Nano Banana). Generate four options per asset, pick one, lock it.
• Contact configurations sheeted. Any two-character contact shot has its own fused reference; hands are empty on turnarounds.
• World elements locked. Lock a key element and let the invideo agent extract wide, close, and side angles from it automatically.
• Blocking references shot. Phone mocks recorded for every POV or complex-camera shot on the list.
• Dense scenes stress-tested. Cut counts per 15-second chunk checked against model limits; over-dense scenes split.
• Model routing decided per shot. Seedance 2.0 for reference-to-video continuity across clips, Kling for native multi-shot sequences, Veo where its look fits the scene — all available inside invideo, so the invideo agent routes each shot rather than you switching tools.
• Always-ask mode on. Every generation approved against the plan before credits are spent.

FAQ

What is an AI shot list?

An AI shot list is a sequenced breakdown of every shot in an AI-generated film, extended beyond traditional fields (shot size, lens, movement) to include the reference assets each shot must carry — character sheets, locked world images, and style constraints. Because each AI clip generates in isolation, the list functions as the continuity system, not just a schedule.

How is an AI shot plan different from a storyboard?

A storyboard shows what each frame looks like; the shot plan governs order, references, and blocking inputs. With multi-shot video models, one planned frame can seed a 15-second sequence containing several cuts, so you board fewer frames and invest more in sequencing and locked references.

How many generations should I budget per shot?

Expect to generate several takes per usable shot — documented productions kept only a minority of the clips they generated. Plan overgeneration as a deliberate budget line rather than a surprise.

Which AI video model should I route shots to?

It depends on the shot: Seedance 2.0 reference-to-video carries character and location context across consecutive clips, Kling generates multi-shot sequences natively, and Veo suits shots where its rendering style fits. All of these run inside invideo, so the invideo agent routes each shot to the right model from one shot plan.

Sources

All production figures in this article — generation counts, selection rates, per-minute costs, and character-lock economics — come from documented invideo productions: a 3-minute animated episode, a 70-second short film, a ~90-second horror short, and a multi-day team production.

Watch these to see the techniques in action: