Cinematic shots
Take a screenplay or scene to shots an image or video model can make, checking each stage before the next one spends money. Blank forms for every file named here are in templates.md.
The filmmaker conducts; the agent plays. The agent generates, logs and prices options. Every choice of what to keep, reject or change belongs to the user. Where the script leaves something open, the agent leaves a blank or offers options. It never fills the gap with its own taste.
Rules that hold everything else
- Cost order. Words cost almost nothing, then draft stills, hero stills, motion, then sound and the cut. Fix each problem at the cheapest stage that shows it. A bad still makes a bad clip, so never pay for motion to find out what something looks like.
- Ask before spending. Before any paid batch, state the model, count, settings and estimated cost, then wait for a clear yes. "Nice" or "maybe" is not a yes, and cheap is no exception. Stop at the session cap.
- The +1 is intent, not taste. When output departs from what the user said they want, flag it once; the user decides, and the decision is logged. A departure from history or from the agent's taste is the user's call, not a failure.
- Never delete. Weak stills and clips move to
superseded/. Files get version numbers and are never overwritten. - Log everything. Every job in the cost log, every decision in the production log.
Set up the project
One folder per film: shot-table.md, look-rules.md, register.md, production-log.md, cost-log.md, and folders references/, drafts/, heroes/, clips/, superseded/. The user sets a spending cap at the start of each session; the agent never sets it. API keys stay in the environment, never in chat, prompts or project files.
1. Notes on the script, then the shot table
Notes first. A screenplay is written for a crew that fills in what the page leaves out. A model fills those gaps with its own defaults. Before converting, list once each: format problems; continuity inside the script (names, counts, props that change); likely departures from the stated period; rights problems such as song lyrics or quoted text. Ask the writer to state their intent before the notes, because coverage-style notes otherwise pull toward the familiar. The writer answers each note in their own words, and each answer is logged. Voice and subtext are the writer's; a declined note stays declined.
Then the conversion. Screenplay to scene list, then to a shot table, which is the generation plan:
- Scene headings give the place and the time of day.
- Action lines become pictures, one shot per picture.
- Parentheticals become performance notes.
- Transitions become cut notes.
Each row holds one subject action and one camera move; a sentence that needs "and then" is two shots. Each row names its references, first frame and end state, so the next shot starts where this one stops. For a dialogue scene, add the floor plan and the side of the line the camera keeps, and carry both into every prompt of that scene. Cells come only from the script or the user's answers; a row with a blank that generation needs does not go to generation. The user corrects rows in their own words. No images yet.
The register is the breakdown: every character, place, prop and state the table uses, each with one fixed description. Change a description in the register and nowhere else; a look written in two places will contradict itself.
2. Look rules, written once
A model sees one prompt at a time and remembers nothing between shots. Write the look once, in the user's words, and paste it word for word into every prompt, even within a batch; one source reports that a look stated once at the top drifts by the later shots.
- Light: source, direction, quality, time of day, colour. Hold it fixed across a sequence.
- Palette and grade: one per sequence. A different grade per shot breaks the sequence.
- Single-frame camera properties: lens feel, depth of field, grain. Not a sequence of moves: a block in every prompt that says "wide, then close" asks every shot to do both.
- The hard constraint: period, place and what never appears in frame (for example: black sails, no modern objects). State it in every prompt. A frame that breaks it is regenerated; do not hope the video model fixes it.
- Shape: where each sequence rises and rests, as a note the user writes.
Write visible properties, not praise words such as "cinematic" or "stunning": ask of each adjective whether a picture can be formed from it. Sources disagree on whether naming a film, an artist or a camera brand helps; write the concrete properties regardless, so the rules survive a model change.
Write what should be in frame ("an empty meadow", not "no buildings") and gather the few bans in one closing line. Two sources report that naming a thing, even to forbid it, can bring it in; test that on one shot. Some models take bans in a separate field.
Aspect ratio, resolution and clip length are settings on the job, not words in the prompt.
When a user note applies beyond one shot, offer to add it to the look rules, and add it on a yes.
3. Characters, places and props
A model rebuilds every character, set and prop in every shot. Continuity is carried by references and fixed wording.
- One description, word for word. Paste the register description wherever the subject appears. Use the same noun every time ("the old king", never "the monarch" in shot 4) and never a bare pronoun. Change the scene around a subject, not the subject.
- Reference sheets. From approved drafts, make each character's sheet from four or five angles plus a close-up; on one film, three video models each held a face given five such stills. Ask for distinct faces, ages and builds; identical figures read as computer graphics. Distinguishing marks (a scar, a worn collar) can help hold an identity.
- Places. A reference of the place, empty of main characters, anchors the backdrop for every shot set there.
- Hero props. One reference per shot. For an asymmetric object, compare both ends with the reference before anything else: on one film, two of four image models drew at both ends a feature the reference had at one.
- Keep each reference set pure. One identity and one look per set. No second hero in a character's set, no main character in a place's set. Separate states (clean and bloodied, day and night) get separate sets.
- Name each reference's role in the prompt ("the face from reference 1, the room from reference 2"), keep their order fixed, and do not pass a reference the prompt does not name. Check that a reference exists before using it; never substitute a near match.
- Choose the model for a sequence before starting it. Switching partway can change faces and look.
4. Draft stills, then hero stills
Draft stills are fast, cheap images for exploring. Make the whole scene as drafts and show them on contact sheets. The first approved draft sets the look; use it as a style reference for the rest. To choose an image model, try several on one shot with the same prompt, references and ratio, and keep the one that looks like a film rather than a 3D render. An image prompt describes one still moment, so take any timeline out of it.
When a picture is wrong, the user revises the description, not the picture. Regenerate from the original prompt and references rather than editing a derivative again and again, which compounds drift.
Hero stills are paid, high-quality stills, made only for frames the user has approved as a clip's first frame. The clip can only be as clear as the still it starts from.
No text, labels, arrows, logos or speech bubbles in any still. Overlays are added in post (film-edit, motion-graphics).
5. Check every still before motion
Compare each still with its shot-table row, the look rules and its references. Report what is visible separately from guesses about why. Look for:
- breaks of the hard constraint, and counts of people and objects against the script;
- ghosts and duplicates (a second sail), missing parts (masts with no sails), simplified shapes, mushy impacts, both ends of asymmetric objects;
- faces that are all alike, hands, held props;
- the wrong object (a spyglass where the script says monocle), edge bars, stray text, stray background elements that will move once animated.
The agent flags; the user keeps, rejects with a reason or asks a question, and marks which keepers become hero stills. Redo rejects as variants; the rejects go to superseded/. The agent's check is never the user's approval.
6. Image-to-video: one action, one gentle move
The hero still is the first frame, so the prompt describes what changes, not what the frame already shows. A shape to start from:
[Look rules, pasted]. [Subject, in register wording]. [One action, as a verb].
[One camera move]. [Light and air]. [End state: "... then settles."]
Keep the exact style, faces, wardrobe and props of the first frame.
[One closing line of bans.]
- Animate gently. One slow camera move, rigid and plausible motion. Fast, complex motion smears: on one film, oars in a stroke smeared until the prompt said they never bend or merge and asked for a moderate pace.
- One shot per clip. Some models add cuts by default; ask for one continuous shot when the row is one shot.
- Say how it ends. A shot in which something changes state needs its end state written. A last-frame image, where the model accepts one, pins it.
- Short over long. One source traced objects appearing mid-clip to long single shots. Split a row that needs several actions.
- Timing inside a clip ("at 3 seconds she turns") is reported to be followed loosely; treat it as guidance.
- Resolution. Test motion at the lowest resolution that shows the problem. Raise it only on a yes; "final" or "cinematic" is not a yes.
- Audio. If the video model makes sound, say what it should be or ask for none. Everything heard belongs to film-sound.
Camera moves, plainly: static, slow push in, slow pull back, pan, tilt, tracking, crane, handheld drift. Each needs a reason in the shot.
7. Check several frames per clip
One frame says nothing about motion. Make a four-frame contact sheet (first, two middle, last) for every clip:
N=$(ffprobe -v error -select_streams v:0 -count_frames -show_entries stream=nb_read_frames -of csv=p=0 clip.mp4)
S=$(( (N - 1) / 3 ))
ffmpeg -i clip.mp4 -vf "select='not(mod(n\,$S))',scale=480:-1,tile=4x1" -frames:v 1 sheet.png
Check the face in all four frames, motion faults (bending, merging, smearing), cuts inside the clip, objects that appear or vanish, the hard constraint and the end state. Look at more frames when something moves fast.
When a clip fails:
- Name the one most likely cause: the still, the references, the action, the camera, the length, the model, or a contradiction in the prompt.
- If something extra is in the clip, look at its first frame. If the still has it, fix the still; no wording argues it out. If the still is clean and the thing grows mid-clip, simplify the action or shorten the shot.
- Change one thing and generate once. Never re-run an unchanged prompt; some models give near-identical results for the same prompt. Rewrite only the failing part and copy every line that worked back verbatim.
- After two failed tries, stop and offer redesigns: fewer actions, a split shot, another angle, the state built into the still, or another model. The user chooses.
- Test the fix on one shot before a batch. When clips chain, fix the earliest bad shot first.
The user keeps or rejects each clip with a reason; rejects go to superseded/.
Money: the cost order, the cap, the cost log
Price the shot table on three lines: draft stills, hero stills, motion. Run a test budget first: take one shot through drafts, a hero still and one clip, then price the rest from what it cost. For each shot the user decides whether to generate it, shoot it live or cut it.
Before each paid batch, show the model, the count, the settings, the unit price (from the service's current prices or quote, not memory), the estimated total and what is left under the cap. Run one or two samples before a larger batch. Wait for a clear yes.
- The cap is a ceiling, not a target. At the cap, stop and ask.
- Never re-submit a job already sent. A timed-out job may still be running; check it by job ID. An empty result is a failure: report it, never retry quietly. Report partial batches as they are ("6 of 8 ready").
- Read the model's input fields before calling it; never guess parameters. Download outputs promptly; hosted links can expire.
The cost log records every job, failed ones included: date, shot, stage, model, settings and seed, quoted and charged cost, job ID, output file, status. A logged job can be regenerated or reused as a reference.
The production log
The production log records who decided what: what was shown, the decision (keep, reject, question, hero), the reason in the user's words, each flag the agent raised with the user's answer, and where superseded files went. Write a decision only when the user makes it. A finished generation is not an approval, and neither is silence.
Rights and consent
- A real person's likeness needs that person's consent, including before their photo is uploaded as a reference. No public figure's likeness without rights.
- Ask before uploading anything to a service, and keep generated media private by default.
- No trademarks, logos or copyrighted characters in prompts unless the user holds the rights.
- Check the service's terms for commercial use before paying for hero stills and motion; at least one service's free tier does not allow it.
- Disclosure and credits are part of release (film-edit).
Not covered yet: screenplay file formats such as Fountain, and production scheduling.
Tools as of 2026-11-17: On the films behind these rules, generation ran through Higgsfield over MCP. Image models: Nano Banana 2.1 for crowds, interiors and distinct faces; Cinema Studio 2.5 for ships, hulls and macros; GPT Image 2.6 for cheap tests. Video: Kling 4.0 for image-to-video (1080p, 3 to 30 s, with an ambient audio track). In the five-reference face test, Kling 4.0, Seedance 2.0 and Seedance 2.5 all held the face. Logged costs: a still about 2 to 4 credits; an 8-second clip 26 credits on Kling 4.0 at 1080p, 96 on Seedance 2.0, 128 on Seedance 2.5; a 37-frame draft storyboard about 75 credits; the whole first film about 1,000 credits (dollar price per credit not recorded). Batches were capped at 12 items; earlier job IDs could be reused as references; upload links expired after 24 hours.