AI Video Camera Work

The source post's lesson in one line:

Name the camera move, then the subject, then the direction and speed. “More impactful” is not a camera instruction.

This skill turns vague video ideas into a shot list and a prompt that a video model can follow.

When to use

• A prompt like “make it more cinematic / powerful / dynamic” did not work.
• The model substitutes a zoom for a dolly, ignores an orbit, or changes the shot size.
• You need a 60-second AI video with a real blueprint instead of a one-shot prompt.
• You need to describe a shot to Veo, Sora, Kling, Runway, Pika, or a local video model.
• You need consistent camera language across several scenes.

The method

1. Define the shot's job.
2. Choose ONE camera move from the 100-term index.
3. Add shot size + angle + lens/DOF + composition.
4. Add subject + action + start/end position.
5. Add direction + speed + duration.
6. Add light + continuity + negative constraints.
7. Generate, inspect, repair one variable at a time.

The source sheet's own formula:

技法名 + 被写体 + 方向・速さを指定
technique + subject + direction/speed

Prompt formula

[Shot size], [angle], [lens / depth of field], [composition].
Camera: [one movement] [direction] at [speed] over [duration].
Subject: [who/what], [action], [start position] → [end position].
Environment: [place, time, weather, era].
Light: [source, direction, contrast, palette].
Continuity: [wardrobe, props, screen direction, identity anchors].
Negative: no cuts, no text artifacts, no morphing, no extra limbs, no camera shake unless requested.

Four ready prompts

Dolly in

Medium shot, eye level, 50mm, shallow depth of field, centered composition.
Camera: slow dolly in toward the subject over 5 seconds, constant speed, no zoom.
Subject: a lone courier in a rain-soaked alley, walking toward camera, start mid-frame, end chest-up.
Light: cool practical neon from screen-left, soft rim light.
Continuity: same coat, same bag, screen direction toward camera.
Negative: no cuts, no text, no morphing, no extra people.

Orbit

Medium wide shot, chest height, 35mm, rule of thirds.
Camera: 360-degree orbit around the subject at constant radius over 8 seconds, smooth gimbal.
Subject: a warrior standing still in a ruined courtyard, centered, cloak moving in wind.
Light: low sun from screen-right, long shadows.
Continuity: same armor, same sword, horizon level.
Negative: no cuts, no zoom, no camera shake, no text.

Rack focus

Close-up, eye level, 85mm, shallow depth of field, locked-off camera.
Camera: rack focus from the foreground hand to the background doorway over 2 seconds.
Subject: a hand holding a key in the foreground; a silhouette appears in the doorway behind.
Light: warm practical lamp on the hand, cold blue background.
Negative: no camera movement, no cuts, no text, no morphing.

Whip pan

Wide shot, eye level, 24mm.
Camera: fast whip pan from screen-left to screen-right over 0.6 seconds, strong motion blur, then settle.
Subject: empty street on the left; a speeding car enters from the right.
Light: dusk, wet asphalt reflections.
Negative: no cuts, no text, no warped car shape, no extra vehicles.

The 100-term index

Full bilingual table with prompt cues:

references/100-camera-movements.md

Four groups:

60-second blueprint workflow

Full workflow and templates:

references/60-second-ai-video-workflow.md

Short version:

1. Blueprint — 10 scenes, second-by-second; 20 subtitle lines with display windows; factual rules; visual rules.
2. Approve — do not render before the blueprint is checked.
3. Voice — direction separate from body; generate the whole script at once; keep narration shorter than subtitles.
4. Picture — one camera move per shot; use the formula above.
5. Assemble — narration at scene heads; duck music; consistent screen direction.
6. QA — run the same blueprint twice; compare structure and inspect text frame by frame.
7. Package — blueprint, shot prompts, voice script, QA notes, sources.

Repair table

Guardrails

• One dominant camera move per shot. Split complex moves.
• Do not stack dolly + zoom + orbit + handheld unless the model is known to handle it.
• Use English camera terms for most models, with a plain-language restatement.
• Do not use “more powerful” as the only direction.
• Do not claim a shot succeeded without inspecting frames.
• Label source claims: the linked article's model versions, timings, and free-tier numbers are the author's claims, not independently verified.
• Do not claim the LINE-gated PDF was read; it is GATED.

Source and assets

• Post: https://x.com/Codex_jisen/status/2110505177245777955
• Article: https://x.com/i/article/2103440981387587742
• Follow-up: https://x.com/Codex_jisen/status/2110636237405188466
• PDF post (gated): https://x.com/Codex_jisen/status/2111262714144379348
• Notes: references/codex-post-notes.md
• Source links: references/source-links.md
• Four original infographics: assets/images/camera-work-1.jpg … camera-work-4.jpg

The images were copied at full resolution (2048×1638). The 100 terms and section labels were transcribed by direct visual reading of those images; no OCR engine was available. Hashes are in references/source-links.md.

Completion checklist

• shot size chosen
• angle chosen
• one camera movement chosen
• subject + action + start/end position
• direction + speed + duration
• lens/DOF + composition
• light + continuity
• negative constraints
• generated and inspected
• one variable repaired, not five
• source/caveats recorded