방방구석컴퍼니
← 이야기 목록
AI 도구2026-10-02

클로드 오퍼스 5.5로 30초 AI 액션 단편 만들기 — 단계별 지시문과 실제로 넣은 영상 프롬프트

클로드 오퍼스 5.5가 힉스필드로 30초 염력 액션 단편을 만든 과정의 단계별 지시문, 실제로 제출한 영상 프롬프트 원문, 밋밋한 초안을 액션으로 바꾼 규칙 7가지를 그대로 공개합니다.

방

방구석컴퍼니

혼자 일하는 사람 · 작업 기록

[광고] 이 글의 영상은 힉스필드(Higgsfield)로부터 대가를 받고 만든 유료 협찬 영상입니다. 내용은 직접 만들어 보고 잰 그대로예요.

클로드 오퍼스 5.5(Claude Opus 5.5)에게 좋아하던 30초 액션 영상 하나를 보여 주고 "이 스타일 좋아"라고만 했어요. 클로드는 힉스필드 MCP로 그림과 영상을 직접 뽑고, 검수하고, 효과와 소리까지 코드로 입혀서 「염력소녀 도아: 지하철 진상 참교육」을 끝까지 만들었습니다. 하루, 581크레딧이 들었어요.

여기에는 그때 클로드에게 건넨 단계별 지시문, 실제로 제출한 영상 프롬프트 원문, 그리고 밋밋했던 첫 초안을 액션 영화로 바꾼 규칙을 그대로 담았습니다. 내 장면으로 바꿔서 바로 써 보세요.

이렇게 바뀌었어요

첫 초안 — 주인공은 앉아서 손만 쥐고, 때리는 장면이 없었다

첫 초안은 주인공이 여덟 장면 중 한 장면에만 나왔고 때리는 장면이 0번이었어요. 피드백 세 마디 뒤 세 번째 초안은 열한 장면 전부 카메라가 달랐고, 맞는 순간이 세 번 화면에 잡혔습니다.

세 번째 초안 — 덕킹 뒤 뒤돌려차기

준비물

  • 클로드 코드(모델 Opus 5.5) 또는 클로드 앱 + 힉스필드 커넥터(MCP 주소 https://mcp.higgsfield.ai/mcp)
  • 힉스필드 유료 구독 — MCP로 만들 때마다 크레딧이 빠집니다
  • 좋아하는 짧은 영상 링크 하나

단계별 지시문 (그대로 복사해서 쓰기)

각 단계를 새 대화가 아니라 같은 대화에서 이어서 시킵니다. 단계가 끝날 때마다 결과를 보고 다음으로 넘어가요.

0단계 — 좋아하는 영상 뜯어보기 (크레딧 0)

이 영상을 받아서 뜯어봐줘: <영상 링크>
1) 해상도·길이 2) 컷 검출로 장면 수와 평균 장면 길이 3) 1초 간격 모아보기 이미지를 직접 보고 장면마다 무슨 일이 일어나는지 표로
4) 대사 받아쓰기 5) 주인공이 상대와 몸이 닿는 순간이 몇 번인지 세고, AI가 약한 부분을 이 영상이 어떻게 피했는지
6) 우리가 가져갈 것 / 베끼면 안 되는 것(창작자 시그니처)을 나눠서 REFERENCE.md 로 저장. 측정값마다 근거 명령을 같이 적어.

1단계 — 디자인 시트 (상한 60크레딧)

이야기(STORY.md)대로 시트를 만들어줘. 생성 전에 꼭 견적부터.
- 주인공 얼굴만 나온 클로즈업(회색 배경) / 전신 의상 / 손 클로즈업
- 상대 인물 각자 상반신 · 배경 장소(사람 없음, 21:9) · 전체 분위기 대표 장면 1장
- 실사, 글자·로고 없음, 유명인 닮음 금지, 나이·얼굴이 서로 확실히 다르게
받은 이미지는 모아보기로 만들어 직접 검수하고, 불합격은 이유를 적고 1번만 다시.

막힌 곳: 주인공을 "teenage girl"로 쓰면 생성이 거부됐어요. 이후엔 "교복 입은 학생(a high school student in a school blazer)"으로만 썼습니다.

2단계 — 컷 설계 + 첫 초안 (480p, 15초 1테이크)

SHOTS.md 에 15초 블록 프롬프트를 써줘. 절 구조(Summary / Style / Cinematography / Lighting / Physics / Acting / Continuity / Editing / Audio / Constraints),
Editing 절에는 정수 초 단위 샷 목록, 한 샷 = 주어 1명·동작 1개.
480p 초안으로 견적 내고 1테이크만. 받으면 컷 검출·받아쓰기·장면별 판정표(계획 vs 실제)까지.

3단계 — 피드백으로 고치기 (여기가 핵심)

초안을 보고 말로 짚습니다. 저는 이렇게 말했어요.

초안 봤는데 액션씬이 너무 부족한데? 직접 타격하면서 염력 쓰는 느낌이 딱 나왔으면 좋겠어.
피하고 때리고 이런 것도 나오게 잘 만들어봐.
다이나믹한 영상 되게 진행하고 효과는 코드로 후편집하자.

4단계 — 뒤 15초는 이어 늘리기

뒤 15초는 앞 초안의 마지막 장면에서 바로 이어지게 video_extension(forward)으로 만들어줘.
새로 나오는 인물은 이미지 레퍼런스로 같이 넣고, 이음새(구도·인물 위치·조명·좌석)가 맞는지 확인해.

5단계 — 효과는 코드로

맞는 순간 프레임을 직접 뽑아서 번호·좌표를 표로 정리하고(VFX_PLAN.md),
히트스톱·충격파 고리·화면 흔들림·번쩍임·스피드 램프를 Python(OpenCV)+ffmpeg 코드로 입혀줘.
좌표는 비율로 저장해서 1080p 확정본에도 그대로 다시 돌릴 수 있게. 전/후 비교 이미지를 직접 보고 판정해.

6단계 — 소리

생성 오디오는 액션음·환경음만 살리고, 음악은 따로 만들어 첫 타격에 가장 크게 터지는 지점을 맞춰줘.
타격음은 맞는 순간 멈춤 시작에, 휙 소리는 날아가는 구간에. 최종 -14 LUFS, 피크 -1 dBTP 이하.

7단계 — 고화질 확정

480p 초안을 draft_job_id 로 1080p 확정해줘. 견적은 길이·모드까지 넣어서 내고,
확정본이 초안과 같은 장면인지 맞는 순간 프레임 몇 장을 비교한 뒤 효과·소리 코드를 그대로 다시 돌려.

막힌 곳: draft_job_id 만 넣고 견적을 내면 길이가 기본 5초로 잡혀 60크레딧으로 나왔어요. 실제는 15초 180크레딧입니다.

밋밋한 초안을 액션으로 바꾼 규칙 7가지

  1. 한 장면 = 한 번 때리기. 접촉하는 사람은 두 명까지.
  2. 맞는 순간에 끊는다. 휘두르는 장면은 닿는 순간 끝나고, 다음 장면은 몸이 이미 날아가는 상태로 시작. AI가 약한 '몸이 닿는 순간'의 팔다리 엉킴을 피하는 방법이에요.
  3. 피하기는 아슬아슬하게. 주먹이 머리카락을 스칠 거리, 카메라는 피하는 방향으로 휙.
  4. 샷마다 카메라를 다르게. 낮은 앵글, 어깨 너머, 휙 돌리기, 맞는 순간 급줌, 반 바퀴 회전, 위에서 내려다보기. 빈 통로만 비추는 장면 금지.
  5. 효과는 생성에서 빼고 코드로. 생성 단계에 빛·파동을 넣으면 480p에서 안 보이고 나중에 입힐 효과와 겹쳐요.
  6. 이어지게 고정하기. 좌석 색, 의상, 쓰러진 사람이 끝까지 그 자리에 남는다는 걸 프롬프트에 박아 둡니다.
  7. 실제 무게. 맞은 몸은 접혔다가 체격으로는 불가능한 거리로 날아가고, 떨어진 자리에 남아요. 피·상처는 없음.

실제로 제출한 영상 프롬프트 (앞 15초)

레퍼런스 이미지 7장(객차·취객·할머니·주인공 얼굴·주인공 의상·트레이닝복 남자·라이브 방송 남자)과 함께 Seedance 2.5, 21:9, 15초, 480p 초안으로 제출했어요.

No BGM, no music, no score, no instrumental, no melody, no synth, no ambient pad. Only diegetic fight sounds, subway ambience and one Korean line.

Summary: In a daytime Seoul subway car, a calm Korean high school student in a navy school blazer under a grey oversized hoodie fights three rude passengers in the aisle. She dodges each attack by a hair and answers with one clean martial-arts strike that folds the attacker and throws him across the car. Dynamic live-action Korean action cinema, cool and serious, not comedy.

Style: photoreal live-action Korean action film, anamorphic widescreen look, subtle film grain, muted natural colors, real-time motion at normal speed throughout.

Cinematography: every shot uses a different camera: low-angle tracking, over-the-shoulder handheld, whip-pans that follow each dodge, crash zooms on each strike, one half orbit, one top-down overhead shot. The camera is always moving with the action; never show an empty aisle.

Physics: strikes have real weight. At the instant of contact the struck body visibly folds around the impact, then snaps backward and is thrown much farther than her size could manage. Bodies land hard and stay down where they fall. Dodges are close: the fist or the stick brushes her hair. Hand straps swing, metal rattles, the door glass cracks. Nobody bleeds and nobody is injured.

Continuity: the same subway car in every shot; all seats are light grey; everyone keeps the outfit from their reference image; each defeated man stays lying where he landed and is visible in later wide shots.

Editing: 15 seconds, 12 shots, HARD CUT between shots. Each strike shot ends at the instant of contact; the next shot starts with the body already flying.
0-2 seconds, Shot 1: low-angle wide shot moving slowly down the aisle. The drunk man leans over the grandmother in the priority seat. Drunk man's line (slurred, rough, in Korean): {비켜, 늙은이가 어딜 앉아.}
2-3 seconds, Shot 2: over-the-shoulder handheld. Doa pulls her headphones down to her neck and stands up.
3-4 seconds, Shot 3: handheld medium close-up. The drunk man reaches to grab the front of Doa's hoodie; she knocks his wrist aside with her forearm.
4-5 seconds, Shot 4: crash zoom. Doa drives an open palm strike into the drunk man's chest; his body folds around her palm; cut at the instant of contact.
5-6 seconds, Shot 5: low-angle camera tracking backward as the drunk man flies down the aisle and lands on the floor by the connecting door, where he stays lying.
6-7 seconds, Shot 6: the tracksuit man swings a big hook; Doa ducks under it, the fist brushing her hair; whip-pan with the dodge.
7-8 seconds, Shot 7: the camera makes a half orbit as Doa spins into a spinning back kick to his stomach; his body folds; cut at the instant of contact.
8-10 seconds, Shot 8: top-down overhead shot from the ceiling. The tracksuit man is thrown upward, hits the hand-strap rail, and drops limp across the grey seats.
10-11 seconds, Shot 9: the streamer swings his selfie stick down at Doa's head; she slips sideways and the stick brushes her hair; whip-pan with the dodge.
11-12 seconds, Shot 10: crash zoom. Doa snaps a front kick into the streamer's chest; his body folds; cut at the instant of contact.
12-14 seconds, Shot 11: side tracking shot. The streamer flies sideways into the glass of a side sliding door; the glass cracks; he slides down and sits slumped against the door.
14-15 seconds, Shot 12: low-angle wide shot from the floor, the camera does not move. Doa stands in a ready stance in the middle of the aisle; the three men lie where they landed; the grandmother and the passengers stare at her.

Constraints: no blood, no injuries, no weapons other than the selfie stick. No visual effects: no glow, no light flashes, no energy waves, no sparks, no particles. Normal human eyes. No subtitles, no captions, no on-screen text, no signs, no logos.

(레퍼런스 이미지마다 역할을 적는 "Image N is …" 줄은 내 레퍼런스에 맞게 앞에 붙이세요.)

든 비용

단계크레딧
디자인 시트 12장41.25
480p 초안 4번(15초씩)180
1080p 확정 2건(앞뒤 15초)360
효과·소리(코드)0
합계581.25 (공식 환산 약 29달러)

탐색은 480p 초안으로 싸게, 확정은 마지막에 한 번만. 이게 크레딧을 아끼는 가장 큰 차이였어요.

새 글 알림

이런 작업 기록, 계속 받아보실래요?

네이버 블로그 이웃추가하면 새 글이 피드로 바로 떠요. 광고 없이 진짜 이야기만.