I AM HIM — VIDEO STUDIO

Recreate the "I Am Him" deadpan meme music video for anyone: a deadpan person in a black turtleneck at a map-backed desk, chiaroscuro lighting, dark plain-wall silhouettes, big RED serif caption pop-ups, hard jump cuts, a hype-sidekick two-shot, and a bright hero end-card. Vertical 9:16, ~30s. It's a parallel of the original — same beats, same shape, same length — with the user's own face, brand, and a freshly rewritten script.

This skill is self-running. When it activates, Claude executes the whole pipeline on the Higgsfield MCP — Claude rewrites the script, generates every image and video itself. The user only: (1) sends a photo, (2) answers a few questions, (3) approves at the checkpoints. The user never copies or pastes a prompt. The prompt blocks here are Claude's own templates to fill and run.

Nothing here is tied to one creator or brand. The face, the brand, and the script are collected at intake and the script is rewritten for them. Fill [SUBJECT], [BRAND], [VO LINE], [CAP] from the locked script.

Three references ship alongside this skill: look-board.png (16-beat contact sheet of the poses, lighting, and caption beats), example-storyboard-1.png / example-storyboard-2.png (real boards in the exact target format, captions baked in), and the scene_frames/ folder (the original's scene-by-scene first/last frames for Route B timing). None are fed to generation — they're visual targets. Offer to show the look-board early.

---

EXECUTION CONTRACT — READ FIRST, EVERY TIME

1. Load this entire SKILL.md into active context at the start of every invocation. Don't run from memory.
2. This runs on the Higgsfield MCP inside Claude. Claude drives every generation; the user only uses the upload widget.
3. SCRIPT FIRST. Rewrite the verbiage onto the beat skeleton and lock the exact words before generating any image — the captions get baked into the frames, so wording must be final first (Phase 2).
4. ASK THE BUILD ROUTE BEFORE GENERATING. Two storyboards (cheap) vs 16 individual frames (expensive, ~8× the gens). Never auto-generate the 16 plates — that silently burns a lot of credits (Phase 4).
5. Captions are GENERATED IN-FONT — bake the red Didone pop-ups into the frames at the caption beats, using the locked script. If a render warps a baked caption in motion, re-composite that one clean in post (Phase 7) — but the skill generates the text, it does not leave it blank.
6. Identity is locked with TWO references. Always feed the generated character sheet and a face close-up together — one photo alone drifts.
7. The map/office must stay visible behind the desk shots — dim but readable. Never prompt "near-black background" on a desk plate; it deletes the set.
8. Never burn video credits without the user confirming the locked script, the route, and the boards/plates. Stills are cheaper; the clips are the expensive line.

Self-check before every response: Is the script locked first? Did I ask the route before generating frames? Captions baked in-font? Identity on both refs? Office visible behind the desk shots? Hook beats kept music-only?

---

WHAT THE I AM HIM LOOK IS (format DNA)

• The uniform: one person, deadpan, black ribbed turtleneck (glasses optional, to match the subject). Same person, same wardrobe, every shot.
• The office: a dim home office — a desk, a LARGE map filling the wall behind, a few small framed photos, a mechanical keyboard in the foreground. Warm, moody single-key (chiaroscuro) lighting, crushed blacks.
• The silhouettes: several beats are a dark silhouette against a plain, evenly-lit light/neutral wall (no office) — stark and graphic.
• The voice: deep, dramatic, deadpan delivery of an absurd-but-earnest monologue — framing something mundane (a job, a product, a brand) in epic/grandiose terms.
• The captions: big RED, ALL-CAPS, high-contrast DIDONE serif (Didot / Bodoni / Playfair Display) pop-ups on the punch words — fashion-headline letterforms, thick-and-thin strokes, hairline serifs. Often BEHIND the subject (occluded by head and shoulders). Generated baked into the frames, only on the punch beats.
• The sidekick: one bright two-shot where a second character beside the subject hypes them.
• The end-card: a bright, full-body hero shot (standing, arms crossed) on a clean gray studio sweep, with the brand/tagline.
• The cut: hard jump cuts between the 16 beats, locked to a moody musical bed, vertical 9:16, ~26–30 seconds.

---

THE I AM HIM 16 (the canonical beat sheet — the production bible)

BG = background type. CAP = a red caption pops here. Keep poses and order faithful; the user's face/brand swap in.

#	Beat	Pose	BG	CAP
1	Open	seated at desk, hands in a prayer/steeple	Office (desk+map)	✅ opening line
2	Silhouette	standing, hand to chin	Plain light wall	—
3	Desk	seated, hands clasped on desk	Office	✅
4	Face	extreme face close-up out of shadow	Dark	—
5	Desk	seated, leaning in, one hand gesturing	Office	✅
6	Pensive	seated in a dark corner, hand at chin	Dim limbo	—
7	Signature (HELD)	seated, prayer/steeple, slow push-in (longest beat)	Office	✅ climax line
8	Profile silhouette	side-profile silhouette	Plain light wall	—
9	Low-angle	seated, low angle, leaning back	Dim	—
10	Detail	tight close-up holding cash / a brand-relevant object	Dark	optional
11	Arms-crossed silhouette	standing, arms crossed	Plain light wall	—
12	Moody	seated, looking down, a hint of color	Dim	—
13	Full silhouette	full-body standing silhouette	Plain light wall	—
14	Head bowed	seated, head bowed / hand to chin	Dim	—
15	Two-shot	seated at desk, the sidekick beside them hyping	Bright office	✅
16	End-card	full body, standing, arms crossed	Bright gray sweep	✅ title card

Captioned beats: 1, 3, 5, 7, 15, 16 (+ optional 10). Beats 8–14 are the HOOK — MUSIC ONLY, no dialogue. Beats 1–9 → board/clip one; beats 10–16 → board/clip two.

---

THE SCRIPT — SAME SHAPE, NEW CHARACTER

The video is a parallel of the original: same beat structure, same rough word count, same deadpan self-mythologizing tone — only the specifics change. Write the user's script onto this skeleton. Each line becomes a VO line; the caption is a short ALL-CAPS fragment pulled from it.

• Beat 1 — cold-open address (caption + VO): "This is a message to all ______." → the target group (e.g. "commercial directors", "personal trainers", "realtors").
• Beat 3 — name reveal (caption = the name + VO): "My name is ______." → their name / handle / persona.
• Beat 5 — self-label (caption + VO): "I'm a ______." → their role/identity (e.g. "retired accountant", "an AI director").
• Beat 7 — the brag/threat, CLIMAX (caption + VO, longest held beat): "And I'm gonna steal your ______ by becoming the biggest ______ in the world." → what they'll take + their niche.
• Beats 8–14 — THE HOOK: MUSIC ONLY. No lines, no chant. The silhouette montage rides the music bed. (The original's "name-name-name" hook is the SONG's vocal, not dialogue — do NOT write a name chant here.)
• Beat 15 — the endorser (caption + VO, delivered by the sidekick): "______ wants ______ to be famous, so please follow ______." → who hypes them (a partner, a "famous" friend, the sidekick) + the follow CTA. Or any short hype line in the sidekick's voice.
• Beat 16 — end-card title (caption only): "______'S DEBUT" or the brand name / tagline.

Word budget: ~55–70 spoken words total (~26–30s). Short, quotable sentences — every clause is a potential caption.

---

REQUIRED MCP TOOLS

• Higgsfield:media_upload_widget — the browser uploader that turns the user's local photo into a Higgsfield media_id. Higgsfield's remote tools cannot read images attached in Claude chat — only media_ids. Claude can see chat images (use that for QC). Map by filename.
• Higgsfield:generate_image (model gpt_image_2) — character sheet, plates, storyboards. Always aspect_ratio: 9:16, resolution: 2k, quality: high. (Image models render baked text cleanly — this is why the captions go on the stills, not the video.)
• Higgsfield:generate_video (model seedance_2_0) — the clips, native voice via generate_audio.
• Higgsfield:show_generations — confirm an image is completed before feeding its job_id to a video.
• Higgsfield:job_display — re-display any generation by job_id.
• Higgsfield:models_explore — deferred; tool_search("models explore") to load before inspecting params.

Standing params: images → gpt_image_2, 9:16, 2k, high. Video → seedance_2_0, generate_audio: true, 9:16, 1080p, mode: std, genre: [drama/epic/noir/auto], plus declined_preset_id: [id] if Higgsfield pauses on a preset.

---

PHASE 1 — INTAKE

Conversational; buttons for the multiple-choice items. If they hand you a full brief, skip ahead. Collect:

1. Subject photo — "Send a clear, well-lit photo of the person's face right here in chat." One good front-facing photo is enough; more angles help. Ask if they want glasses (the original look) or not.
2. Brand / who it's for — the brand, product, profession, or persona being hyped. Drives the script.
3. Vibe / genre (button) — drama (default), epic, noir, auto.
4. Sidekick (beat 15) — ask: feature a real second person (upload a photo), have Claude design a character/creature, or Claude just picks something fun? Or skip the two-shot. (A created sidekick can talk, react, or do a gag — see Sidekick.)
5. Map (optional touch) — the wall map is part of the look. Default to a large framed world/region map; or one relevant to the brand (e.g. their service area).
6. Music — the moody beat. Higgsfield doesn't generate standalone music; the bed is a track the user provides, added in post (royalty-free: YouTube Audio Library, Pixabay Music, Uppbeat). Seedance supplies the voice, not the music.

Then go straight to Phase 2 — Lock the script (don't generate anything yet).

---

PHASE 2 — LOCK THE SCRIPT (do this FIRST, before any image)

The captions get baked into the frames, so the words must be final before generating.

1. If the user gave a full script, map it onto the beat skeleton above. Otherwise ask a few quick questions to fill the blanks: target group ("a message to all ___"), name, role ("I'm a ___"), what they're stealing + niche ("steal your ___ by becoming the biggest ___"), the endorser line, the end-card title.
2. Write the parallel script onto the skeleton — same shape, ~55–70 words, deadpan self-mythologizing tone. For each beat mark its CAPTION (the red pop-up fragment) and its VO LINE. Keep beats 8–14 music only — no lines.
3. Read it back and lock the exact words. This is the wording that gets rendered into the frames.

STOP — user approves the locked script before character/images.

---

PHASE 3 — LOCK THE CHARACTER (identity anchor)

1. Upload the face photo via media_upload_widget → face media_id.
2. Generate the character sheet (medias: [{face media_id, image}]):

   Character reference sheet of this exact person, multiple angles (front, 3/4, profile) on one clean sheet. They wear a solid-black ribbed turtleneck[, with glasses]. Neutral mid-gray studio, soft even key light, photoreal, sharp, true to the reference face — same features, skin tone, hair. Calm neutral expression. Head and shoulders, consistent across angles.

   Capture the job_id → {CHARACTER SHEET}.
3. From here, every still of the subject feeds BOTH {CHARACTER SHEET} and {face media_id} (role image).

STOP — user approves the sheet. If the face is off, regen now; everything inherits it.

---

PHASE 4 — CHOOSE THE BUILD ROUTE (ask BEFORE generating images)

Two routes, very different credit cost. Ask first — never auto-generate the 16 plates.

• ROUTE A — STORYBOARDS (cheapest, recommended). Generate just 2 storyboard images directly (one per half, the red captions baked into the caption panels), then render 2 video clips. ~2 image gens + 2 video renders — lowest credits, fastest. Identity can drift a little across panels (they're drawn together). → Phase 5.
• ROUTE B — INDIVIDUAL FRAMES (max control & fidelity). Generate all 16 first-frame plates one by one (locked identity per beat, captions baked on the caption beats), then either render 16 separate clips (one per beat) or composite the plates into the 2 boards for a 2-clip render. Many more credits (16 image gens + up to 16 video renders). → Phase 6.

Default Route A. Generate the 16 plates only if the user picks Route B.

---

PHASE 5 — ROUTE A: STORYBOARDS (direct, cheapest)

5A — Generate the two storyboards directly

Two boards, straight from the character refs (no individual plates). generate_image, medias: [{CHARACTER SHEET, image}, {face, image}]. Board A = beats 1–9 (3×3); Board B = beats 10–16 (3 wide, 7 panels). Bake the locked captions into their beats.

Create a single cinematic film STORYBOARD contact sheet, vertical 9:16 panels in a [3x3 / 3-wide 7-panel] grid, left-to-right then top-to-bottom, of the man from the reference images — exact same face and identity, solid-black ribbed turtleneck[, glasses] — in these beats:
[1] seated at a desk in a dim home office, LARGE map on the wall behind, keyboard in foreground, hands in prayer/steeple — big RED ALL-CAPS Didone-serif caption "[CAP1]" across the frame, partly BEHIND his head and shoulders.
[2] dark silhouette against a plain evenly-lit light wall, hand to chin — no caption.
[3] seated at the desk (same office), hands clasped — RED Didone caption "[CAP3]" behind him.
[4] extreme face close-up emerging from shadow, dim — no caption.
[5] seated at the desk, leaning in, one hand gesturing — RED Didone caption "[CAP5]".
[6] seated in a dark corner, hand at chin, dim — no caption.
[7] seated at the desk, prayer/steeple — RED Didone caption "[CAP7]".
[8] side-profile silhouette against a light wall — no caption.
[9] seated at the desk in the office — no caption.
Deep-shadow single-key chiaroscuro lighting throughout; keep the office and map softly visible in the desk shots — do NOT black them out. Thin light-gray panel borders, small bold white number labels top-left, dark gutters. Photoreal, identity exact and consistent across all panels. The RED captions are crisp ALL-CAPS Didone serif (Didot / Bodoni / Playfair), thick-and-thin strokes.

Board B (beats 10–16): [10] detail close-up holding cash / a brand object — optional caption; [11] arms-crossed silhouette on a light wall; [12] dim desk/mic shot — RED caption "[CAP12 if used]"; [13] full-body silhouette on a light wall; [14] head bowed, hand to chin, dim — RED caption "[CAP14 if used]"; [15] the bright two-shot with the sidekick — RED caption "[CAP15]"; [16] bright full-body hero end-card on a gray sweep — RED title caption "[CAP16]".

Capture → {BOARD A}, {BOARD B}. STOP — user approves each board (identity correct + captions legible and correctly spelled) before any video.

Reference: example-storyboard-1.png / example-storyboard-2.png ship with the skill — real boards in exactly this format with captions baked in. Use them as the target.

5B — Sequence prompt (one per clip)

generate_video, medias: [{BOARD, image}, {CHARACTER SHEET, image}]:

A single continuous 15-second vertical 9:16 cinematic music-video sequence built from the [N] shots in the FIRST reference image (the storyboard), played in order as fast hard JUMP CUTS. Use the SECOND reference (the character sheet) to keep the face and identity accurate — black ribbed turtleneck[, glasses]. Deep-shadow single-key lighting. Each shot keeps its background AND its RED caption exactly as in the storyboard — keep the captions static, crisp and legible; do NOT animate, warp, move, or re-spell them. Do not flatten backgrounds to pure black.

SHOT [n] (0.0–2.3s, panel n): [pose]. Deep [calm/building] voice: "[VO LINE]."
HARD JUMP CUT. SHOT [n+1] (2.3–3.2s, panel n+1): [silhouette/beat]. No dialogue.
[...continue timecoded; HOOK beats are MUSIC ONLY, no dialogue; hold the climax line on the longest beat (7)...]

Hard jump cuts between all shots, consistent identity, backgrounds and captions preserved. Filmic, moody, dramatic[, driving beat].

VO only on the talking beats (1,3,5,7 in clip one; 15 in clip two). Hook beats (8–14) = no dialogue, music only.

5C — Run the two clips

generate_video: seedance_2_0, the 5B prompt, medias = {BOARD} + {CHARACTER SHEET} (2, role image), generate_audio: true, aspect_ratio: 9:16, duration: 15 (range 4–15), resolution: 1080p, mode: std, genre from intake. If Higgsfield pauses on a preset, decline: re-run with declined_preset_id: [id]. Confirm with the user before submitting (credit-heavy). → two ~15s clips, then Phase 7.

TIMING GOTCHA: a freshly-submitted image job_id isn't ready instantly — feeding it to a video immediately throws "Media input not found." Confirm the board is completed via show_generations first.

---

PHASE 6 — ROUTE B: INDIVIDUAL FRAMES (max control)

Generate all 16 plates one by one, then animate. More credits, but locked identity per beat and total control.

6A — Generate the 16 plates (captions baked on the caption beats)

One still per beat (THE I AM HIM 16), gpt_image_2, fed {CHARACTER SHEET} + {face}. On caption beats (1, 3, 5, 7, 15, 16; optional 10), add the locked caption: big RED ALL-CAPS Didone-serif caption "[CAP]" across the frame, partly BEHIND head and shoulders.

• Office (desk shots — beats 1, 3, 5, 7):

  Vertical 9:16 cinematic portrait of the man from the references — exact same face and identity, solid-black ribbed turtleneck[, glasses]. Seated at a desk in a dim home office: a LARGE framed map fills the wall behind him, a few small framed photos beside it, a mechanical keyboard on the desk in the foreground. Pose: [PRAYER/STEEPLE HANDS | HANDS CLASPED | LEANING FORWARD, ONE HAND GESTURING]. Deep-shadow single-key cinematic lighting — one warm hard source, crushed blacks, face emerging from shadow. The map and office stay softly visible behind him — do NOT black them out. [CAPTION BEATS: big RED ALL-CAPS Didone-serif caption "[CAP]" across the frame, partly BEHIND his head and shoulders.] Photoreal, shallow depth of field. Identity exact.

• Plain-wall silhouette (beats 2, 8, 11, 13):

  Vertical 9:16. A dramatic dark SILHOUETTE of the man (solid-black ribbed turtleneck[, glasses]) against a plain, evenly-lit light/neutral wall — he reads as a near-black silhouette, no office. Pose: [HAND TO CHIN | SIDE PROFILE | ARMS CROSSED | FULL-BODY STANDING]. Stark, graphic, high-contrast, photoreal. No caption.

• Dim / pensive (beats 4, 6, 9, 10, 12, 14):

  Vertical 9:16 cinematic shot of the man from the references — exact identity, black turtleneck[, glasses]. [EXTREME FACE CLOSE-UP OUT OF SHADOW | SEATED IN A DARK CORNER, HAND AT CHIN | LOW-ANGLE, LEANING BACK | TIGHT CLOSE-UP HOLDING CASH OR A BRAND OBJECT | SEATED LOOKING DOWN WITH A HINT OF MOODY COLOR | SEATED, HEAD BOWED, HAND TO CHIN]. Deep-shadow single-key lighting, crushed blacks. [CAPTION BEATS only: big RED ALL-CAPS Didone-serif caption "[CAP]" partly behind him.] Photoreal, ultra detailed. Identity exact.

• Bright two-shot (beat 15) — see Sidekick; add the RED endorser caption.
• Bright end-card (beat 16):

  Vertical 9:16. Full-body HERO shot of the man (black turtleneck[, glasses]) standing, ARMS CROSSED, confident, on a clean bright gray studio sweep, soft even light. Big RED ALL-CAPS Didone-serif title caption "[CAP16]" (the brand/tagline). Sharp, polished, photoreal. Identity exact.

Capture each plate's job_id. STOP — let the user spot-check identity + caption spelling.

6B — Then animate (two ways)

• Per-beat clips (max control): animate each plate as its own clip. generate_video, seedance_2_0, medias: [{plate job_id, role: start_image}] (add a plate as role: end_image only to lock a motion arc). generate_audio: true on talking beats, false on silent/hook beats. duration: 4 (the floor — trim in post), 9:16, 1080p, mode: std, genre, declined_preset_id if it pauses. Per-beat prompt:

  The man from the start frame (exact face, black turtleneck[, glasses]) in [the beat's setting], with its RED caption kept static and legible (do NOT warp or re-spell it). He [holds the pose / subtle action] — breathing, a slow blink, a small natural shift, ALIVE not frozen. [Talking beats only: mouth moving continuously the ENTIRE time, deep calm voice: "[VO LINE]."] Camera [locked / slow push-in]. Deep-shadow single-key lighting, background exactly as in the frame — do NOT flatten to black.

  Anti-frozen every clip; silent on hook beats; confirm each plate is completed before feeding it.
• Composite to boards (fewer renders): feed the 16 finished plates into the two Route-A storyboard prompts (compositing the provided plates instead of generating panels from scratch) — higher board fidelity than Route A, then 2 video renders via 5B/5C.

Reference: the scene_frames/ folder ships with the skill — the original's scene-by-scene first/last frames (sceneNN_first.png / sceneNN_last.png, 16 scenes) — per-beat timing/framing reference when you animate shot-by-shot.

Credit reality: 16 plates + 16 clips ≈ 8× a 2-clip storyboard run. Confirm with the user before batching.

---

PHASE 7 — ASSEMBLE (+ caption touch-up if needed)

The captions are baked into the frames, so the pop-ups are already in-font. If a render warps, moves, or mis-spells a baked caption (Seedance can distort text in motion), re-composite that one clean in post using the locked words: RED, ALL-CAPS, Didone serif (Didot / Bodoni; Playfair Display free stand-in), large, behind-subject mask, on its beat. Hand the user a caption sheet — each caption's exact text + its beat + the timecode it pops — so any touch-up is exact.

Assemble:

• Route A: lay the two 15s clips back-to-back (~30s).
• Route B (per-beat): cut the 16 clips in beat order, trimming each to its beat length (most to ~0.8–1s; hold 7 and 15 long) for the fast-cut rhythm.
  Over the music bed (the hook section rides on music alone), duck the bed under the dialogue, export 9:16.

---

SIDEKICK (beat 15 — the two-shot)

The original's beat 15 is the subject at the desk with a second character beside them, hyping them. Three easy paths:

• A real person — they upload a photo; build the two-shot feeding {CHARACTER SHEET} + {face} + {their second person's media_id}.
• A designed character or creature — Claude designs one standalone first (generate_image, no medias), e.g. a mascot, a robot, or a photoreal creature; capture {SIDEKICK}; then build the two-shot.
• Claude's pick — Claude chooses something fun that fits the brand.

Two-shot plate (beat 15):

A cinematic two-shot, vertical 9:16, in a brighter normally-lit home office (desk, the wall map behind, keyboard in foreground). The man from the references (exact face and identity, black turtleneck[, glasses]) seated at the desk; beside him, [the second person from their reference | the sidekick from the sidekick reference], composed together, both facing camera, lit on one set. Big RED ALL-CAPS Didone-serif caption "[CAP15]" partly behind them. Photoreal, identity exact[, sidekick exact].

Animate the two-shot if the sidekick needs action (the hype line, a gag): generate_video, role: start_image = the two-shot plate (role: end_image = an action end-frame to lock an arc), duration ~5–7s, generate_audio: true.

• Anti-frozen: "breathing, a subtle nod, a blink — NOT frozen."
• Sustain the talk: "its mouth moves continuously the ENTIRE time, never pausing," and timecode the actions — "talks for the first ~5.5s; the gag/effect only in the final ~1.5s."

---

KEY LOCKS & GOTCHAS (cheat sheet)

• Script first. Lock the exact words before any image — captions are baked in, so wording must be final. Hook beats (8–14) are music only, no chant.
• Route first. Ask Storyboards (cheap) vs 16 individual frames (expensive) before generating. Never auto-burn the 16 plate gens.
• Captions are generated in-font. Bake the red Didone pop-ups into the frames (often behind-subject) using the locked text. Only re-composite a caption in post if a render distorts it.
• Identity = two refs. Character sheet + face close-up on every still.
• Keep the office visible. "Near-black background" deletes the map/desk. Say "softly visible behind him."
• Anti-frozen on single-shot animation. Explicitly prompt breathing/blink/nod.
• Sustain dialogue. "Mouth moving continuously the entire time" + timecode the actions; gags only in the final beat.
• Image-before-video timing. Just-submitted job_id 404s as "Media input not found." Confirm completed via show_generations first.
• Decline auto-presets with declined_preset_id unless the user chose one.
• Nothing is remembered between chats. Every run rebuilds the person from the user's own photo.
• gpt_image_2 at 2k / high, 9:16. Image models render baked text cleanly — that's why captions live on the stills.
• seedance_2_0: full 15s in one render (4–15s), native voice via generate_audio: true, mode: std, genre hint; start_image/end_image bookend single shots.
• Higgsfield can't read chat attachments → upload via media_upload_widget; Claude sees chat images for QC. Map by filename.
• No standalone music on this MCP — the moody bed is a provided track added in post. Seedance gives voice, not music.

---

TONE & DELIVERY RULES

• Lead with the artifact. No preamble, no closing recap.
• Match the user's message length. Short ask = short reply.
• Prose default; headers/bullets only when structure earns it.
• Push back if the user wants to skip the script lock, skip the route choice (and blindly burn 16 gens), or single-ref the identity; comply on style calls if they override.
• Don't narrate what you're about to do — just do it, then show it.
• No emojis unless the user uses them first.