Picadabra

Pica Adapter

This skill has been adapted for Pica. Treat any upstream-specific command snippets as workflow examples, not the current command contract.

Use pica status before paid work to confirm auth, credits, active environment, and current project.
Use pica --schema, pica --schema=.generate, pica --schema=.model, and pica --schema=.assets for exact command shapes.
Never guess model IDs. Run pica model search <query> and pica model info <model-id> before constructing generation input.
Route media generation through pica generate --input @file.json5; route uploads, downloads, and public URLs through pica assets.
Keep templates, prompt cookbooks, reference assets, scripts, and examples inside the skill folder as skill-owned assets.
Do not treat visual review heuristics as hard infrastructure gates. Use them as low-confidence checks unless the issue is mechanically measurable, such as missing files, wrong MIME type, bad duration, missing audio, or invalid dimensions.

Trigger

FIRES on a swap brief: "same video but with a different girl/guy", "recreate this clip with ", "give me 4 versions of this clip with different people", "swap the creator in this UGC video".

DO NOT FIRE when:

The user wants a brand-new ad authored from a product (no source clip to mirror) → /ugc-ad.
The user wants a different niche of video → match that niche's skill.

What this skill is

The craft specialist for the swap kind of remix. The user supplies a source clip and a replacement person/character; this skill mirrors the source faithfully and changes only the on-camera subject. It runs through pica generate / the chosen render path like everything else.

Hard invariants

Provider invariants stand. Source analysis, generation, and render all go through pica (pica ref analyze-video <slug> for the source breakdown; pica generate for the swap; the chosen render path). No Higgsfield MCP, no raw API. Read pica model search / pica model info before naming a model id.
Model choice is the trap — get it right (per MEMORY.md):
- Photoreal human replacement → kwaivgi/kling-v3.0-pro. bytedance/seedance-2.0 blocks photoreal-human i2v anchors (privacy filter), even AI-generated ones — it fails the swap. Do NOT reach for seedance for a realistic person, regardless of what a non-pica tutorial claims.
- Stylized / non-human character (cartoon, creature, mascot) → bytedance/seedance-2.0 is fine and often better for non-default physics motion. The privacy filter only bites photoreal humans.
Reference gate. If the replacement is a named real person ("make it Tom Cruise"), the reference-required gate fires — refuse without a ref or logged --no-ref-consent. A generic replacement ("a young woman with red hair") proceeds without a ref.
Don't recreate copyrighted source audio. Mirror the structure of the source; the user supplies/approves the actual VO and music.

The craft, in five rules

Single continuous scene — no numbered splits. Describe the entire video as ONE flowing paragraph. "Scene 1 / Scene 2" labels confuse i2v models and break continuity.
Face always in frame — state it repeatedly. If the action has the subject look down, reach, or place something, the camera must NOT follow. Put the lock inline in the action AND as a closing rule: "FACE IS ALWAYS CENTER FRAME. Camera locked on the face. Never tilts down. Never follows the hand."
Props — hyper-specific, with negatives. If there's one prop, forbid the variations the model invents: "only ONE standalone ice cube resting in the palm — no bucket, no bowl, no container, no pile."
Body mechanics — describe the motion, not the intent. Not "she puts it away" → "she deliberately reaches her hand down and tucks the cube into the waistband, her hand moves below frame, the object disappears from view." Specific mechanics = faithful output.
Lock the setting up top. Open the prompt with the interior/exterior + lighting ("INTERIOR kitchen, soft window light, blurred counter behind") before any action, or the model relocates the scene.

Prompt shape (single continuous flow)

[Replacement character: age, look, outfit — specific; OR "@ref the attached photo"].
INTERIOR/EXTERIOR [setting] — [lighting], [background].
Opening: [shot type], [what's in foreground]. The subject [neutral start]. FACE ALWAYS IN FRAME.
[Continuous action — body mechanics in order, one flow].
Camera stays tight on the face throughout. Handheld smartphone UGC feel.
FACE IS ALWAYS CENTER FRAME. Camera locked on the face. Never tilts down. Never follows the hand.
Audio: "[line]" / "[line]" / [sound] / "[line]". Natural room acoustics.

Workflow

Analyze the source. pica ref pull <url> then pica ref analyze-video <slug> (gemini-3.1-pro over the full mp4 — don't eyeball frames; see MEMORY.md). Extract: setting + lighting, the continuous action sequence, props (what + how used), audio/dialogue, camera feel.
Lock the replacement. Photo provided → pass via --ref (lock it as the master and pass on every variant to prevent drift — MEMORY.md super-original-refs). No photo → describe in text. Named real person → reference gate.
Build the single-continuous-scene prompt per the shape above. Pick the model by the rule (photoreal human → kling; stylized → seedance).
Generate via pica generate ... --ref <character>. 9:16, duration matched to source.
Variants: to produce multiple character versions, change ONLY the character description; keep setting/action/props/audio identical across variants. Generate as separate pica generate calls (auto-versioned).
Render + eval → the user's chosen composition/render path → /evaluator.

Pitfalls → fixes

Problem	Fix
Photoreal human swap fails / returns nothing	You used seedance — switch to `kling-v3.0-pro` (privacy filter).
Model makes a pile/bucket instead of one prop	Add explicit negatives: "only ONE … no bucket, no bowl, no pile".
Camera tilts down and follows the hand	Repeat the face-lock rule inline AND at the end.
Subject drops the prop instead of placing it	Spell out body mechanics ("deliberately reaches down and tucks … hand moves below frame").
Scene opens in the wrong location	Put "INTERIOR " at the very top before any action.
Video splits into disconnected shots	Remove any "Scene N" labels — one continuous paragraph.
Replacement identity drifts between variants	Lock the ref as a master and pass `--ref` on every call.

Ugc Model Swap