Stagefight

stagefight

A two-call pika pipeline: stage still (generate_image, gpt-image-2) → fight clip (generate_video, image-to-video, ~10s, 16:9). The clip is fan-filmed POV phone footage of a choreographed cosplay performance on a themed stage — two performers clash, a live-stage effect bursts, the crowd reacts. One run produces one clip for one matchup.

The core job of this skill: figure out WHO is fighting (ask the user, or suggest matchups), then AUTHOR fresh prompts in the structure of the bundled examples and generate a new video. The examples in examples/ are the pattern to follow, not a fixed catalog — every run should produce a new matchup composed the same way. The look is calibrated and the safety framing is load-bearing: keep every fixed sentence of the two templates and fill only the marked [SLOTS].

★ Why it works — the method (this is the actual research, do not drop any of it)

Raw "two people fighting" prompts get moderation-blocked. These four moves are what make fight content reliably generate, and together they ARE the trend's aesthetic:

1. Frame it as a wholesome theatrical performance. Every clip prompt opens, verbatim: "Wholesome theatrical stage performance, choreographed cosplay, advertiser-safe." and closes with "Believable live-stage effects, not cartoon magic." This reframes combat as a stunt show, which clears the safety pass. Never describe injury, blood, real weapons connecting, or "cartoon magic" — describe choreography, props, and stage pyrotechnics.
2. Fan-filmed audience POV. "POV handheld phone footage from the audience, raised above other people's heads filming the stage," with "a large audience in the foreground — dark silhouetted heads, many holding up phones recording." This is both the viral "I was there" aesthetic AND part of the safety (it reads as documenting a performance, not real violence).
3. Costumed ORIGINAL characters — never real people, never named IP. Use archetypes ("a lean samurai in a straw hat", "a horned demon in black robes", "a spiky-blond-haired ninja in an orange tracksuit"). Real-person faces get moderation-blocked on both backends; named/trademarked characters drift and trip flags. Describe the cosplay (hair, costume, props), not a brand.
4. Live-stage effect, grounded. The "wow" beat is a believable practical effect — a puff of stage smoke, green stage fire and sparks, a wall of water / geyser. Both performers recoil/brace dramatically; nobody gets hurt.

Moderation ops: re-roll the same call on a block (don't immediately swap providers). If a video gen runs past ~4 min before failing, the content already cleared the safety pass — the failure is infrastructural, just retry. Do NOT bake music in via `reference_audio` — that path trips moderation; add music/SFX in post. Different Seedance backends (ark / fal) have different moderation pipelines if you need to retry.

Prerequisites

pika MCP available. Tool prefix varies by host — use whatever is exposed. Tools: generate_image, generate_video, task_status; optional for finishing: generate_music / SFX + edit_concat / edit_audio_mix.

Stage 0 — Who's fighting? (settle this first) — One matchup → one clip.

1. If the user already named two fighters — in the /stagefight args or their message (e.g. "frost mage vs ember knight", "my cat vs a dragon", "two robots") — use them, skip the menu, go straight to Compose. Don't re-ask.

1. Otherwise ask with ONE `AskUserQuestion` (header: Matchup). It MUST give the user a way to supply their own characters — make that the first option — plus exactly 3 template matchups:

• "Pick your own two fighters" — I'll name them → when chosen, ask "Who are your two fighters?" and use the answer. (The auto-added Other field also lets them type a matchup straight in.)
• Ninja vs ninja — smoke
• Samurai vs horned demon — green fire
• Sorceress vs vampire (water-stage) — geyser Never auto-select a template — the whole point is the user chooses or types their own. Do NOT just print 3 suggestions as prose and proceed; present the real 4-option menu and wait.

1. Aspect: default to 16:9 silently (the trend look). Only ask about 9:16 if the user mentions vertical/TikTok/Reels.

Once the matchup is set, compose the prompts yourself (next section), confirm in one line ("Generating your staged-fight clip — frost mage vs ember knight…"), and run end-to-end — no further yes/no gates.

Compose the prompts (the authoring step, NEVER skip it)

You are composing for the USER'S two fighters. Start from THEM and DERIVE everything else. The ninja / samurai / sorceress examples are illustrations of the formula, not the output — if the user said "X vs Y", the video is about X vs Y. Never silently default to one of the example matchups, and never substitute a different fight than the one chosen. Defaulting to ninja/samurai/water when the user asked for something else is the #1 failure of this skill.

Derive each of the four variables FROM the chosen fighters:

• Stage = a themed set those two characters would believably duel on — what world are they from? — + 2–4 concrete props + a backdrop + time-of-day/light. (Wizards → a stone arena; mecha → a neon dome; cowboys → a saloon street. Not a temple unless they're temple characters.)
• Performer A (LEFT) / Performer B (RIGHT) = each as live-action cosplay: hair, costume, a prop, a stance. Archetypes only — never real people; translate any named/trademarked character into its describable costume (no brand names in the prompt).
• Signature effect = the one believable live-stage payoff that fits THIS fight (smoke / colored fire + sparks / water geyser / dust + lantern-glow / ice shards / embers / confetti-cannon…), that both performers react to.

Worked derivation (arbitrary matchup, to show the move): user asks "robot boxer vs sumo wrestler." → Stage: neon-lit night arena, holographic banners, ring ropes, sponsor lanterns, city skyline behind. → LEFT: a tall chrome humanoid robot in red boxing gloves, joints glowing, guard up. → RIGHT: a massive sumo wrestler in a ceremonial mawashi, stamping low. → Effect: a shockwave of dust and sparks as glove meets palm. → then drop those into the two templates. (Notice: nothing ninja/temple/water — every variable came from the actual fighters. Do this for whatever the user picks.)

Step 1 — Stage still (generate_image, gpt-image-2)

Bakes the stage, both performers, and the filming crowd into frame 0 so Step 2 only has to animate the clash. Template — keep every fixed sentence, fill the `[SLOTS]`:

A recording from the audience at an outdoor theater performance at [TIME-OF-DAY/SETTING]. [STAGE: a wooden stage recreating <themed set> — <2-4 concrete set pieces, lanterns/gate/fog/water/etc>, <backdrop>]. On stage, two performers in high-quality live-action cosplay face off several meters apart: on the LEFT, [PERFORMER A — cosplay: hair, costume, prop, stance]; on the RIGHT, [PERFORMER B — cosplay: hair, costume, prop, stance]. A large audience fills the foreground — dark silhouetted heads, many holding up phones recording with glowing screens. The camera films from far back in the crowd, slightly off-center, performers small on the wide stage, slight handheld tilt. Realistic phone-camera look, faint digital noise, [natural light note]. High resolution, no text, no UI.