--- title: "Museum Installation — place-as-protagonist 2-minute museum-bench piece (no characters)" reliability: 3.9 # 5 − keyframe img gen .3 − 12× kling-v3-i2v .6 − tts .2 − music .2 − 8-stage path .3 + .5 proven showcase (ATELIER · Vermeer's Light rendered E2E); no-characters = zero cast-drift risk (the usual worst failure mode is absent) --- # Museum installation — agent playbook ⭐ flagship > Paste this whole file (with the BRIEF filled in) into Claude cowork, chat, or Code. **This is the no-characters showcase that proves a single workhorse i2v cap can carry a 2-minute museum-bench piece.** The protagonist is the place. The example is [ATELIER · Vermeer's Light](/chapters/atelier-vermeers-light-example.html) — a 17th-century Delft atelier at the moment of arrival, just before the painter steps in. ## What you'll get A museum-grade installation piece, long-form, place-as-protagonist: - **12 keyframes** of one place rendered in a locked visual style — same window, same easel, same beam of light across every shot. - **12 disciplined i2v animations** — every shot has exactly ONE allowed motion (dust drifts OR lace stirs OR light brightens OR pigment-glint pulses, never two). - **Single workhorse cap** — all 12 shots are `kling-v3-i2v` 10s. No premium-cap mixing; the discipline is in the prompts, not the cap budget. - **Scholarly TTS narration** over baroque chamber music — curator-bench voice, not Hollywood narrator. - **Cinema 1920×1080 reel**, ~120s, ready for a lobby loop, exhibit wall, or museum bench monitor. - **Place bible** preserved as a re-pasteable 5-axis spec (no cast — the bible is the room, not a person). **Time**: ~60 minutes wall clock. **Your attention**: ~12 minutes across 2 audit gates. **Cost**: $7–$10 e2e. *Cheapest premium showcase you can ship* — "no characters" means zero cast-lock drift risk. ## Tell the agent about the place ```yaml piece_title: # e.g. "ATELIER · Vermeer's Light" piece_subtitle: # 1-sentence about the piece. e.g. "An empty 17th-century atelier at the moment of arrival." # 5-axis place spec (NOT cast — there are no characters) place_identity_anchor: # 1 short sentence — the agent's seed for every prompt. # e.g. "17th-century painter's studio in Delft, oak-beam ceiling, plaster walls, wide plank floor, single tall north window" place_signature_object: # the one object that anchors the room's identity across shots. # e.g. "single tall leaded-glass window with lace curtain caught against the frame" place_palette: # ≤5 colour names that define the painterly palette. # e.g. "ultramarine lapis, lemon-ochre, lead-white, vermillion, plaster cream" place_light_direction: # where the light comes from, what time of day. # e.g. "north window left, low warm morning angle, golden beam diagonal across floor" place_change_policy: # always "SAME ROOM across every shot" + describe an arc. # e.g. "SAME ROOM across every shot · light intensity grows ~10% per shot through the 12-shot arc" absence_invariant: # what MUST NEVER appear. Critical for "no characters" pieces. # e.g. "no figures, no hands, no shadows of a body — only the room and the light" # Style visual_style: # e.g. "Vermeer painterly", "Hopper light", "Hammershøi interior", "Wyeth tempera". # Pick a single painter / movement reference, not a list. # Narration + music narration_voice: # gemini-tts voice id. Recommended: "Gacrux" (warm scholarly female). narration_paragraphs: # 3-4 paragraphs of curator-voice text. The agent will TTS the whole script. music_brief: # 1-sentence brief for minimax-music. # e.g. "baroque chamber music, lute and viola da gamba, period-appropriate, instrumental" # Output total_shots: # 12 (default for 120s). 6-9 for shorter pieces, up to 18 for 180s. piece_slug: # kebab-case for filenames. e.g. atelier-vermeers-light. ``` --- ## Quality controls (auto-applied) - **Cost upfront.** 12 × krea-2-large (~$0.76) + 12 × kling-v3-i2v (~$6.30) + 1 × gemini-tts (~$0.25) + 1 × minimax-music (~$0.30) = ~$7.61. Approve before fire. - **Place spec re-pasted verbatim** in every prompt. Same room, same window, same easel — the bible is unchanged. - **Absence-of-figures invariant.** Every prompt includes the absence clause: "no figures, no hands, no shadows of a body". This is the museum-installation guarantee. - **ONE motion per shot rule.** Each i2v prompt has exactly ONE allowed motion. The agent will refuse to write a prompt with two simultaneous motions; this is the discipline that makes a single cap carry the whole piece. - **Audit gate after keyframes.** All 12 laid out in a grid: same room? same palette? same light arc (S1 to S12 ~10% brighter)? Zero accidental figures? Pass → fire animations. - **Audit gate after animations.** Did the easel drift? Did the stool move? Did a figure appear in a render? Re-fire offenders before concatenation. - **Long-form audio mix covers the full runtime.** Voice 1.0 + music ducked 0.15, muxed via `ffmpeg-mux` with `audio_fill:"loop"` so the instrumental bed runs under the whole piece (it bridges the gaps between narration paragraphs — no silent stretches). Voice fades in 0.4s / out 2.0s; music fades in 1.5s / out 2.0s; both trimmed to total runtime. > **Why one cap, not premium-mixed:** in a place-as-protagonist piece, the camera does very little — usually no fancy moves. kling-v3-i2v handles "camera holds, ONE small thing moves" beautifully at $0.05/s. Spending grok-imagine-video money on shots that are 90% stillness is wasted budget. --- ## How the agent should run this (interaction contract) 1. **CONFIRM (one message, ≤1 question):** restate in 1 line ("12-shot, 120s, 1920×1080 place piece on , kling-v3-i2v workhorse, ~$7.61, ~60 min wall clock") + real choices: 9 vs 12 vs 18 shots · the single painter/style reference · narration on/off. If the place spec is filled in, state assumptions and start; only ask when the 5-axis spec is missing. 2. **PREVIEW CHECKPOINT:** the 12-keyframe grid (~$0.76) is the first audit gate — same room? same palette? light arc S1→S12? ZERO accidental figures? — and it MUST pass before the ~$6.30 of i2v fires. The second gate (post-animation drift check) gates concat. 3. **NARRATE:** animate in batches of 3 (12 × kling 10s ≈ 30–45 min total); per batch: "shots 4–6 animating, job …, ~5 min"; poll `get_create_media` every 10–15s, never silent >2 min. 4. **FAIL GRACEFULLY:** `kling-v3-i2v` hang/422 → retry once → `ltx-q-i2v` (`num_frames:241`, `resolution:"auto"`) for that shot → Ken-Burns slow-zoom on the keyframe (stillness suits this format); a figure appears in a render → re-fire with the absence clause restated FIRST in the prompt; music comes back sung → re-run with `lyrics_prompt:"[Instrumental]"`. ≤2 retries per shot; an 11-shot piece with a note beats no piece. 5. **DELIVER:** the 1080p reel + one honest line ("shot 7 fell back to ltx-q; light arc held S1→S12; zero figures verified across all 12") + ONE next step ("export the lobby-loop version, or run the next room with the same brand kit"). ## The 12-shot beat sheet (canonical) | Shot | Role | Motion budget | |------|------|---------------| | S1 | ARRIVAL | light brightens 5%, signature object stirs once | | S2 | ABSENCE | dust mote drifts across an empty surface | | S3 | OBJECT-A | one element catches a breath of air | | S4 | DETAIL-A | a glint travels across a surface | | S5 | TEXTILE | drape edge lifts subtly | | S6 | DUST | three motes drift through the beam | | S7 | TOOL-A | light gleam moves across a tool | | S8 | TOOL-B | secondary tool catches + releases a glint | | S9 | BREATH | the room exhales — curtain billows once + settles | | S10 | FLOOR | the light diagonal advances ~3 inches | | S11 | WALL | a warm shadow softens + lifts | | S12 | HELD | the room breathes once, held in full light | Substitute role labels for your place type. For a kitchen: ARRIVAL/STEAM/KNIFE/SPICE/FIRE/etc. For a workshop: ARRIVAL/BENCH/CHISEL/WOOD-DUST/HAND-VICE/etc. The shape is: arrival → small details → atmosphere → held finish. --- ## Reproduce via MCP 1. `brand_kit_create` — palette + style keywords. The kit lives across future installation pieces. 2. `generate_project({ project_id, scenes: 12, style: visual_style })`. 3. 12 × `create_media({ action: "generate", model_override: "krea-2-large", prompt: STYLE + PLACE_SPEC + ABSENCE_CLAUSE + SHOT_DETAIL })` parallel. 4. Audit gate. 5. 12 × `create_media({ action: "animate", model_override: "kling-v3-i2v", source_url: keyframe_url, prompt: ONE_MOTION + PLACE_SPEC + ABSENCE_CLAUSE, duration: 10 })` — fire in batches of 3 to stay polite to the orch. 6. `create_media({ action: "tts", model_override: "gemini-tts", voice: "Gacrux", prompt: narration_script })` parallel. 7. `create_media({ action: "music", prompt: music_brief, lyrics_prompt: "[Instrumental]" })` parallel. **Critical:** museum-installation scores are always instrumental — pass `lyrics_prompt: "[Instrumental]"` explicitly. NEVER omit it; minimax-music sings whatever is in `lyrics_prompt` verbatim, so without this you'd hear `music_brief` sung. See `public/playbooks/music-video-mtv.md` §"The lyrics-vs-style bug". 8. Audit gate. 9. `ffmpeg-concat` the 12 clips → silent base reel, then `create_media model_override:"ffmpeg-mux"` with `audio_fill:"loop"` (voice 1.0 + music ducked 0.15, both trimmed to runtime) → `ffmpeg-export` 1920×1080. Audio covers every second. --- ## Honest limits - **No characters means no character-arc story.** If your brief secretly wants a protagonist, run beat-driven-opera.md instead. - **Long-form (>3 min) needs more shots.** The "12 shots × 10s = 120s" math is the safe default. For 180s use 18 shots; for 90s use 9. Beyond 3 minutes consider chunking the piece into multiple installations. - **Voice script length matters.** gemini-tts pace is ~140 wpm — a 120s narration is ~280 words. Write to the runtime; don't try to fit 500 words into 120s.