Cardstock Diorama Myth Animation (Unsora MCP)

One request — a myth, a premise, or nothing at all — becomes a finished narrated mythic short film in the layered cut-paper diorama style: stacked matte cardstock worlds, three colours and no others, a storyteller's voice, one final MP4. This is animation-movie storytelling — gods, monsters, thresholds, transformations — not a data explainer.

The style, in one line: flat matte cardstock worlds — visible cut edges, soft contact shadows, a strict three-colour palette and no others — a tiny figure dead-centre of a huge symmetrical set, rendered in 3D, not real stop-motion, scenes packed with incident and joined by paper wipes, native impact SFX and paper foley, landing on a face.

The look is a specification, not a vibe. Palette, exposure and pacing are numbers, and every failure is a number drifting off target. State targets as numbers in the prompt, then measure the output (scripts/qc.py). Adjectives don't survive the trip to the model; percentages do.

The failure this skill exists to prevent is not an ugly clip — it's a boring one: one slow scene per block, one camera drift, half the runtime doing nothing. It looks correct and feels dead. Density is the deliverable. The second failure, specific to story work, is a film that narrates instead of showing: if the voice is describing what the picture already says, the picture isn't working hard enough.

The tool split (read this — it is the point of the pipeline):

• Everything generated runs on Unsora. create_image (nano-banana-pro) for the style key, the reference sheets and any hero frame → create_video (seedance-2.0) for every block → create_voiceover (ElevenLabs Eleven v3) for the narration → create_music (mureka-7.5) only if a bed is wanted. Each create_* is async: poll its matching wait_for_*.
• Assembly is ffmpeg, locally. Unsora has no assembler, so the cut, the narration placement inside each block window, and the mix are ffmpeg commands you run yourself. This is an upgrade over any server-side assembler: block windows, VO centring and wordless blocks are all under your control instead of a server's.

Hard preconditions

• Unsora MCP connected and authenticated. Needs create_image, create_video, create_voiceover, list_voiceover_voices, and the matching wait_for_image / wait_for_video / wait_for_voiceover; create_music + wait_for_music only if a score is requested. If they're missing, stop and say so — do not substitute another generator.
• Local ffmpeg and ffprobe. Not optional here — they are the assembler. Without them you can generate clips but cannot deliver a film.
• Web search for the lore pass, whenever the story is a real tradition.
• Recommended: python3 with numpy, pillow, scikit-learn plus curl for the QC pass. If unavailable, run the eyeball checklist in references/style-dna.md and say QC was visual only.

Operating mode

Hands-off. The user delegates story choice, palette, script, voice, casting and assembly.

• Honour anything they specified — the myth, the tradition, the length, the palette, the voice, the orientation. Decide everything else yourself.
• Right before the first paid generation, post one short plan message (the story and which variant, palette, block count, the beat sheet in one line each, voice, rough credit estimate) so the user can interrupt — then proceed immediately without waiting for approval, unless they asked to be consulted. Note that the story shortlist below is a separate, earlier stop: by the time this plan message goes out, the myth is already settled.
• Never stop mid-pipeline to ask what a default answers. Delivering loose clips instead of one assembled MP4 is a failure.
• One exception: which myth. If the user gave no story, present a shortlist and wait for their pick (Phase S). Everything downstream — palette, voice, framing, casting, beats — you still decide. A topical explainer works with any trending subject, so picking one unasked is fine; a myth is the whole creative premise, and guessing it wastes the entire run. This is one stop at the front door, not a habit of asking.

Two style names collide, so be exact: this is the three-colour cardstock diorama, not the sepia-newsprint censor-bar diorama documentary. Do not use its letterpress prop text, and do not use any Mixed Media collage preset. This skill generates its own style key in Phase 1.

Defaults

Story engine

Four things carry a mythic film:

• Every block delivers one nameable event. "What does the viewer see happen that they didn't a second ago?" — a blade is drawn, a wave swallows a boat, a body becomes a tree, the sky splits. If you can't name the event in one clause, the block is a postcard: pack it, fold it, or cut it.
• A 10s block needs two internal paper wipes. A wipe every 3–4 seconds is what keeps this style dense, so each block runs roughly: event → wipe → event → wipe → event, plus its join bookends. Never let a 10-second block ride one slow camera move.
• One talisman travels the whole film. A single physical object carried through every block and escalating — a thread, a feather, a cracked mask, a lantern, a seed. In myth this is the story's memory: it enters in the first beat, changes hands or changes state at the crisis, and pays off in the last frame. Design it before writing block 1.
• The arc, mapped to blocks. For N = 6 use the short column; for N = 8–12 expand the ordeal.

Find the ring. The top-down concentric beat needs a circular geometry to fall into — an arena, a whirlpool, a coiled serpent, a ring of standing stones, a crater, a spiral stair, a circle of mourners, a labyrinth. Myth is full of them. If the story seems to have none, find the ring hiding in it (a duel becomes a ring of onlookers; a funeral becomes a ring of fire). A ring you only look at is a postcard; one something happens inside is a film.

Narration — sparser than an explainer, and that is the point

~12–18 words per block (an explainer runs 20–24). Mythic register: concrete nouns, plain verbs, no hedging, no modern idiom, no jokes. Present or simple past, held consistently. The voice names what cannot be seen — a lineage, a bargain, a consequence — and stays silent about what the picture already shows.

You centre each take inside its fixed 10s window in the ffmpeg pass, and for a myth that is a feature, not the bug it is in an explainer: a 16-word line runs ~6.5s and centres with ~1.75s of pure paper foley at each end — the film breathes, and the impacts land in silence. So aim short deliberately. The ceiling is ≈9.5s per block; past that there is no air left, and you have to speed-fit the take (atempo ≤ 1.15) and the storyteller starts to sound rushed. Reshape long lines into one flowing sentence or cut a clause; narrator voices pause ~0.7s at every period, so two short sentences cost more time than one longer one.

One or two blocks should carry no narration at all — the crisis and often the ring shot play stronger silent, on SFX alone. On this pipeline a wordless block is trivial: generate no voiceover for it and lay nothing over it in the mix. No workaround, no test needed.

Joins — paper-wipe bookends

• Tail of block N — the final 0.4s: the camera rushes into a blank sheet of the join colour filling the frame; the clip ends on flat paper, no detail.
• Head of block N+1 — opens on ~0.25s of the same flat colour, then pulls back into its tableau.

Concatenated, the cut lands between two flat identical-colour frames, so the seam is invisible; the halves read as one continuous ~0.5s wipe. Ground/light wipes into bright beats, accent wipes into dark ones — the colour previews the beat it opens, which in a myth means the wipe carries dread or relief before the image does. The join colour is a property of the join and must be written into both sides. Never chain image → lastImage across blocks; the bookends are the join. Wording: references/style-dna.md §The paper wipe.

This trick is battle-tested against exactly the assembler used here — a local ffmpeg concat — so there is no uncertainty about how the seam lands. Still inspect the seams after the concat and report what you see.

Pipeline

Read references/style-dna.md before Phase 1 and references/unsora.md before Phase 4. Phases S, L, P, 2, 3 and 6 are free; 1, 1b, 4, 5 cost credits.

Job model: every create_* returns a generation.id. Poll the matching wait_for_*; a completed job carries a result URL. That URL — not the id — is what you pass onward: as image and inside referenceImages on create_video, and to curl for the local copy. Everything from Phase 6 on works on local files. Check get_credits once before Phase 4 if you want a balance for the plan message.

Phase S — Story

Story given → use it, go to L.

No story → propose 4–6 myths, show them, and wait for a pick. Do not choose one yourself and start generating; this is the one front-door stop. If the user answers with a vibe rather than a title ("something with a sea monster"), that counts as a pick — take the closest candidate and go.

Bias hard toward action-dense spectacle: monsters, sieges, floods, descents into the underworld, shapeshifting, oaths and bargains with something older than the hero. What makes a myth work in this style:

• A physical world to build in card — a mountain, a sea, a forge, a maze, a storm. Abstract theology needs a physical proxy.
• A creature or force with a silhouette — cut paper lives or dies on silhouette, so a serpent, a wolf, a giant, a wave read beautifully; an unseen dread does not.
• A transformation — this style does metamorphosis better than anything, because a body becoming a tree is literally one stack of card replaced by another.
• A circular geometry somewhere, for the ring beat.
• An ending that lands on a face — a cost paid, a change worn.

Scope each candidate to the runtime. One myth, or one episode of a larger cycle — not a whole epic. "The binding of the wolf" is a film; "the Norse myths" is not. If a tradition only offers an epic, propose the single strongest episode and say which.

Spread the shortlist. Vary the tradition, the scale and the tone — not five Greek myths, not five monster fights. Include one lighter or stranger option; the trickster tale is often the one people pick. Lesser-known myths outperform the famous ones: the audience doesn't know where it's going, and there's no animated-film version in their head to compete with.

Proposal format — one entry each, so a pick takes seconds:

**[Title] — [tradition]**
[One-line logline: who, the turn, and what it costs.]
Ring: [the circular geometry] · Creature: [the silhouette] ·
Transformation: [what becomes what] · Talisman: [the travelling object]

If a candidate is missing one of those four, say so rather than inventing it — a myth with no transformation can still be the best pick, and knowing what it lacks tells the user what the film will lean on instead.

Calibrated seeds (a spread that works; don't just recycle these, but match this level of concreteness):

One caution. Some stories are living sacred tradition rather than public folklore — especially Indigenous and closed-practice material. Prefer widely published, public-domain myth; if the user asks for something that reads as sacred or closed, say so plainly once, offer a near neighbour, and follow their call. Don't lecture, and don't refuse a myth just because it is religious.

Phase L — Lore

Never write a myth from vibes. Search the tradition and collect: which variant you're telling (most myths have several irreconcilable versions — pick one and say which), the names as commonly transliterated, the iconography the tradition itself uses (what the creature actually looks like in its own art, what the hero carries, what the place is described as), the ending as the source gives it, and any detail specific enough to be worth stealing — a number of doors, a colour of a horse, the thing said at the threshold. Specifics are what make a mythic film feel found rather than generated.

Then decide what you are changing and own it. Compression is fine; contradicting the tradition's ending is a choice to make deliberately and mention in the delivery message.

Do not describe characters as they appear in copyrighted animated films. Render from the tradition's own iconography. See the moderation and IP notes in references/unsora.md.

Phase P — Palette

Pick one of the five tested palettes and hold it for the whole film: Reference (cream/vermilion/navy), Sable (lilac/violet/moss), Cobalt (cream/vermilion/teal), Iris (cream/violet/forest), Glacier (ice/vermilion/bitumen). Match the myth's world — the mapping is in references/palette.md §Picking a palette for a myth. State the choice and the three hexes in the plan message. The maths, the full prompt-ready ranges, and the rule that something must oppose the accent by ~140° are in the same file.

Phase 1 — Style key

create_image, model: "nano-banana-pro", aspectRatio = the run's ratio, with the STYLE KEY prompt from references/style-dna.md §Style key, palette hexes substituted. wait_for_image, then keep the result URL — that URL is the style key, and it goes in referenceImages on every clip. QC it before spending on clips: a bad key poisons the run and re-rolling costs one image.

Phase 1b — Casting (do not skip this)

A myth follows the same figure for its whole runtime, so character drift is the single biggest technical risk in this skill — a bigger one than palette drift. In the explainer version sheets were optional; here they are the job.

Generate, before any clip: a character sheet for the protagonist, one for every creature, and a location sheet for every place appearing in 2+ blocks. Formats in references/style-dna.md §Reference-sheet formats — always 16:9. The talisman gets its own sheet too, at 1:1, from the format in §The talisman / prop sheet: it must stay recognisable at every scale, and a second sheet if it changes state at the crisis.

Then pass each sheet's URL in referenceImages on every block that uses it, alongside the style key, and refer to it in the prompt by slot — "the same figure from @Image2". Mind the slot maths: on create_video, image is @Image1, so the first referenceImages entry is @Image2. Pass only the sheets that block actually uses; a dangling reference invites drift, and Seedance caps referenceImages at 9. QC every sheet and show them in the plan message: a flawed sheet propagates into every block that references it, and this is the cheapest moment to re-roll.

Phase 2 — Beat sheet + narration

N beats on the arc table above, labelled Block 1 … Block N. For each: the event in one clause, the wipe-in colour, which sheets it uses, and its narration line (~12–18 words, or none). Plain spoken text only — no stage directions, no parentheticals, numbers spelled out. Read the whole narration start to finish aloud: it should sound like one voice telling one story, not N captions. Cut every line that describes what the picture already shows.

Phase 3 — Block prompts

One video prompt per block. Each block is a 10-second create_video call that takes the style key and this block's sheets in referenceImages. Write all N with the block template in references/style-dna.md, then run the per-block checklist there. (For the ring block and the final face block, consider the optional hero-frame lock in style-dna.md: generate that block's opening frame with create_image first and pass it as image, which buys precision on the two shots that carry the film.) Non-negotiable in every prompt:

• The medium sentence (stacked matte cardstock, visible cut edges, soft contact shadows, rendered in 3D) and the AI-default negative.
• The palette stated three ways — hexes with ranges, percentages, and the exclusion naming the colour you removed.
• The exposure block. Never a global exposure lock — the swing is the style.
• Paper logic named for every element, especially anything entering mid-block, and for every creature, flame, wave and wound.
• Every camera move pinned to a timecode and forbidden to pause.
• Both wipe bookends with colour and duration (except block 1's head, block N's tail).
• Sheet binding by slot for every figure, creature, place and the talisman.
• Never the word "hold" — the model freezes on a full-detail frame. Named micro-motion instead.
• The audio block: native paper foley and impact SFX, No music. No voice. The narration is a separate track; a voice inside the clip fights it.

Phase 4 — Clips

Submit N create_video jobs — model: "seedance-2.0", duration: 10, resolution: "720p", generateAudio: true, aspectRatio identical across the run, style key plus that block's sheets in referenceImages. Exact call template and the moderation and IP map are in references/unsora.md. Poll each with wait_for_video, curl the result to clips/block_NN.mp4, and record every generation id against its block in the manifest; re-submit only failures.

Violence, the way this style does it. Myth is violent and the style handles it natively because it is abstract: a wound is a vermilion paper shard, a beheading is a silhouette and a scatter of punched confetti, a flood is stacked wavy card. Describe the paper event, never the gore — that keeps the film true to the style and clears moderation at the same time. Explicit blood and viscera will flag and are off-style anyway.

QC each clip if the local pass is available: python scripts/qc.py video clips/block_NN.mp4 --ground '#…' --accent '#…' --void '#…'. Report numbers honestly, failures included — luma swing (target ~150, under ~80 is a dead clip), ground share near 45%, no fourth hue, and the flat head/tail runs that prove the bookends landed. Re-roll only on a fail; never re-roll a good block "for consistency" — a re-roll comes back different and breaks the bookend match with its neighbour.

Phase 5 — Voice

1. list_voiceover_voices → auto-pick a storyteller from the twenty Eleven v3 voices: deep, unhurried, warm, slightly formal, comfortable with silence. Read the register off the myth — a Norse saga and a trickster tale don't want the same voice. Note the exact voice_id and reuse it on every block. If the user wants to pick, the tool renders a picker with previews: show it and wait, and don't repeat the voices as a list in your reply.
2. One create_voiceover call per voiced block. stability: 0.6 for a steady narrator (lower gets theatrical, higher gets flat); leave similarity at its default. Generate nothing for a wordless block. Poll wait_for_voiceover, curl each mp3 to vo/vo_NN.mp3.
3. ffprobe every take for its real duration and record it. A 12–18 word line landing at 5–8s is correct — the block is 10s and the assembler centres the take so it opens and closes on foley. Over ≈9.5s there is no air left: cut a clause or merge two sentences and re-voice that block only, or speed-fit it in the mix (atempo ≤ 1.15, pitch-safe).

<#0.5#> between words inserts a half-second pause and is worth using at a threshold line — but a pause spends the same window the silence at each end does, so check the take still lands under the ceiling.

Phase 6 — Assemble (automatic, mandatory)

The moment all clips and takes are done, assemble in the same run without being asked. ffprobe the finished clips' real width/height and encode to those — if the clips rendered in a different aspect than planned, the assembly matches the clips, not the plan.

1. Concat the clips in manifest order. Every join is wipe-bookended, so the cut lands between two flat identical frames and needs no crossfade.
2. Place each take at its block's window: offset_ms = block_start_ms + (10000 − vo_duration_ms) / 2. That centring is what gives the myth its air, and on this pipeline it is your arithmetic, not a server's. Wordless blocks get nothing.
3. Mix VO over the clips' native SFX (kept low so impacts read), plus the music bed ducked under the VO if there is one.
4. Verify duration, dimensions and audio streams before presenting.

Exact commands — concat, loudnorm, the centring offsets, the sidechain duck, verification and resume — are in references/unsora.md §Assembly. Then present final.mp4 with the host's native file delivery.

Delivery

Final message: the video, the myth and which variant you told in one sentence, the talisman and what it paid off into, the locked palette (name + three hexes), the full narration so the user can reuse it, the QC numbers if measured, anything you deliberately changed from the source, and where you read the tradition. Then offer — never run unasked — social posting (Unsora create_post, its own explicit yes) or a titles pass.

Failure handling

Reference files

• references/style-dna.md — the full cardstock visual system: locked slots, paper logic (including creatures, water, fire and wounds), exposure swing, camera, motion, audio vocabulary, the AI-default negative, the wipe wording, the style-key prompt, the block-prompt template with a worked mythic example, sheet formats, and the per-block checklist. Read before writing any prompt.
• references/palette.md — the palette maths, the five tested palettes plus one documented failure, prompt-ready ranges for all five, how to state a palette three ways, palette-by-myth-world, and how to build a sixth with scripts/palette.py.
• references/unsora.md — exact Unsora call templates for every phase, the job and polling model, the manifest schema, the moderation and IP map, cost arithmetic, every ffmpeg command for the assembly, verification, resume-after-failure, and the optional posting flow.
• scripts/qc.py — measures frames (frame) and clips (video) against the palette, exposure and wipe targets. python scripts/qc.py --help.
• scripts/palette.py — solves a new three-colour palette at the right luma values.