skills/faceless-video/references/prompts.md

# prompts.md — faceless-video prompt templates

All image/video prompts in **English**. Only narration switches to the user's
language. `{STYLE}` = the ONE style formula, pasted byte-identical everywhere.

## 0. STYLE FORMULA (write once, reuse verbatim)
80–100 words, ≤2 sentences: line/surface work · shading · palette · one signature
accent · background treatment · motion timing · render discipline (positive form).
Always non-photorealistic. **House styles ship pinned formulas — take them verbatim
from `style-editorial-collage.md` (Editorial Motion Graphics — default for History AND
Explainer) / `style-paper-diorama.md` (History's named alternate), including their
PALETTE LOCK lines and {MOTION} mappings; do not re-derive.** Stickman Cartoon
(Explainer's second main direction) uses this generic formula (never name a real
comic/brand/IP):
> flat 2D webcomic cartoon, extremely minimal — uniform thin even-weight black
> outlines, egg-shaped heads with tiny dot eyes and a single line mouth, plain
> noodle limbs, solid flat color fills with NO shading, NO gradients, NO texture,
> deadpan minimalist design, plain flat solid-color backgrounds.

## 1. Style key (seedream_v5_pro, 1k, chosen aspect — default 16:9)
Make it **pretty & legible**, not formless blobs — a small representative vignette so
the look reads clearly: "A clean {STYLE} STYLE SAMPLE: one simple recognizable subject
(a small friendly object or mini-scene) rendered in the style so its line weight,
shading and colours read at a glance. Balanced, attractive, easy to read. No palette
strip, no colour chips, no swatch bar, no colour squares, no labels, no text, no
watermark, no reference-sheet layout." (It's still a look anchor, not a final frame.)
**Tell the user** when showing it: a style key can look a little odd/abstract — that's
normal, it's only a reference for the look.
From uploads (style only), verbatim prefix: "Take only the visual render style and
color grading of the input image(s)… never use the characters, inscriptions, etc."

## 2. Assets (seedream_v5_pro, always 1k; pass style key as `image`; {STYLE} verbatim)
- **Character (2:3, 1k):** "Full-body character centered on a plain flat solid-color
  background, in THIS EXACT style: {STYLE}. Character: {desc — distinctive, readable
  silhouette}. No text, no watermark."
- **Location (CHOSEN aspect, 1k):** "{interior/exterior} … in THIS EXACT style: {STYLE}.
  {key furnishings + ONE named anchor object with a position}. Empty room — no people, no
  characters, no figures. Wide establishing. No text, no watermark."
- **Prop (1:1, 1k):** "A single isolated prop centered on a plain flat solid-color
  background, in THIS EXACT style: {STYLE}. Object: {desc}. No hands, no scene, no
  other objects. No watermark."

## 3. Block = one 10s H3 clip with FIVE hard-cut shots (the core template)
Each `generate_video_batch.requests[].params` uses model `minimax_h3`,
`duration:10`, `aspect_ratio` = the CHOSEN aspect
(default `"16:9"`; this workflow uses ONLY 16:9 / 9:16), `resolution:"2K"`, `medias` =
location → characters → props (role `image`), **max 7 refs per call** (keep the production packing even though H3 supports more) — only this block's assets; trim extra props/coverage first, never the
location or an on-screen character. If the response is a preset
recommendation instead of a job → resubmit the same call with `declined_preset_id` = the
recommended preset's id. Follow the Phase-4 live-schema/native-audio preflight;
only send `generate_audio:true` when that field is declared.

```
Style: {STYLE}; {MOTION} — the visual style is EXACTLY as in the reference images, same rendering, same surface treatment.
PALETTE LOCK: use ONLY the colors and the background treatment of the reference images ({e.g. flat off-white/clean webcomic background}). Do NOT introduce any new or foreign colors, no colored/gradient/painted backgrounds that aren't in the references, no recoloring of characters or objects.
A single 10-second scene of FIVE hard-cut shots. Do NOT open the video on any reference image or show a sheet/swatch — stage everything fresh, matching characters, room, colors and background to their references. Characters only emote and gesture, they do NOT talk. Motion starts on frame 1 (no opening freeze).
REFERENCES (look, identity, palette): @Image1 = LOCATION ({desc}). @Image2 = {CHARACTER A} ({desc}). @Image3 = {CHARACTER B}. @Image4 = {PROP}.
SHOT 1 — 0.0s to 2.0s — {SIZE+ANGLE}: {beat}.
HARD CUT.
SHOT 2 — 2.0s to 4.0s — {DIFFERENT SIZE+ANGLE}: {beat}.
HARD CUT.
SHOT 3 — 4.0s to 6.0s — {DIFFERENT SIZE+ANGLE}: {beat}.
HARD CUT.
SHOT 4 — 6.0s to 8.0s — {DIFFERENT SIZE+ANGLE}: {beat}.
HARD CUT.
SHOT 5 — 8.0s to 10.0s — {DIFFERENT SIZE+ANGLE}: {payoff}.
Five hard-cut shots at 2.0s, 4.0s, 6.0s and 8.0s, no dissolves, no fades. {MOTION}, continuous motion within each shot, never freezes.
```
**~2s per shot is the LAW: no shot longer than 2.5s** — a frame hanging 3–5s reads
as a slideshow, and every shot needs visible ACTION, not a held pose. Degradation:
a block failing generation twice at 5 cuts drops to 4 (2.5s grid) — that block only.
**Kids blocks use FOUR cuts** (2.5s each: `SHOT 1 0.0–2.5 … SHOT 2 2.5–5.0 … SHOT 3
5.0–7.5 … SHOT 4 7.5–10.0`, cuts named at 2.5s/5.0s/7.5s) with the WIDE → CU character
(reaction) → ECU detail → MEDIUM pattern, order varied every block — full pattern +
interplay staging in `${FACELESS_STYLES_DIR}/references/kids-styles.md`. Everything below applies to both
templates:
```
AUDIO: {diegetic SFX only} — no voice, no narration, no music.
NEGATIVE: opening on a reference image, a sheet/swatch or a static first frame, leading freeze, dissolves or fades, NEW or foreign colors, colored/gradient/painted background not in the references, recolored characters, style drift, extra people, cloned characters, characters talking, lip-sync, on-screen text, captions, photorealism{2D-ONLY: , 3D render}, watermark.
```
**{MOTION} — match the style family.** Flat 2D styles (webcomic / stick-figure / flat
cartoon, Editorial Collage): `simple limited animation on twos` (collage prefers its
own token — see `style-editorial-collage.md`). Dimensional styles (Paper Diorama, 3D Papercraft, Fluffy
Toy, claymation…): `smooth simple handcrafted motion, subtle stop-motion feel` — and for
these DROP `3D render` from the NEGATIVE (it fights the look; the photorealism ban always
stays).
Vary shot sizes and ANGLE on every cut (WIDE / MEDIUM / CU / ECU / OTS / low / high).
**OTS is valid only when a named on-screen character's shoulder/head is intentionally
visible in the foreground.** In object-only, diagram, empty-location, or any other
characterless shot, never write OTS; use overhead/top-down, low/high, macro, lateral, or
another coverage angle instead. OTS without a referenced character makes the model
invent a person.
Only the FIRST block of a new location may open on a full establishing WIDE; later
blocks in the same location open on a fresh close/medium/coverage angle — never
re-establish the same wide. **≤2 consecutive blocks per location**, then move (new
location / coverage angle / variety insert). Vary the character's distance & screen
position; never park them back at the opening framing. **Every block carries at least
ONE impact beat** (a slam, stamp, snap, whip-pan hit, collapse) — name it in a SHOT line
and echo it in AUDIO; reference-grade explainers land an accent roughly every 3 seconds.
**Action grammar (all styles):** write SHOT beats as pure CHOREOGRAPHY — the {STYLE}
formula owns the look, so no style/color/material words inside beats; elements enter
STAGGERED ("A, then B, then C" — never simultaneously) with landing verbs (snaps into
place, stamps down, drops with a bounce, settles); exactly ONE camera behavior per
shot, stated once (slow push-in / static / gentle drift / whip) — two moves in one cut
reads as AI soup; the FINAL shot eases into a stable, still micro-moving final frame
(settled ≠ frozen — rule 21 holds).
On-screen TEXT is unreliable →
use SYMBOLS (⚠ / $ / ✓ / ✗), let the VO carry words. **Sole exception — Paper
Diorama's letterpress prop label** (ONE 1–2-word label per scene, fenced as
`no text anywhere except "<LABEL>", no gibberish letters, no captions` in place of the
NEGATIVE's `on-screen text, captions` pair — see `style-paper-diorama.md`); every other
style keeps the full text ban.
AUDIO line per block: **SFX follows choreography 1:1** — every motion verb gets at most
ONE cue, 2–4 cues per block + one room-tone bed (whooshes on moves, a stamp hit on the
impact beat, ticks on data steps, sparkles/dings for Kids, ambient of the place);
nothing sounds that didn't move; percussive, never musical — still "no voice, no
narration, no music" (voice + music bed are added in the Phase-6 mix, not inside the
clip).

**Compose from assets, full scenes.** Every clip is staged from the approved Phase-2
assets (location + characters + props as `image`) — a real dressed scene with
depth, matching the assets' colors/background, never a lone object on a foreign background.

## 4. Retry ladder (H3 nsfw / preset)
Preset recommendation instead of a job → resubmit the SAME call with `declined_preset_id`
= the recommended preset's id. On `nsfw`: resubmit ≤3× (change `seed` only if declared); if still failing,
reword (drop child/kid/childlike, avoid tight animal-face CU, calm the pose). On terminal
`nsfw` after retries: swap the beat's framing. Never drop a block.

## Scriptwriter

Phase 3 owns the mechanics (N blocks, 5 varied ~2s shots per block — Kids 4, one VO line
of 20–23 words, Kids 17–21, paced toward a full-block read, humor per channel type).
This file owns the SHAPE. The
failure mode of a faceless script is not "boring" — it is **a list**: fact,
fact, fact gives the viewer no reason to keep watching. Facts are material;
shape is the product.

## 1. Through-line first

Before Block 1 exists, name ONE physical object or process that appears in
every block and **escalates monotonically** (shorter / taller / fuller —
never merely recurring as decoration): a fuse burning down, a stack growing
until it blocks a door, a map filling with dots, a plate scraped into an
overflowing bin.

- It must be renderable in the chosen style — **give it a prop asset in
  Phase 2** so it stays identical across blocks (diorama: the burnt-orange
  prop; collage: its own cutout).
- Each block's shots show it at its current state; the **payoff block
  resolves it** (the fuse reaches the keg). If the finale doesn't pay it off,
  pick a different object.
- Name the through-line when showing the script at SCRIPT LOCK.

## 2. The arc, sharpened (hook → build → turn → payoff)

- **Hook — cold open, SHORT.** The most surprising concrete thing, stated
  flat — understatement makes a big number bigger. **Keep block 1's opening
  line PUNCHY: the first sentence is ≤8 words** (ideally the raw fact — "Most
  of the ocean has never been seen."), THEN the block's line fills out to the
  normal 20–23-word density. No greeting, no "in this video", no
  throat-clearing, no windup clause before the hook lands. (Kids keeps its
  warm host manner and catchphrases — the cold-open brevity binds History and
  Explainer.) A sharp direct question is a legal hook (also ≤8 words); if
  used, withhold the answer until the payoff and let it live in the visuals.
- **Early stakes.** One beat answering "why should I care": the misconception
  ("everyone blames X — wrong") or the proximity (it touches the viewer this
  week).
- **Build — evidence, escalating.** ONE idea per block (the cuts are angles on
  that idea, not separate ideas). Every block anchored to a concrete:
  number, date, place, named thing, comparison. **The reorder test:** if the
  build blocks could be shuffled without loss, you listed instead of
  escalated — each must be bigger, weirder, or more specific than the last.
- **Turn.** The counterintuitive reveal, usually block N−1. Test: does it
  change what the viewer thought two blocks ago, or just summarize? A
  summary is not a turn.
- **Payoff.** Land the answer, then a kicker that **reframes the hook** —
  echo its image or number with new meaning, don't repeat it. Through-line
  resolves here. Humor styles still apply (deadpan setup → absurd punch).

## 3. Research targets (factual topics — before scripting)

History and factual Explainer topics: quick web research first, never script
from memory alone. You're done when you hold: the **hook stat** (the number
that stops the scroll), **3–5 concrete facts**, the **counterintuitive turn**,
and vivid physical specifics the shots can stage. Cross-check spoken numbers
against a second source; keep a Sources line for delivery. Fantasy/Kids
topics: skip research, invent freely (but nothing fake about the real world).

## 4. VO line craft (works with vo_and_captions.md)

- **20–23 words, Kids 17–21, comma-light, paced to fill most of the block** (the
  `[00:00-00:09]` timecode carries delivery but does not pace ElevenLabs — see
  vo_and_captions.md), **at most two sentences** — TTS
  pauses ~0.7s at periods and ~0.5s at commas, so one flowing clause fits where
  three clipped sentences both overrun AND sound pausey. Performed brackets
  (`[scoffs]`) cost ~1s each — budget them.
- **Pacing quality:** prefer a line that fills most of its block. Rewrite a clearly
  sparse first take denser, but do not chase an exact measured duration or fail a run
  for minor underfill after bounded retries. Never pad with filler or stretch audio.
- Numbers spelled out. No sentence or near-identical phrase in two blocks.

## 5. Rewrite pass (run before SCRIPT LOCK / before voicing)

1. Does the hook work standalone, zero context?
2. Is every number/date traceable to the research?
3. One idea per block; every cut serves that idea?
4. Do build blocks fail the reorder test (i.e., escalate)?
5. Does the turn surprise rather than summarize?
6. Does the payoff's kicker echo the hook and resolve the through-line?
7. Every full-block line 20–23 words (Kids 17–21), flowing/comma-light and
   naturally paced toward a full-block read; humor lands per
   the channel type's voice?
8. Does every block's SHOT text stage the CONCRETE nouns/quantities of ITS OWN line
   (spices on screen when the line says spices)? No action motif repeated back to back,
   none beyond its budget (long-form: the motif ledger in history-longform.md)?