video-prompt-dialects · diff

git:20260929.e9767cf to git:20261006.4886132

12 added, 4 removed. Audit A to A.

---

name: video-prompt-dialects

description: >

Writes the prompt for a video or still generation in the DIALECT the venue speaks — Seedance 2.5 /

2.0 (the 【】 template, @ImageN roles with exclusions, cause-before-reaction event script), MiniMax H3

(typed Subject labels, verbatim dialogue), the Create-a-Meme front-end (@refN, 2000-char cap),

kie gpt-image-2.5 / nano-banana-pro stills — from a shot document and its accepted references, then

lints it before the gate. Use when a prompt is being authored or rewritten, when a venue refuses or

misreads one, when "cinematic" or a negative list is about to be written, when a shot must continue

a keeper or match a start image, or when the same shot moves to another venue. Triggers — "write the

prompt", "prompt for this shot", "rewrite for Higgsfield", "the model widened the shot", "it rendered

what I forbade", "nsfw on the prompt words", "lint the prompt". Not for the reference set or the

start image the prompt cites — use video-refs-continuity. Not for the cost line and the submit — use

video-gen-cost-gate.

allowed-tools: Read, Glob, Grep, Write, Edit, Bash(python3*), Bash(ls*)

---

# Video Prompt Dialects

A prompt is compiled, not written: the shot document supplies the facts, the accepted references supply

every visible attribute, the venue's **dialect** supplies the shape, and the prompt says only what neither

a reference nor a parameter can carry — who is which reference, what happens in what order, what the

camera does, what is heard. Three words carry the method: **dialect** (one venue, one schema — a

front-end's feature set is not the model's contract), **contract** (what the prompt may say: the

relationship between references, the action, the exclusions; never a description of a cited reference,

never a parameter), **tail** (the trailing exclusion region — the only place a prohibition steers instead

of injecting).

**What varies.** Each dialect is one model's contract at one version: its layout, its caps, the words its moderator

refuses and the beat density it renders were measured there, and a new model or version is read from its own docs before

its first prompt. The style prefix belongs to the campaign — its genre, aspect and look; a creator-style spot replaces it

with the phone-native genre dialect ([`references/PHONE-NATIVE.md`](references/PHONE-NATIVE.md): the device named as the

camera, never as a prop; positive-spec only; the video-call register; a stressed word in capitals; the prompt locked after

the first line), layered on the model's own contract. Pointing at references,

affirmatives in the body and the lint do not move — `video-production/references/WHAT-VARIES.md`.

## 1. Resolve the venue, mode and layout

|---|---|---|---|---|

| Seedance 2.5 (Higgsfield `omni_reference` / `t2v` / `video_extension`; fal r2v / i2v) | `@Image1…` per type, in upload order; a start image is passed separately | style prefix + the 【】 sections (below); numeric cut budgets when the shot document pins beats, stage form otherwise | ≈ 2.5–3.2 k chars works; 6 k wire; > 5 k warned | dialogue in `{ }` per speaker, or quoted in sentence case; supplied-audio words must NOT appear; edit/extend are inferred from wording (a plain reference request bills a whole new video); parameters never in the prose |

| Seedance 2.5 on monid (`bytedance /v1/video/seedance-2.5`, the pay-as-you-go route) | the same `@Image1…` dialect in the PROSE, but the call is a `content[]` array with a `role` per picture; ordinals count per type in array order, so a `first_frame` takes `@Image1` and the refs shift | as 2.5 above | **6000 chars, the model's own** | `ratio` only on t2v/r2v — first/last-frame, edit and extend REQUIRE `adaptive`; edit and extend are phrasing, and a plain reference request bills a whole new video |

| MiniMax H3 (fal, 768p min) | `<Subject N>` / `<Picture N>` defined once, cited in the definition; `Image 1` in prose | named fields in fixed order, blank-line separated; `[Shot N] At 00:MM.mmm` strictly increasing | 7 k | dialogue VERBATIM in `<d>[English] …</d>` — the opposite polarity; `prompt_expansion_mode` rewrites and invents — `fast`, persist and diff `expanded_prompt` |

| Create a Meme (web front-end) | `@ref1…@ref9` by upload order; `[Video1]` motion clip; start + end frame outside the slots | one paragraph | **2000 chars, hard** | never press "Enhance prompt"; the 【】 layout written out does not fit |

- | kie gpt-image-2.5 (i2i, flare / sunburst — the default) · nano-banana-pro (the fallback) | ref 1 = the image being edited; "@ref2 the product" | prose: a transformation ("must appear 100 % identical to its own reference; only the pose, light and setting change"), the preserve list, the conflict priority, the exclusion per ref | — | gpt2 composes and copies; nano edits surgically and IGNORES sheets when composing; the words a moderator refuses go through the copy route |

+ | kie gpt-image-2.5 (i2i, flare / sunburst — the default) · nano-banana-pro (the fallback) | ref 1 = the image being edited; "@ref2 the product" | prose: a transformation ("must appear 100 % identical to its own reference; only the pose, light and setting change"), the preserve list, the conflict priority, the exclusion per ref | — | gpt2 composes and copies; nano edits surgically and IGNORES sheets when composing; the words a moderator refuses go through the copy route; a photoreal still is Sunburst at Flare's kie price, written as a capture, clean of grain (REALISM.md § 2) |

The full contracts, wire limits and moderation classes: [`references/DIALECTS.md`](references/DIALECTS.md).

+ **A live-action job reads [`references/REALISM.md`](references/REALISM.md) whole before its first prompt** — stills on

+ GPT Image 2.5 (Sunburst, the capture sentence, night exposure, a clean image, realism edits that keep an approved still)

+ and footage on MiniMax H3 (the style first, the camera vocabulary, grain in the video, the safe negative form, singing,

+ a front-end's one scene text). A governed project's skill gate denies a prompt write until this skill is loaded.

Parameters — resolution, duration, aspect, audio toggle — never appear in the prompt; they ride on the call.

**Done when:** the dialect is named, the reference slots are listed in upload order with their names, and

- the layout (cut budgets vs stage form vs single line) is chosen from the shot document.

+ the layout (cut budgets vs stage form vs single line) is chosen from the shot document, and a live-action job has

+ read REALISM.md.

## 2. Compile the house template from the shot document

Seedance 2.5 ads shape, section by section — the worked examples are in

[`references/HOUSE-TEMPLATE.md`](references/HOUSE-TEMPLATE.md):

1. **Style prefix**, glued to every prompt of the campaign (photoreal live-action commercial, 9:16, the

light, clean exposure, no grain, no on-screen text, "the camera is already moving on frame one").

2. **【Generation Goal】** — the script's own sentences, quoted; a continuation opens "The shot begins on the

start frame and continues from it without any cut".

3. **【Reference Asset Roles】** — one line per reference: `@ImageN is X: use only its A, B, C; never its

D` (a shape sheet: "the silhouette and colours only, never the size shown"; a size ref: "the

size relative to the tie knot only"; a room ref: "its ground, furniture and daylight only"); the

start frame sentence fixes the camera, the people, their places and clothes "exactly"; a video ref

carries "his VOICE only — he stays off-camera"; a look plate's line is look-only ("take only its

light, palette and contrast; never its content"). Unused uploads are listed as unused.

4. **【Subjects and Relationships】** — the ledger, per person: age, build, hair, clothes, seat, hands;

"only these N people"; who faces whom; the geometry as THIS camera sees it ("the white front door in

the far wall at the centre of the frame, directly facing the camera"); "nothing added to him" for every

untransformed character. At most ONE anchor coordinate; everyone else by relationship.

5. **【Event Script】** — `Cut N (t–t s):` the framing FIRST on its own clause, then the action: cause

before reaction (contact → motion → sound → response); the start = the previous shot's last state

(standing stays standing; after a burst "blanketed and still raining, NO bang — the device already

fired"); every eyeline a target ("eyes locked on the doorway at the right edge, never at the lens");

hands placed ("the free hand up beside the head, held there"); the hit named with its sound in `< >`;

a runner described from behind; the last frame keeps the subject in it, small.

6. **【Audio】** — room tone, the named sounds, the lines: `{line}` per speaker in sentence case, a stressed

word in capitals (a whole line in capitals reads as shouting; profanity trips the filter → a mouth-shape twin, dubbed later), "no words, nobody

shouts" for everyone else, "no music" unless scored.

7. **【Maintain Consistency】** — the locked facts: the door sentence copied verbatim from the geometry

record; seats, wardrobe, counts, the product "only ever @Image1"; the lock line for a fixed frame ("the

camera never pulls back, tilts or widens; no object of any kind appears in the frame").

Rules that hold in every section:

- **Point, never describe.** A cited reference gets a role and an exclusion; any adjective about what it

shows can only contradict it.

- **Referent + result, never mechanics.** One familiar, filmed motion ("moves like a cross-country skier")

and the visible outcome; a category ("big wild animals") resolves to the wrong animal; joint-level gait

words compose into a hop.

- **Affirmatives in the body; the tail carries the prohibitions.** `no head bounce` in the body rendered a

bounce; `THEY DO NOT MOVE IN UNISON` rendered unison. Sanctioned tail negatives on Seedance 2.5:

subtitles, music, logos, watermarks, and what the shot document explicitly forbids.

- **Negatives are scoped to the generation's length.** "No hidden splice, no unmotivated cut, no abrupt push-in"

belongs to one continuous shot of a few seconds. A long generation — a whole continuity partition — composes its

own coverage, and a cut inside it carries consistency by construction; those words forbid the reason it was

generated long (HOUSE-TEMPLATE § The negative tail; LINT L34).

- **A clip with several cuts or several people is directed for the model** — a lock line closing every cut (light, each

principal, position, facing, the camera's side), one action per cut with idle business for the rest, reactions as

reverses, crossings on screen; and a spoken clip is sized on the load ladder before it is bought (DIALECTS § Multi-shot,

multi-person direction; LINT L37).

- **Never NAME an off-frame object** — the model widens to show it and invents it; write the look

("jabs a finger down, out of the bottom edge"); the name may appear only inside a quoted line.

- **Bind the camera to one subject or one fixed geometry**; a move that cannot hold its named subjects

(a push-in while three diverge) is resolved by the model discarding the part you cared about. When a scene

proxy exists, the clause is derived from the authored move (`scripts/camera_clause.py`) and written per

cut — never one camera sentence for a clip that cuts.

- **Blocking in metres, on named floor and furniture sides** ("two metres in front of the cab glass",

"along the left side of the table, on the carpet") — a step "toward camera along the table" walked on it.

- **Beat density ≤ ~1 beat per 3 s**; the model renders locomotion and physical state, not intent — plan

comedy around situation, and shrink the beat count before shrinking the seconds. A reference that cuts faster

(`ad-spot-preprod` REFERENCE-TEARDOWN.md's grid) is realised by cuts — composed or edited — never by more actions per

shot. ⚠ UNDER TEST since 2026-09-18, not rules: a timestamped beat script as a long generation's prompt and an

identity block at the prompt's end — the house template stays the default until a costed A/B lands.

- **The verb must be the scripted action** ("he drinks from the bottle", not "the others hold him") and the

reaction's emotion matches its cause (shock, not joy, at a collapse).

**Done when:** every noun and action in the prompt traces to a script sentence, the ledger or a

reference role; every reference in the slot list has exactly one role line; the start state equals the

previous keeper's last state; and the lines are quoted in sentence case, a stressed word in capitals.

## 3. Lint before the gate

```

python3 scripts/prompt_lint.py --dialect seedance-2.5 --refs ROOM,W1,PRODUCT-sheet prompts/r2v/S02-G4.txt

```

Mechanical checks, each from an incident ([`references/LINT.md`](references/LINT.md)): slot ranges

expand with no double assignment, no gap, within the venue cap; negations in the body; ALL-CAPS tokens

outside the stoplist and outside a quoted line, and a whole quoted line in capitals; the venue's moderation words; parameters in the prose; a named off-frame object;

category referents and gait mechanics; "aerial view"; camera move vs diverging subjects; a framing that

is not on its own clause; beat density; the char cap; a prompt that repeats a reference's description;

single-shot negatives on a long generation (`--duration`); a per-line prompt that drifted from its locked first line

- (`--locked`); and by eye, a look plate restated in prose and a

- citation that resolves to no asset.

+ (`--locked`); the photoreal rows — quality words, grain asked of a still, an H3 negation naming a scene object, H3's

+ style not first (`--selftest` proves each one fires and stays silent on MiniMax's own examples); and by eye, a look

+ plate restated in prose and a citation that resolves to no asset.

The lint is advisory except the slot and cap rows — but a warning row is answered in the prompt or in the

GO ask, never ignored.

**Done when:** `PROMPT-LINT OK` (warnings named) and the prompt file is the one the gate and the submit

will read.

## 4. When the venue answers wrongly

| symptom | cause | rewrite |

|---|---|---|

| `nsfw` within a minute | a prompt WORD, not the picture (an anatomical noun, "thrusting", a suggestive look described) | paraphrase the shape ("the exact @Image silhouette"); the profane line becomes a mouth-shape twin and is dubbed in the character's own cloned voice later; frame chest-up or knees-down, never the waist of a bare torso |

| a fresh bang at frame 1 | the prompt said "a few last pieces drift down" | "the room is blanketed, still raining, NO bang — the device already fired" |

| the shot widened / pulled back | an off-frame object named; a camera move that cannot hold the subjects | write the look, not the thing; bind the camera; lock line in the tail |

| the line-deliverer faces the lens | the default | write the addressee: "turns to his right toward the friend at the mantel and shouts at him" |

| the first word of a line dropped (2/2 under one layout) | layout-sensitive | put the line in `{ }` on the audio line; A/B a layout change same-seed before adopting it; whisper the takes |

| a label or arrow rendered in the scene | a schematic map as a reference | blocking in prose; geometry from frames |

| the model invented a description of the reference (H3 expansion) | `prompt_expansion_mode` | `fast`; diff `expanded_prompt`; roles with exclusions |

**Done when:** the rewrite names the cause row, the changed words are the only change, and the GO ask

says which row it answers.

## 5. Record and hand off

The prompt file lives beside the shot document (`prompts/r2v/<SHOT>.txt`, `prompts/stills/<ID>.txt`)

with its reference names in slot order in the shot list; the gate and the submit read the file, never a

pasted string. Hand to `video-gen-cost-gate` with the refs and the start image named.

**Done when:** the prompt file, the slot list and the shot-list row agree, and the cost-gate ask cites

the file path.

## ❌/✅

```

❌ "@ref2 the middle one, red-ginger, long face" ✅ "@ref2 — take his complete appearance from his sheet and change nothing"

❌ "cinematic, dynamic, epic" ✅ "a slow push-in on a steady gimbal from seated eye level"

❌ "no head bounce" in the body ✅ "the head travels level"; prohibitions in the tail

❌ "moves like big wild animals" ✅ "moves like a cross-country skier: long strides, slow cadence, level head"

❌ "x 20 %, x 52 %, x 63 %, staggered in depth" ✅ one anchor; "bunched, overlapping, one hand up on the red one's arm"

❌ "has just spotted his bag on the carpet (below the frame)" ✅ "looks down at the floor below the frame; nothing else enters the frame"

❌ "an aerial view of the ridge" ✅ "a wide landscape photograph from a high ridge top"

❌ "the camera pushes in as the three run apart" ✅ "the camera stays on the largest one and makes ONE slow push"

❌ {WHO'S READY FOR SOME <profanity>?} (caps, the word) ✅ {Who's ready for some… company?} → the real line dubbed in post

❌ "packed with dancing guests and the other women" ✅ "the ONLY people in the shot are …; no other guests, no men"

❌ "medium shot, waist up" buried in the event sentence ✅ "Cut 2 (4–7 s): medium on the lead from the front at eye level: …"

❌ "a selfie video, she holds up her iPhone" ✅ "handheld iPhone shot, the front camera at arm's length" — the device as the camera; as a prop it is drawn in her hand

❌ "SO WEAK FOLLICLES WAKE BACK UP" (a shouted line) ✅ "so weak FOLLICLES wake back up" — one word in capitals carries the stress

❌ line 4's prompt reworded while its line changes ✅ the first line's prompt locked; only the quoted words change (`--locked`)

❌ "authentic UGC style, no studio lighting" ✅ "window daylight from camera-left, a warm bulb over the counter, the frame drifting as she talks"

+ ❌ "a hyperrealistic 8K cinematic still of the camp" ✅ "a real, unretouched photograph taken on a full-frame camera, 35 mm at f/2.8, ISO 3200, a few seconds"

+ ❌ H3: "a small fire, no sparks, no embers" ✅ "the flames stay small and low, the air above them clear" (REALISM.md § 3.5)

```

## Failure behavior

- A lint FAIL on slots or the cap blocks the gate; a warning row is answered explicitly.

- A refusal is diagnosed by listing what PASSED first and diffing against what failed — never by three

hypotheses in a row.

- Three rewrites of one prompt without convergence → stop; the operator gets the alternatives as

numbered questions (a different framing, an extension, an insert).

## Cross-references

- `video-refs-continuity` — the references and start image the prompt points at; the ledger the

subjects block is copied from.

- `video-gen-cost-gate` — the venue table, the cost line and the submit.

- The vendor's own optimizer spec (`sd25-pe`) and the ten-model dialect library are cited in

`references/DIALECTS.md`; read them when a venue not listed here appears.

- Deeper context — the upstream projects behind the rules here: `video-production/references/CONTEXT-MAP.md` § Where the deeper context lives.