skills/faceless-video/references/intake-and-dispatch.md

21.4 KB · Oct 2, 2026 · 00:02 UTC

# Intake and platform dispatch

This reference contains the complete Phase 0 contract. The caller must resolve
the helper-root variables defined by `faceless-video` before following any
cross-skill path below.

## Contents

- Interactive ChatGPT intake and blocking popup rounds
- Style, voice, duration, aspect, and subtitle locks
- Platform-dispatched headless jobs

### Phase 0 — Intake · native ChatGPT path, fixed semantic order

Resolve a structured product rendering alias before checking for missing intake:

- `animation_mode:"fully_animated"` locks the ordinary animated-block pipeline;
- `animation_mode:"scene_based"` locks Picture Story/stills mechanics.

Record the resolved mode and skip any conflicting inference or question. These
aliases are canonical whether they arrive from ChatGPT, a dispatcher, or an eval
harness; never default a supplied `scene_based` value back to Animated.

**INTERACTIVE INTAKE IS A BLOCKING PRECONDITION.** At the start of every ordinary
ChatGPT run, classify the request as either interactive or explicitly hands-off.
Imperatives such as “make,” “create,” “generate,” “produce,” “run,” or “go for it”
do **not** mean hands-off. Lock every parameter explicitly supplied by the user and
collect every missing parameter through the popup rounds below.

Before calling any media-generation, research, narration, assembly, or subtitle tool,
evaluate this exact condition:

If true, render the next missing intake round as an actual ChatGPT popup and
**immediately end the turn**. Do not continue Phase 0, announce production, apply
defaults, or enter Phase 1 in that turn. Resume only after the widget response is
present in conversation state, then repeat the same check. The only tools allowed
while satisfying this precondition are `get_faceless_channel_presets` for the style
picker and `list_voices` for the voice picker. Reaching any generation call with an
incomplete interactive intake is a workflow failure.

Collect only missing values in this order: (1) channel type, (2) style, (3) topic /
duration / aspect / subtitles / thumbnail, (4) pasted-script title when required,
(5) Kids sound mode when applicable, (6) voice. Use ChatGPT's unversioned GenUI
`ask_user_input` widget for every closed-choice intake question. Never substitute
Codex Plan mode's `request_user_input`: it is unavailable in normal ChatGPT answer
turns and causes the intake to collapse into prose. Never call legacy
`ask_user_question` or `AskUserQuestion`, and never write or call
`ask_user_input_v3` unless the current host explicitly exposes a callable tool with
that exact name.

When a callable `ask_user_input` elicitation tool exists, call it. Otherwise emit this
raw GenUI payload directly in the assistant response; ChatGPT renders it as the
interactive widget:

```text
genui{"ask_user_input":{"questions":[
{
"question":"...",
"options":["...","..."],
"type":"single_select",
"free_text_placeholder":"..."
}
]}}
```

The fenced form above is documentation only. During execution, never put the payload
inside a Markdown code fence. Every question requires `question`, at least two short
`options`, `type:"single_select"`, and a specific `free_text_placeholder`. Emit no
more than three questions per widget. Add one short conversational line before the
widget, then output the widget and end the turn; do not add questions or prose after
it. A tool result or raw payload that renders as a popup is the success condition;
printing question text or JSON as ordinary prose is not. If the current host exposes
the exact callable name `ask_user_input_v3`, it may be used for that host; never assume
or fabricate the name when it is absent. If raw GenUI is rejected by the host rather
than rendered, use one concise
normal-chat question as the last-resort fallback and state that the widget surface was
unavailable.

- **Ask ONLY for what's missing.** Parameters the user already stated in their message
(duration, aspect, subtitles, thumbnail, style/preset name, topic, channel type) are LOCKED from
the prompt: do NOT re-ask them and do NOT confirm them — restate the locked ones as a
plain chat STATEMENT (no question mark, no "confirm or correct", no options) and ask
only the gaps. A "here's what I gathered —
all good?" round IS a re-ask and is forbidden. If a round has no missing parameter,
SKIP that round entirely. Only a merely INFERRED value (e.g. type guessed from the
topic) still gets confirmed — as a pre-selected option inside the relevant question,
never as its own extra round.
- **The INTAKE question set is CLOSED.** During intake, the ONLY things this skill may
ask are type, style, topic, duration, aspect, subtitles, thumbnail yes/no, the Kids
sound mode, the voice widget, and one choice among four titles only when a pasted
script has no title of its own.
The three post-generation review questions defined above are the only later
exceptions. NEVER invent extra intake
questions — no cover/thumbnail image, no video title outside that pasted-script case,
no language (write narration in the user's language automatically), no
character/mascot-design question, no "anything else?".
- **Round 1 — channel type (the niche), alone, first:** chips **Explainer
(recommended)** first, then History, then Kids, then **Fairy Tale & Myth**
(retellings of myths/fairy tales/legends — cinematic storybook look,
`${FACELESS_STYLES_DIR}/references/style-cinematic-storybook.md`; also auto-locks when the user says
"fairy tale / myth / legend / folklore", in any language incl. their local
equivalents). Picture Story is NOT a
chip — it auto-locks when the user says "picture story / stills / slideshow story /
storybook video / frame-by-frame" or picks the "Frame by frame" card in Round 2
(narrated stills — `${FACELESS_MODES_DIR}/references/picture-flow.md`). Render this round as one
`ask_user_input` question with exactly those four options and
`free_text_placeholder:"Choose a channel type"`, then end the turn.
- **Round 2 — style, IMMEDIATELY after the type:** call
`get_faceless_channel_presets` to render the unified **Faceless channel presets** catalog.
If the host can capture a structured selection, use it. Otherwise ask the user to
reply with the visible card title; do not manufacture a separate preset list. The
picked card is the LOCKED
style for the run — lock it and move on (CROSS-CHANNEL PRESET RULE below; the
"Frame by frame" card locks Picture Story). Skip the widget when the style is already
decided: a style/preset NAMED in the prompt resolves by name (below), uploaded
style images (≤3) take the custom path, and a long-form-locked run offers its own
LONG-FORM style set instead (below). HOW a style was picked never changes its
mechanics: house styles have no preset id and generate the key from their pinned
FORMULA (stickman uses the generic webcomic formula in `${FACELESS_STYLES_DIR}/references/prompts.md §0`);
Kids styles pin 2–3 canonical ref images. Pass authorized HTTPS references directly
as `medias:[{role:"image",value:url}]`, or use `media_upload_and_confirm` for local
images, then make ONE `seedream_v5_pro` style-key call with the FORMULA. Refs are
style donors only, never final frames.
- **Per-type DEFAULTS + long-form:** when nothing is picked (hands-off, briefs):
Explainer → **Editorial Motion Graphics** (Stickman Cartoon = the second house
direction); History → **Editorial Motion Graphics** (named alternates **Paper
Diorama**, **Mannequin** — `${FACELESS_STYLES_DIR}/references/style-mannequin.md`, clay-render
reenactment figures); Kids → **Studio 3D** (then Pastel Flat 2D / Colorful 3D /
Hand-drawn Ink / Poster Vector — `${FACELESS_STYLES_DIR}/references/kids-styles.md`; Fluffy Toy is a
legacy card in the same catalog); Picture Story → **Flat 2D Papercraft** (then
Stickman / Hand-drawn Ink — one-liners verbatim from
`${FACELESS_MODES_DIR}/references/picture-flow.md`; adjacent asks map to the closest and confirm in one
line); Fairy Tale & Myth → **Cinematic Storybook** (the only style;
`${FACELESS_STYLES_DIR}/references/style-cinematic-storybook.md`, canon-refs → unique seedream key like
the Kids flow). **LONG-FORM AUTO-LOCK: if the request already says ≥10 minutes and/or
"documentary", the LONG-FORM direction is LOCKED from the prompt** — never
offered as an option (offering what the user already chose is a re-ask). A locked
long-form run: duration options become 10/15/20 min (+ Other) if not already
stated; the style round offers the LONG-FORM set — **Watercolor Chronicle
(recommended, first)** / Paper Diorama / Editorial Motion Graphics / Upload —
with descriptions VERBATIM from the style files (`history-longform.md` carries
Watercolor's one-liner); and the mandatory time/cost warning + ERA-MAP flow from
`${FACELESS_MODES_DIR}/references/history-longform.md` apply.
- **ONE catalog — "Faceless channel presets"** (backend consolidation 2026-07-27: the
12 channel cards were moved into the explainer catalog and it was renamed). List it
with **`get_faceless_channel_presets`** and turn a card into a style-reference `media_id`
with **`resolve_faceless_channel_preset`** — both work for every card, house styles
included (verified live: `Fairy Tale & Myth` → `media_id`). There is no separate
faceless catalog and no separate resolver any more.
- **The 12 channel cards, mapped to their style files** (CMS title on the left —
titles differ slightly from the file names, so fuzzy-map, same rules as
user-typed names): Editorial Motion Graphics · Stickman Cartoon (generic §0
formula) · Paper Diorama · Mannequin · Watercolor Chronicle · Studio 3D · Pastel
Flat 2D · Colorful 3D · **Hand Drawn** = Hand-drawn Ink · Poster Vector ·
**Frame by frame** = the Picture Story direction card · **Fairy Tale & Myth** =
the Cinematic Storybook look (`${FACELESS_STYLES_DIR}/references/style-cinematic-storybook.md`).
- **Two cards are DIRECTION cards, not just looks.** "Fairy Tale & Myth" locks
channel type = Fairy Tale & Myth (on-twos `--stepped 12` + mysterious-calm bed).
"Frame by frame" locks channel type = Picture Story with the Flat 2D Papercraft
look (`${FACELESS_MODES_DIR}/references/picture-flow.md` — formula unchanged, only the card name differs).
- **Legacy explainer cards live in the same catalog** (Fluffy Toy, 3D Papercraft,
Mixed Media, Whiteboard Doodle, Pixel Art, Claymotion, Low Poly, Isometric Flat
Vector, 3D Mix, 2D Illustrator, Dynamic Motion Design, Vintage Documentary,
Custom Template). They are valid on any channel (CROSS-CHANNEL PRESET RULE) but
have NO pinned style file: resolve the card to its `media_id`, use it as the look
anchor, and write the locked formula from the card art's visible traits.
- Where a card DOES have a style file, that file wins: the pinned FORMULA +
canonical static.higgsfield.ai refs ARE the preset (self-sufficient — a
`media_id` is optional extra anchoring). Kids styles always keep the
canonical-refs → unique seedream key flow; the FORMULA goes byte-identical into
every prompt regardless of anchor source.
- **CROSS-CHANNEL PRESET RULE — any catalog preset is valid on ANY channel type.**
A preset named on input (user message, brief `preset`, or a card pick) is the
LOCKED style for the run even when it is not among that channel's defaults —
History in Colorful 3D, Kids in Editorial, Explainer in Watercolor Chronicle are
all legal. Never re-ask, never "correct" the choice, never silently substitute the
channel's default. **A style brings ONLY its look, never its home channel's
mechanics:** the style file contributes the FORMULA, canonical refs / anchor
mechanism (Kids styles keep their unique-key flow anywhere), palette lock,
{MOTION} + negatives, and style-inherent laws (e.g. Mannequin's cast/identity
rules). Everything narrative stays with the CHOSEN channel type: cut pattern
(Kids' 4-cut belongs to the Kids CHANNEL — a History run in Studio 3D cuts the
standard 5), narrator↔character interplay, beat grammar, documentary skeleton,
script rules. **Kids-catalog styles carry ONE style-inherent extra: the default
wordless music bed** (`${FACELESS_STYLES_DIR}/references/kids-styles.md §Kids music bed`) — a history or
explainer run in a Kids look still gets the bed, with the MOOD matched to the
channel's tone (playful-light for the look, not babyish). **"Fairy Tale & Myth"
(Cinematic Storybook) carries TWO style-inherent extras anywhere it is used: the
on-twos cadence (`--stepped 12`) and a mysterious-calm music bed**
(`${FACELESS_STYLES_DIR}/references/style-cinematic-storybook.md`).
The ONE exception is **"Frame by frame"** (or any explicit stills/picture
preset), which IS a direction card: it locks Picture Story MECHANICS even when
the channel/brief says history or kids — the channel keeps only its TONE.
Watercolor Chronicle outside long-form is just the watercolor look — no
long-form skeleton, no ERA MAP unless the run is long-form.
- **Pick another preset** → if the user has ALREADY NAMED a preset (in their message
or by choosing a named option), resolve it BY NAME: `get_faceless_channel_presets` → exact
match, else FUZZY match (case/word-order/partial — "fluffy toy" hits "Fluffy Toys",
"3d paper" hits "3D Papercraft") → `resolve_faceless_channel_preset`, confirm in one line.
If nothing plausibly matches, offer the 1–2 closest names in the SAME breath and only
then fall back to the widget — never jump straight to the widget over a typo (asking
twice for the same choice is a bug). Browsing = the Round-2 widget, which serves the
same single catalog the resolver does. (Never enumerate presets as a plain text
question.)
- **Upload ≤3 style images** → collect via the media upload widget (style donors only).
- **Round 3 — compact intake for everything else (only the missing ones; skip when
nothing is missing):** render `ask_user_input` widgets with no more than three
questions each, then continue in the next compact round if necessary. Collect
(a) topic — free text / channel-link /
"randomizer"; (b) duration — 1 / 2 / 3 min (+ Other), **Fairy Tale & Myth offers
2 / 3 min with 2 as the default** (a myth needs room to breathe); **on KIDS runs
(Kids channel or a Kids-catalog style) the duration question ALSO offers "Music
video — a real sung song (1 or 2 min)"** — picking it locks SONG MODE (`${FACELESS_MODES_DIR}/references/kids-song.md`:
the song is generated FIRST, the blocks are staged to it; direct asks like "kids
song / sing-along / music video" lock it from the prompt); (c) frame aspect —
**16:9 (default)** / 9:16 and NOTHING ELSE (the video model supports only these
two — never offer 1:1 or other ratios); (d) subtitles — yes / no (not offered in
SONG MODE); (e) thumbnail — yes / no.
When the topic input is a pasted script, first accept an existing `Title: …`, heading,
or bare leading title line. If none exists, offer exactly four short titles derived only
from that script: blunt claim, question, number, and surprise. A hands-off or unanswered
round takes the first. Lock the title for Phase 8b; never compose a second thumbnail hook.
- **Round 4 — Kids sound mode, only for a Kids run:** skip when SONG MODE is already
locked or `talking_characters` was explicitly supplied. Offer exactly
**Narrator only (recommended)** and **Narrator + talking characters**. The second
sets `talking_characters:true` and loads
`${FACELESS_MODES_DIR}/references/kids-talking-characters.md`; the first sets it
false. An explicit song request wins and forces it false. This selects who speaks,
not the cast design, so never ask a separate character/mascot question.
- **Round 5 — voice (the narrator), LAST — use the voice library (no generated
samples):** SKIP in SONG MODE. Otherwise call `list_voices` so the user can see and
preview current voices, then end the turn and wait for its selectable widget. If the
host surfaces the voice list without selectable UI, render one `ask_user_input`
question in the next turn using the returned voice names as options and
`free_text_placeholder:"Choose a listed narrator voice"`; map the chosen name back to
the exact returned id/type. Lead with a ONE-LINE RECOMMENDATION for the
channel (plain chat text next to the widget): **History → Arthur or Callum; Kids → Remy; Explainer →
Remy, Roxie or Cillian; Fairy Tale & Myth → Arthur or Callum (deep storyteller),
or Remy for gentler tales; Picture Story → match the tone** (kids-warm → Remy,
history-witty → Arthur/Callum, slice-of-life → Roxie/Cillian). The user still picks
freely. Do NOT generate `seed_audio` audition samples (wasted
credits, confuses the flow), do NOT list voices as text options with invented
descriptions. **Record the picked voice's exact `voice_id` + `voice_type` — LOCKED for
the whole video** (rule 17); never re-ask it later. **If the picked voice ERRORS on
first use ("didn't resolve" / not found): do NOT re-open the widget** — look the id up
via `list_voices` to recover the correct `voice_type` (preset vs element is the usual
culprit) and retry with the exact pair; re-ask the user ONLY if the id truly does not
exist in the library. Intonation and MOOD are not chosen
by voice: the script shapes delivery to fit the channel type + topic (line writing +
performed brackets — `${FACELESS_FLOW_DIR}/references/vo_and_captions.md`); the voice itself never changes.
- **Planning locks do not stop; media reviews do.** STYLE LOCK, SCRIPT LOCK, and
long-form OUTLINE LOCK remain notification-only. ASSET LOCK is folded into the
completed IMAGE review. In interactive mode the completed IMAGE, VIDEO, and AUDIO
stages each show one list and stop on the review question from the batch contract.
These are the only production approval stops. Named failures (retry ladder
exhausted, assembly assert, BUDGET_CAP) also stop the run.
- **AUTO / hands-off mode** (a platform flag, or the user says "no approvals /
end-to-end / don't ask" / "pick the voice yourself" / "surprise me"): SKIP the
intake rounds INCLUDING the voice widget — missing parameters take the documented
defaults: type Explainer, aspect 16:9, subtitles off, thumbnail yes, duration 1 min, the type's
default style (Kids → Studio 3D), and `talking_characters:false`. **VOICE IS NOT ASKED in this mode — do NOT open
the voice widget. Auto-pick the CHANNEL'S RECOMMENDED voice by name via
`list_voices` and lock it (record `voice.lock`), then name it in one line:**
- History → **Arthur** (or Callum)
- Kids → **Remy**
- Explainer → **Remy** (or Roxie / Cillian)
- Fairy Tale & Myth → **Arthur** (or Callum); Remy for gentler tales
- Picture Story → by tone: kids-warm → Remy, history-witty → Arthur/Callum,
slice-of-life → Roxie/Cillian
These are preset voices — `list_voices` resolves the name to its `voice_id` +
`voice_type`. Opening the widget in hands-off is a bug. **Stopping for a review
approval in auto mode is a bug — auto mode exists precisely so the user gives no
approvals; nothing in it waits.**
- **"Randomizer" topic path:** WebSearch what is trending NOW (2–3 angles: "trending
topics this week {month year}", "most searched questions this week", plus one vertical
the user cares about if known). A good pick has a **why/how question** at its core, one
**surprising number or reversal**, strong **visual potential**, broad appeal ("Why X is
suddenly everywhere", "The real reason X costs so much"). Avoid breaking tragedies and
active disasters, gossip with no data angle, anything unverifiable by two sources.
Present the pick + one runner-up on the topic step; proceed with the pick unless the
user swaps.

**GATE 0:** you have {type, aspect, subtitles y/n, thumbnail y/n,
`talking_characters` for Kids,
**voice_id+voice_type (the locked
pair — WRITE it to a `voice.lock` file next to the outputs, one line:
`voice_id voice_type`)**, topic, duration, style}. Compute **N = duration_seconds / 10,
rounded half-UP (45s → 5), minimum 3**. If a channel profile was saved earlier (memory /
project notes: style key + voice + type), reuse it and ask only the topic. Style option
descriptions come VERBATIM from the style files' intake one-liners — never improvised,
never naming third-party brands/studios, never promising on-screen text.

### Structured hands-off requests

When the request already supplies a structured brief, treat every supplied value as
locked and skip the matching popup. Supported locks are `topic`, `duration_seconds`,
`aspect`, `channel_type`, `voice_id` + `voice_type`, `preset`, up to three
`style_reference_urls`, `subtitles`, `thumbnail`, `title`, `music_url`, `talking_characters`, and
`channel_dna`. Resolve `animation_mode` before all of them. A supplied style reference
URL is authoritative look input; a simultaneous preset is lineage metadata only.

Use `N = ceil(duration_seconds / 10)` and keep `requested_duration_seconds` immutable for
assembly receipts. Explicit song intent clears `talking_characters`; otherwise the flag
is legal only for Kids. A hands-off request skips media review widgets and delivers the
same confirmed final media URL as an interactive ChatGPT run.

SHA-256: 16c364465b6068d93ec7bec5cd8068ee824a4d6b26bb613be1afe1d8fec6fa98