Video Production

video-production

Core principle: control is everything. Every guard-rail you put in front of the engine is drift you do not pay for later. Lock the shape, the cast, the location and the words before spending, then gate every clip on evidence rather than impression.

This skill covers composing gflow into a finished video. It does not restate the command surface — that is the gflow-cli skill, which owns per-command syntax, flags and single-shot recipes. Load that one for "how do I call t2v", this one for "how do I turn a script into a film that holds together".

How to read the claims in this skill

Every non-obvious statement is tagged. This is one calibrated approach, not the only one. Other people drive this engine with techniques not tracked here; absence from this document means untested, never forbidden.

When NOT to use this skill

A single clip with no continuity requirement, one image, or a pure command-syntax question — use the gflow-cli skill. Editing or grading existing footage — this is generation, not post. Any engine that is not Flow.

Prerequisites

Everything gflow — Python 3.11+, uv, an installed gflow, Playwright Chromium, a signed-in profile, Flow access — belongs to the gflow-cli skill's Prerequisites. Run its checks; do not restate them.

This skill adds three:

1. ffmpeg and ffprobe on PATH, 5.0 or newer — ffmpeg -version. The assembly step uses -fps_mode, which does not exist before 5.0.
2. An ffmpeg carrying libass and libfreetype — the same banner lists enabled libraries. Burned subtitles (subtitles) and title cards (drawtext) are absent from minimal or "essentials" builds. Usual Windows trip.
3. faster-whisper, only if you want the transcript gate — python -c "import faster_whisper". Nothing in gflow installs it. First run pulls base.en, ~75 MB, once. Without it you lose the word-hit check and keep every other gate.

clip_qa.py, beside this file, needs only ffmpeg, ffprobe and the standard library — so the fluidity, lip-sync and A/V-drift gates still run where the transcript gate cannot.

Step 1 — intake, before anything else

The plan changes completely with the answers. Ask, or read them from the brief; do not assume.

Veo always generates audio [CONSTRAINT]. There is no silent mode and omitting sound from the prompt does not produce silence. A "silent" montage returns invented room tone under every shot. If the audio is unwanted, strip it at assembly (-an) rather than hoping for a quiet clip.

Step 2 — check the host, or the whole plan is dead

Load https://labs.google/fx/tools/flow/project/<id> in the profile's Chrome and read the final URL.

Lane B grows as forms are ported, so confirm the row rather than trusting it: the maintained list is CONFIGURATION § GFLOW_CLI_FLOW_HOST and the #639 entry in KNOWN_ISSUES. Exit 36 on a form the table says is ported is a regression worth filing, not the environment.

An unported command on lane B exits 36, non-retryable, with a message naming the migration. A RecaptchaError instead means one of two other things: on gflow ≤ 0.68.0 the migration guard ran after the reCAPTCHA mint on the image path, so image commands died as exit 1 there (gflow-cli#673, fixed); on any build, a RecaptchaError right after gflow auth login is the cookie harvest keyed on the old host (gflow-cli#644). Either way the first move is the same: read the final host URL before anything else. Do not plan entities or plates before this check.

There is a gflow project create [CONSTRAINT]. gflow project create --name <piece> --json > out.json, redirected to a file, never piped through head, which truncates the process before the JSON prints. Do not mint a project by burning a placeholder generation, and do not scrape the id from a browser URL.

Step 3 — the production shape

Pick one; each has a worked command chain in composition.md.

Step 4 — lock the assets before spending

4a. The identity resolution ladder — deterministic, and you record the rung

A person in a shot is anchored by exactly one of these. Take the highest rung the host and path allow, and write down which rung you used and what blocked the one above. This is not a preference order, it is the production record: a film whose log does not say how each shot was anchored cannot be debugged when a face drifts.

Rung 4 does not hold a person. Each generation invents someone new who merely matches the description. Two shots on rung 4 are two different actors. Never plan a multi-shot piece on it and never let it be the silent default.

Rung 3 carries contamination. v1's opening take came back with 72 px letterbox bars and its plate carried those bars into every shot that referenced it. A rung-2 plate is generated clean by the character editor; a rung-3 plate is only as clean as the take it was cut from.

Descending a rung is a decision that gets written down, in production.json next to the shot, naming the blocker. "I used an image" is not a record; "rung 2, because --reference-entity exits 36 on this host (#639)" is.

Written from a run that skipped its own ladder. On 2026-09-07 a five-shot film created two real character entities and then anchored every shot on rung 2, because gflow refuses entity references on the migrated host. The refusal was never questioned — and a $0 probe the same day showed the host takes an entity mention perfectly well (data-reference-type="entity" with the real entity_id); it is gflow that has not ported the gesture. The run went a rung lower than it had to and recorded it only in passing.

4b. Entity beats media when a name matches both [CONSTRAINT]

Flow's @ picker offers character entities and media assets in one list, and a query matching both can resolve to either — measured 2026-09-07: the same @Kael returned reference_type="entity" on one gesture and reference_type="media" (a JPEG that happened to be named after him) on another. Flow does not rank them.

So you rank them. When a name matches a character and a file, the character wins, and if you meant the file, reference it by path (--ref <path>) rather than by name. A shot that silently binds a still image where you asked for a character produces footage that looks right and drifts on the next cut.

4b-bis. A cast with VOICES — the whole recipe, end to end

A character entity is the only thing that carries a voice. A plate does not: it carries face and wardrobe as pixels, and the engine invents a new voice for every clip. Measured 2026-09-07 on one character across three plate-bound takes: 88 Hz, 103 Hz, 118 Hz — three different actors — against an engine noise floor of 4.3 Hz on an identical prompt repeated three times. No re-shoot fixes that, because nothing in a plate was ever carrying the voice.

So if anyone speaks more than once, you need rung 1. Here is the whole path.

1 — pick the voice before you create anyone. gflow character voices lists 29 presets. Each has a public sample you can actually listen to, so audition rather than trust the descriptor — five of the 29 descriptors disagree with the measured pitch of their own sample:

gflow character voices --json
# every voice has: https://gstatic.com/aitestkitchen/voices/samples/<Name>.wav

2 — create the character with the voice bound. Face, body and voice in one call:

gflow character create --project "$PROJECT" \
  --name    "<UniqueName>" \
  --face-prompt "a man, <face description with a GENDER WORD>" \
  --body-prompt "<outfit; no print, no logo>" \
  --voice   "<VoiceName>" \
  --personality "<how they behave>" \
  --json > cast_<name>.json

Three traps, each measured:

• The name must collide with nothing in the media library. Flow's @ picker searches characters and media together and does not rank them (4b), so an uploaded kael_ref.jpg wins the query @Kael and your character becomes unreachable by name. Never name a plate after a character; plate_a.png, not <name>_ref.jpg.
• A face prompt with no gender word gets a gender chosen for it. This fires on the prose canon too, not just on character create — a two-hander whose unbound actor is described without one renders the wrong person.
• Keep the entity_id yourself. Read it from --json at creation. Do not plan on recovering it from the catalog afterwards.

3 — attach the character to every shot they appear in.

gflow video t2v "<prompt>" --project "$PROJECT" \
  --reference-entity      "<entity_id>" \
  --reference-entity-name "<UniqueName>" \
  --aspect 16:9 --duration 8

One face-bearing reference per generation [CONSTRAINT] — so in a two-hander, bind the person who speaks and carry the other in prose. Getting that backwards is what produced the 88 Hz stranger above: the beat bound the silent actor, so the speaker fell to rung 4.

4 — expect the submit to outlive your patience, and do not read a timeout as a failure. Flow allows five concurrent generations and throttles per-minute throughput after heavy daily use, so a queued job routinely outlives gflow's submit-reply budget. An entity-bound run can exit 9 TransportTimeoutError while the video is rendering normally — the job is in Flow's queue and will finish. Check the project before you re-submit, or you will double-spend on a generation you already have. (gflow-cli #723, #741.)

5 — verify the voice actually landed, relatively. Never assert an absolute band: a character legitimately speaks differently in an action beat than in a quiet one, and the pitch follows the performance. The two sound comparisons are:

• the same character across comparably-staged shots — the medians should sit close
• a bound take against that voice's own public sample — the same neighbourhood

And stage every dialogue beat in still air. An energy gate is not a voicing gate: on this production's own wordless clips the detector reported a confident 145 Hz and 280 Hz with no speech present at all, and periodicity did not separate them either. A beat shot in wind cannot be checked.

4c. A costume is part of the identity, so a costume change is a NEW entity [CONSTRAINT]

A Flow character entity bundles face and wardrobe — the body reference fixes the outfit. There is no wardrobe axis inside one entity.

So a character who changes clothes is two entities sharing a face prompt, with different body prompts, named for the costume state:

Kael_ridge   face_prompt=<the canonical face>  body_prompt=<dust-brown canvas jacket, sand scarf>
Kael_coat    face_prompt=<the same canonical face verbatim>  body_prompt=<heavy oiled coat, hood down>

The face prompt must be byte-identical across costume states; only the body prompt moves. Then every scene names the costume-state entity, not the character, and continuity becomes a lookup instead of a hope.

This is not what movie.toml does today [CONSTRAINT]. Character.variants is a Mapping[str, str] and resolve_variant() appends a text delta to the prose appearance (composition.py:67-80), while the runner creates exactly one entity per character name (cli_movie.py:589-597). On an identity = "entity" character a variant therefore changes the words while the entity's body plate keeps the original outfit, and the two argue inside one generation. Until that is fixed, express costume states as separate entries in the manifest's characters array, each with identity = "entity" and a shared face prompt — not as variants. See MOVIE.md for the TOML.

Full method in consistency.md. The rest of the short form:

• People are Flow CHARACTER entities, attached with --reference-entity or @Name. Identity.
• Locations and props have no entity type [CONSTRAINT] — they are images attached per shot with --ref. Look.
• One face-bearing reference per generation [CONSTRAINT]. A second entity, or an entity plus a portrait, returns HTTP 400 reported as a wire-format error. Carry other people as role nouns in prose.
• Multi-angle locations must be chained, not generated in parallel [CALIBRATED] — anchor angle by t2i, every other angle by i2i --ref <anchor>. Independent calls from the same paragraph produce different rooms.
• The reference count picks the model, before quality does [CONSTRAINT]. Caps are per model and the entity counts against the same pool: omni-flash 7, veo-lite / veo-fast / veo-lite-lp 3, veo-quality 0 — it accepts no references at all. So a shot carrying any --ref or entity cannot use veo-quality however much you want its quality. For a single generation needing references and quality together, that makes omni-flash the pick; a veo-lite variant when 3 refs is enough. video chain is the exception [CONSTRAINT] — it refuses omni-flash outright (its i2v is wire-verified for single generations only) and exits with a model/mode incompatibility, so chained links take a Veo 3.1 model and its cap of 3. Full table, image models included: consistency.md.
• Attaching by media UUID buys asset identity, never scene coherence.