Human version

render-street-interview

Summary. This is the renderer for the street-interview ad format (goose-studio recipe one-shot-videos/create-street-interview-video). Product guessing retains its visible object and reveal. Conversation supports mic-only, prepared-sample and visible-task script/prompt previews. Choose a coherent situation and earned ad connection before writing words. People are generated; conversation media delivery remains unverified.

---

Agent version

Script first: choose the execution

Read street-script-writing before adapting a brand. The recipe fixes camera, audio, native timing and finishing rules. Its story choices govern premise, participant role, participation reason, visible task, actions, edited opening, hook, supported brand explanation and payoff. Do not force a correct-answer winner into a service conversation or reduce every brand to a routine problem followed by a logo card.

Interaction	Current support	Person image reference	Max participants
product-guess	Existing object renderer; standalone product reference required	forbidden	4
mic-only	Conversation preview; service explanation without a product or device	forbidden	1
product-sample	Conversation preview; prepared sample in a plain cup, no exact package reference	forbidden	1
concept-challenge	Conversation preview; described visible task using non-UI props	forbidden	1

Use [[composes::write-video-ad-script]] with the scoped angle bank, current facts and selected complete commercial street interactions. scripts/prepare_script_context.py requires a brief with offering_type=physical|service|digital and interaction_type=product-guess|mic-only|product-sample|concept-challenge. Inspect the full visual timeline and spoken exchange: setup, participation reason, hook, product role and payoff. Record unseen recruitment or setup as unknown; inference is not observation. Seed snippets are leads. Radio and editorial exchanges cannot fill a street-ad gap. Keep private observations project-scoped and transfer mechanics, not source-brand claims.

Pass participants too: the number of people interviewed on screen, not counting the interviewer (default 4 for product-guess, 1 for conversation). Its route output is binding for any project built on this format, including custom ones. unsupported-route means stop and offer the listed alternatives; it is not permission to go custom. custom is one of those alternatives only when the customer picks it, and a custom video still keeps every route constraint: people in text, no person image references, one take, deep focus, local lettering.

Save a situation brief before dialogue. Write the user's requested count, then map words and actions to ordered shots. An edited participant answer or silent action/reaction may open the ad. cfg.question mirrors the first actual interviewer question, spoken once. Keep both spoken voices and 3–8 shots in 6–15 seconds, at no more than 2.5 spoken words/s; leave time for actions. New configs describe interaction.type, visible_setup, participant_reason and optional props; missing interaction defaults to mic-only. No forced greeting/consent speech, invented use history or instant product efficacy.

conversation currently runs config validation and prompt previews. single_gen.py --yes refuses that mode until a rendered pilot is validated. Natural speech, audio and camera performance remain unverified. The existing product-guess render path is preserved. All conversation subtypes refuse product/scene reference bindings and phone, screen or UI demonstrations. Dry-run success is not a performed sample, challenge or finished video.

Read the bundled model notes before generation. If a required guide cannot be fetched or opened, stop before spending and name it. REFERENCE.md holds the format's historical Critical knowledge entries and the rejected takes behind them. Read it before changing the prompt scaffold or a gate; its older experiments do not override the current recipe or this entry. Use the project take-ledger guidance before reusing a seed. Keep each brand's observed successes and limitations in its own project; a seed is not a quality guarantee.

Opt-in H3 Max generation

Read the H3 adapter guide. Pass --model minimax/h3-max/reference-to-video --resolution 1080P. It keeps prompt lint and the finish manifest, saves request IDs, and recovers downloads without a new order. Historical defaults below remain unchanged.

Run (historical default endpoint)

The paid generation and finishing commands below are for product-guess. Conversation supports brandkit.py validation and single_gen.py dry runs only.

Run everything from the project the video belongs to. Brand-asset paths in the configs (logo, product photo, end-card sting) resolve against that folder, or $STREET_INTERVIEW_ROOT. The run folder is --run <dir> (default projects/street-interview/), with working/ for intermediates and output/ for deliverables.

python scripts/selftest.py                                   # free: the format and the lint hold
python scripts/single_gen.py --brand <slug>                   # dry run: price + the full prompt
python scripts/single_gen.py --brand <slug> --seed <n> --yes  # PAID: one take (~$3.64 at 12 s, 720p)
python scripts/build_episode.py --episode <name>              # free: grade, re-cut, captions, end card
python scripts/check-cut.py --episode <render>.episode.json   # free: the ship gate

• Product-guess brand data: brands/<slug>.json holds the product and its reference photo, the street, the question, the cast and their lines, props, captions, logo and end card. Copy brands/demo-tallgrass-oat.json. No brand appears in format_spec.py.
• An episode (episodes/<name>.json) joins three takes into a ~25-30 s cut. It names the takes, any whole shots to drop (drop_shots, each pair a real shot's start and end), the brand layer and, optionally, brand_layer.end_card_music, a short sting played under the end card.
• Paid calls go through the GooseWorks proxy (scripts/media_proxy.py), never a local key. On a poll timeout, resume with media_proxy.resume_fal(request_id). Never resubmit, since a dropped poll has already been billed.
• A provider policy rejection (likeness of a real person, content policy) makes single_gen.py --yes exit 3: surface the reason, do not retry. The take's input digest covers every input by content (prompt, settings, seed, and the sha256 of the product image and the scene reference), so the unchanged request is refused before anything is uploaded, and a changed image, prompt or seed is sent as a new request.

Prompt length

The final Seedance prompt is built from the project brief and shared shot instructions. BytePlus recommends at most 1,000 English words because lengthy prompts may miss details. This is quality guidance. The current Fal schema declares no maximum prompt length; that does not prove unlimited acceptance.

single_gen.py prints a non-blocking advisory above that guideline. The old 1,200-word refusal is removed: its source-run observation did not prove a precise boundary. Keep the exact approved dialogue and required clauses; do not trim them, reduce the cast or add a paid retry just to meet a count. Missing clauses, invalid inputs and existing spend approval still block generation. The finished-cut gate reports length as advice, without failing on it. Review actual video adherence through the normal gate and full watch/listen pass.

Guarantees

• The prompt carries every format clause; single_gen.py lints it before any spend, and check-cut.py imports the same clause list, so a clause cannot be dropped silently.
• Captions follow the edited picture. An episode derives word timing from its finished assembly. A single-take recut measures words on the original take, aligns spelling to the approved script and moves only kept whole words through the same source-span map as the video and shot boundaries. Dropped speech is not captioned; moved or repeated spans move or repeat their captions. A boundary through a word stops finishing: widen the kept span and rebuild. Never reuse raw-take caption timestamps on an edited video.
• The gate measures the finished file: shot lengths, splices on real cuts, caption timing and safe zone, ambience floor per shot, loudness and true peak, every scripted line audible, and detail and black point. It prints what it can NOT assess (faces, comprehension, how it sounds) on every run. Watch the cut end to end, and do not publish a FAIL.

Cost

A product-guess take is about $3.64 (12 s at 720p, $0.3034/s); a 30 s episode is three takes, about $10.92. Grading, re-cut, captions, looks and every gate are free. Staging prices move, so price the first call of a run and quote from that. Conversation prompt previews send no media call; no conversation media price is verified.

Local finishing

The local finishing scripts use the current Python interpreter and carry --run into child commands. The default grade (--strength 0) needs no colour-reference file. A positive strength requires the real reference.

Set approved colours in brand_layer.palette: accent, text, and background, each #RRGGBB or three RGB integers. Optional brand_layer.fonts keys are black, bold, and regular (paths relative to the project). Without overrides, fonts resolve on macOS, Windows or Linux. End-card rows keep the approved 86px spacing and 66px type. Shrink only an individual line whose visible text cannot fit the safe area; shorten copy if that line still cannot fit. Transparent text-image padding does not set the line spacing.

The subway series bar stays visible through caption gaps. It is also in the caption-free control so the gate measures captions separately from persistent branding.

For a new prompt, use generation.prompt_version: 3 (or --prompt-version 3). It keeps version 2's repairs (no duplicate articles, a top-edge rule for non-can packaging) and keeps the street in focus behind the people: never blurred and never bokeh (see REFERENCE #8). Version 3 is draft until a 720p take measures inside the real-footage detail band; its paid validation take needs its own approval. The bundled demo config stays on version 2 until then. The manifest records the version for the gate. Historical prompts default to version 1, and versions 1 and 2 retain their hashes. Use a new approved seed for a new prompt; do not overwrite an approved take.

Single-take recuts

Check local finishing prerequisites before buying a take: ffmpeg/ffprobe, PIL, NumPy, fonts, and local Whisper with its base model available. The existing episode transcription helper can download an absent Whisper model; prepare it separately before spend. This fix does not call another video or voice model. A reused take may supply previously measured original word times instead of transcribing again.

Keep the original take and generation manifest. Measure its internal cuts, caption source spans and a genuinely speech-free ambience window in brand_layer; these remain source seconds. Remove dead air locally, then add the approved brand layer and end card:

python scripts/recut.py <original-take.mp4> <run>/working/interview-recut.mp4 --brand <slug>
python scripts/build_looks.py --brand <slug> --run <run> --looks <look> \
  --edit-map <run>/working/interview-recut.plan.json
python scripts/check-cut.py --brand <slug> --run <run> --look <look> \
  --edit-map <run>/working/interview-recut.plan.json --take <original-take.mp4> --falsify
# Repeat the same gate without --falsify, then watch and listen to the entire finished file.

recut.py emits the source-span .plan.json beside its edited output. build_looks.py consumes that file, measures source words with the existing free local Whisper helper and saves interview-recut.plan.words.json. To reuse saved timing, pass --word-times <file> to finishing and the gate. Its JSON names the original source and contains ordered [start, end, word] rows. It must come from actual source audio, not estimates from the script. Missing transcription or incompatible timing stops finishing; there is no fallback to the original caption schedule.

Recut masters and caption-free controls have separate -recut- names. The original takes, controls and episode outputs stay intact. Ambience is extracted from the measured quiet window of the original source even if that window was dropped; only that WAV is looped. If the recut already applied its continuous bed, finishing does not mix it a second time. The end card always uses original room tone. The same mapped cuts clamp captions with a half-open end boundary so one person's last words do not appear on the next shot's first frame.

build.py remains an intermediate grade-and-recut helper. To finish its already graded output, use its printed map with build_looks.py --pregraded; do not grade it twice. A changed spoken line, story, paid retry or generation count still needs its existing review and approval. Run the normal finished-file gate and full watch/listen review; a synthetic timing test does not approve faces or the original customer's ad.

Free regression checks:

python -m unittest discover -s tests -p 'test_*.py' -v

Opt-in hold and handover settings

Six generation keys on the brand config, all off by default, so every recorded prompt and its hash are unchanged. They were paid for on Olipop and Graza (REFERENCE.md items 37-42); start a new brand from brands/olipop-ep3-a.json (a can) or brands/graza-ep1-a2.json (a bottle).

key	what it does
handover_grammar	every person's first shot opens on the interviewer passing the product
natural_grammar	the product is held like a drink someone was just handed, not presented
real_grammar	chest-height relaxed hold, one consistent interviewer, nothing printed carried
level_camera	removes the glance down from speaking shots, which is what tips the product
no_signage	no shops, signs or lettering behind the cast (names park-corner landmarks)
can_sealed_grammar	existing flag; holds with the keys above, so the can stays closed

Episode files take extra_head_s (seconds of run-up kept at the head of each take, so the handover is not trimmed as silence) and brand_layer.caption_centre_y. Transcribe every take with word timestamps before assembly: a fast take can repeat, slur or invent a line.

Known limits

• Seedance refuses some photoreal faces (its likeness gate). Faces here come from the prompt, not a reference photo; see REFERENCE.md.
• A multi-take episode can't make three generations be the same corner. A scene-reference still from take A (--scene-ref) couples later takes to its mic, street and light. Check by eye.
• Ambience can differ between takes. The gate's check G fails a shot whose street bed sits within a few dB of the speech; that needs a re-take, not a mix.

Provenance

Ported from goose-studio skills/molecules/create-street-interview-video on 2026-10-02, including the episode-2 v3 fixes (hook caption, the mispronounced line, the silent end card). The port changed only path resolution and routed the paid calls through the proxy. selftest.py passes from an empty folder, and episode 2 v3 rebuilds identically (11 shots, 22.64 s).