ugc-ads: handheld, phone-shot, deliberately unpolished

Turns a product and a situation into a clip that reads as something a real customer filmed on a phone — an unboxing, a first impression, a use-it-once review. Amateur texture is the deliverable, not a compromise.

This skill is a thin, scenario-specific layer over ofox-video-core. It owns the UGC prompt craft, the brief, the defaults and the pre-generation cost estimate; ofox-video-core owns talking to the Ofox API correctly and safely (the OFOX_API_KEY handling, the no-resubmit rule, error-code mapping, download/verification, and reporting the downloaded file's absolute VIDEO_PATH). Read that skill's safety contract before using this one — it is not restated here.

The shared prompt craft — the vendor's formula, timestamps, transitions, camera vocabulary, consistency locks and negative lists — is in ../ofox-video-core/references/prompt-structure.md; the pre-prompt question rules in ../ofox-video-core/references/creative-brief.md; the spend rule in ../ofox-video-core/references/approval-gate.md. This file carries only what is specific to UGC.

The inversion — read this before writing anything

Every other video scenario in this repo is aimed at polish, and so is most of the shared prompt guidance. Its style vocabulary is full of cinematic anchors, studio keys and hero frames; the ad skill's own archetypes are Luxury, Playful and Minimalist-tech; its pacing shape is hook → showcase → slow-motion climax → hero freeze. All of that is correct for those scenarios and wrong here.

The default, everywhere else	Here
Cinematic colour grade, warm rim light	Whatever the room's own light does, including bad white balance
A studio key, or golden hour	One practical source — a window, a ceiling lamp, a phone's own screen
Smooth, motivated camera moves	Handheld drift, a small re-frame, an arm that gets tired
Perfect framing, subject centred	The product half out of frame until a hand corrects it
A slow-motion climax and a hero freeze	A flat action chain with no climax, ending "naturally, not promotionally"
Skin that reads as clean	Skin that reads as skin
A seamless backdrop	An actual kitchen counter with other people's objects on it
AVOID is a backstop	AVOID is half the brief

The failure mode is silent, and nothing in the tooling catches it. A clip that comes back beautifully graded, stabilised and lit has not errored; it has rendered exactly what it was asked for. It is just no longer UGC, and the only place that gets noticed is in front of the audience it was meant to fool. So the check has to happen in the prompt, before the money moves.

The three habits that leak in

1. Style words carried from another scenario. cinematic, 8K, professional, studio, premium, glossy, film grain, shallow depth of field, rim light, colour grade. Every one of them is a defect here. film grain is the subtle one: it is an anti-plastic anchor in the shared guidance and it is still film, which a phone is not.
2. The ad beat structure. The moment a timeline grows a climax and a hero frame, the clip is an ad with handheld camerawork bolted on. UGC-style prompts in the gallery are a flat chain of ordinary actions (cases 24–27).
3. A camera that behaves. "Handheld" written once in a style line, and then every beat describing a clean push-in, produces a clean push-in.

Before submitting, re-read your own prompt for the words in habit 1. That takes seconds and is the single highest-value check in this skill.

Where the anti-polish vocabulary lives — load it, don't rewrite it

It already exists, derived from real gallery cases, and it is deliberately not copied into this file: one copy gets corrected and the other keeps the old wording. In ../ofox-video-core/references/prompt-structure.md:

What you need	Section, in that file
The AVOID list itself	Consistency locks and the negative list → Negative list — the frequent items → the UGC anti-polish row (cases 25, 40). It covers colour grading, beauty filters and skin smoothing, dramatic slow motion, music, perfect lighting, and the CG look — quote that row, don't paraphrase it
Phone-capture style anchors	Style anchor — the texture layer → the Consumer capture row (cases 20, 25, 40)
Handheld camera phrasing	Camera language → Camera movement → the Handheld / breathing row
The viewpoint — never the phone's location	Camera language → Camera position / angle, and the Phone POV row under Camera movement. ⚠️ Read "The capture is part of the fiction" below first: those rows quote gallery prompts that place a phone somewhere in the room (static phone propped on bathroom sink), and measured here, that phrasing renders a phone in the shot. Take the viewpoint from them, not the prop
Autofocus and exposure misbehaviour	Camera language → Focus → the Autofocus hunting (UGC texture) row
The beat shape	Pacing → the closing note: UGC clips are a flat action chain with no climax
Why a prohibition alone is not enough	The plastic look is designed out, not forbidden — the four faults that produce the AI-render look and the positive sentence that crowds each one out

That last row is the one to actually read. Its argument is the whole reason this skill needs a positive half: a prohibition removes a class of thing without supplying what stands in its place, and "not cinematic" leaves the model to pick its default, which is the cinematic one. The scenario-specific positive form is below, under "What actually reads as UGC".

What has been tested, and what has not

This scenario has two paid runs, both bytedance/seedance-2.5, 8 seconds, 480p, 9:16, 88 cents each, and both with the prompt filled in from this file's own template by an agent working from this file and nothing else — so what they tested was the whole chain (skill → agent → command → artifact) rather than the API. Both were judged on frames extracted from the delivered clip against criteria fixed before the run, not on STATUS completed.

Run	Shape
2ecbedec (2026-09-15)	text-to-video, no attached photo, hands only
ede33e6d (2026-09-17)	a real product photo attached as the first frame, and the corrected CAPTURE wording

Run 1, 2ecbedec

What was checked	Result
Does the anti-polish steering actually hold? The criterion, written before the run: the clip must not look like the ad seedance-ad-creative would produce — no centred hero framing, no dramatic lighting, no shallow-focus beauty shot	It held. Off-centre, loosely composed framing; a single practical window light with real falloff; no grade
Hands only, no invented face	held — hands throughout, no face anywhere in the clip
The product	rendered as described, no drift
The capture line	failed — the line said where the phone was, and the model put a phone in the room. See the ⚠️ under "The capture is part of the fiction"

Run 2, ede33e6d — the corrected CAPTURE line, and three new findings

The correction made after run 1 had never itself been generated. It has now.

What was checked	Result
The corrected CAPTURE wording — does converting the device from an object in the room into a viewpoint keep it out of the shot?	✅ It held. Six sampled frames across the clip contain no phone, no camera, no tripod and no lit screen. One candidate object, enlarged, is an out-of-focus wooden desk item; the left-hand shape is a lamp arm with a blue practical behind it, which is the SCENE line's own "a warm lamp on the left"
The product, against the attached photo	held — wordmark, copper cap and gunmetal brushed finish all match the photo
Hands only, no invented face	held for the whole clip
The beat chain	ran in the written order
❌ The opening	a flash frame and a hard cut — see below
⚠️ Anti-polish	weaker than run 1, and this run cannot say why — see below

❌ The opening flash cut. The attached photo was a white studio shot; the SCENE was written as a desk at night. Background brightness at the top-left: 250 at 0.00 / 0.04 / 0.08s, 33 from 0.12s onward. So the clip opens on about a tenth of a second of the product photo and then hard-cuts into the scene. This is a general mechanism, measured across three runs and written up once in ../ofox-video-core/references/prompt-structure.md → "If the attached frame's background disagrees with your SCENE". It hurts this scenario more than any other, because an unbroken, un-staged single take is the entire product and the cut lands on frame one. Fixes, cheapest first: write the SCENE as the place the photo was actually taken; or use a photo already shot in the target setting; or trim the first 0.15s before delivery.

⚠️ Anti-polish was weaker, and two variables changed at once. Run 1's pre-registered criterion was that the clip must not look like seedance-ad-creative's output. Run 2 leans toward polish: the bottle sits near centre, the background is shallow-focus, the key light is warm with falloff, and the steam is flattering. But run 2 changed both the photo and the CAPTURE wording, so the cause is not attributable. A plausible hypothesis — attaching a clean studio product photo pulls the whole clip toward polish — is worth testing and is not a finding. The single-variable run that would settle it: the same prompt, the same corrected CAPTURE line, with and without the photo. Until then, when a photo is attached, look specifically for polish creeping in and treat a re-roll without the photo as a live option.

Read those as two clips, not a guarantee. One model, one duration, one tier, one product, hands only, no dialogue. The anti-polish inversion is the claim run 1 genuinely supports, because it was designed to falsify it and did not — and run 2 is the reason that claim is now stated for the text-only path rather than for the skill as a whole.

What it is built on besides that run:

• Gallery practice. The collected creator-style prompts (cases 23–27, plus 30–33 for texture phrasing) are prompts somebody kept, with a finished video attached — on platforms and with parameters that are mostly unrecorded. That is evidence about what a good UGC prompt looks like, not about what this API does with one.
• Mechanism facts measured in this repo, which apply here because they are about the API rather than about the genre: an attached photoreal-person frame is refused at submission; asking the model for music has failed output moderation on copyright; a prohibition on a class of object holds better than one on a tempo; a camera move written only as a verb tends not to happen. Each is cited where it is used below.

Practical consequence: two clips is still thin evidence. Anything outside those runs' parameters — a person on camera, a spoken line, a longer clip, a different model, a higher tier — is still an experiment and should be priced and described as one. The attached-photo path is no longer unrun, but its one run carried two defects (the flash cut, the polish drift), so it is the path to draft cheaply rather than the safe one. Draft small before anyone pays for a deliverable — see "Several takes, and the cheap vertical route".

Where the core skill lives

Resolve once, before the first call:

for d in ../ofox-video-core \
         ../ofoxai-skills-ofox-video-core \
         ~/.agents/skills/ofox-video-core \
         ~/.agents/skills/ofoxai-skills-ofox-video-core \
         ~/.claude/skills/ofox-video-core; do
  [ -f "$d/references/ofox-video.sh" ] && echo "$d" && break
done

Examples below are written as ../ofox-video-core/... (the skills.sh / ClawHub / npx ofox-skills layout, where a skill's directory is named after the skill). If the probe found a different directory — LobeHub unpacks each skill as ofoxai-skills-<name>, so the sibling there is ofoxai-skills-ofox-video-core — substitute it, in the ofox-video.sh commands and in the references/*.md links alike.

That ../ is relative to this skill's own directory, which is also where the probe has to run. From anywhere else nothing resolves — use the absolute path the probe printed (candidates 3–5 are absolute already), or, in a clone of this repo, skills/ofox-video-core/references/ofox-video.sh from the repo root.

Nothing found → the core skill isn't installed; see "If the script isn't found".

This skill needs ofox-video-core 2.0.0 or newer. From that version the billable subcommands refuse to run without --approved, and every real-run command below passes it. An older core does not know the flag and stops with unknown option '--approved' before any request — nothing is submitted and nothing is billed, so the fix is to update the core, never to drop the flag.

Before generating: the availability check

Run this once per session (not on every request):

bash ../ofox-video-core/references/ofox-video.sh check

If it fails, follow ofox-video-core's guidance (install curl/jq, or get an OFOX_API_KEY at https://app.ofox.ai) — don't dead-end the conversation, and don't re-run this check on every subsequent request once it has passed.

check reports whether the key is present, not whether it is valid, and makes no network call. It reads an ordinary environment variable in the shell that runs the script — nothing is loaded from a .env file, so a project that keeps the key in one has to source it into the same shell first.

A failing check is not a stop sign: models, providers and generate --dry-run all run without a key, so the brief, the prompt and the price can all be done first. See "Pricing a job with no API key".

Before writing the prompt: the brief

The shared rules — the three tiers, one round of at most four questions, the shape of a question, the "Let the AI decide" discipline, the order with the approval gate, the fallback without AskUserQuestion — are in ../ofox-video-core/references/creative-brief.md. This section adds only this scenario's question set.

Tier	UGC axes
must-ask	is there a product photo? It changes the whole route (see "People, products and what the API refuses"), and a photo that does not exist cannot be invented
ask-if-open	who is on camera, which action chain the clip follows, whether anyone speaks
never-ask	resolution, model, provider, duration when the user gave one, aspect ratio (vertical is this skill's default and saying so in the recap is enough). Two resolutions are two rows in the cost table, not a question

Five axes, four question slots. If the round is full, ask in this order: Photo, On camera, Beat, Talk — and let anything left fall to a default named in the recap.

The question set

#	Tier	header	Question	Options (1 = recommended, last = delegation)	Ask when
1	must-ask	Photo	Do you have a photo of the actual product?	Yes — a local path (recommended): the real label and shape are locked into the opening frame, and everything the gallery collected that had to match a real SKU used one. The photo must have no person in it — not a hand, not a foot — or the job is refused at submission. / No — describe it in text: fine for a generic item; on a real product the label may drift. No AI option — a photo is a fact, not a taste.	No image attached and the user did not say there isn't one.
2	ask-if-open	On camera	Who is in frame?	Hands only (recommended): the most common real unboxing shape, and it keeps faces out of a generated clip entirely — fewer anatomy risks, no identity to hold. / A creator, filming themselves: an arm's-length or a fixed low viewpoint — written as a viewpoint, never as a phone standing somewhere in the room (see "The capture is part of the fiction"); they are generated fresh from text, so they cannot be the same person in a second clip. / A creator, filmed by someone else: a friend holding the phone; allows a wider frame. / Let the AI decide — which means hands only.	The request doesn't say.
3	ask-if-open	Beat	What happens, in order?	Unboxing (recommended): parcel → cut/open → lift the item out → turn it over → first use. / First impression: already unpacked, one use, one reaction. / Before / after: the state that motivated the purchase, then the result. / Haul: several items shown in one sitting. / Let the AI decide.	The request names a product but not an action.
4	ask-if-open	Talk	Does anyone speak?	No — ambient sound only (recommended): room tone and the sounds the actions make; captions get added in an editor afterwards. / One line at the end: a single spontaneous-sounding sentence. / Talking through it: continuous commentary — budget the words per second, see "If the creator speaks". / Let the AI decide.	The request doesn't say, and the duration is long enough for a line.

Not asked: aspect ratio (9:16 unless the user says otherwise), resolution and duration (rows in the cost table), the model, the provider, the audio flag.

Skip rows specific to UGC

On top of the generic rows in creative-brief.md:

Input says…	Axis	Value
a photo is attached, or a path is given	Photo	have one — and check it for hands, feet and faces before attaching
"unboxing", "just arrived", "opening the parcel"	Beat	unboxing chain
"review", "honest opinion", "what I think"	Beat + Talk	first impression, with a line
"no face", "just my hands", "POV"	On camera	hands only
"TikTok", "Reels", "Shorts", "vertical"	Aspect	9:16, already the default
"for the landing page", "for the website"	—	that is probably seedance-ad-creative; check before assuming UGC
"polished", "cinematic", "premium", "hero shot"	—	the user is asking for the opposite of this skill — see "When NOT to use"

From answers to prompt — traceability

Answer	Lands in
Photo: yes, path	--frame-first-image PATH (product alone), plus the anchor sentence at the top of the prompt; no --aspect-ratio flag — ⚠️ so crop the photo to 9:16 first, see below
Photo: no	a text PRODUCT line, and --aspect-ratio 9:16 becomes effective
On camera	the CAPTURE line and what each beat says is in frame
Beat	the timeline's action chain
Talk	the AUDIO line, and a quoted line with a delivery note
Duration	--duration and the timestamps

⚠️ Those first two rows contradict each other on a landscape photo, and the table used to hide it. This file says 9:16 is the default and never asks about it. But with a frame attached, bytedance/seedance-2.5 forces aspect_ratio: adaptive and the clip takes the photo's shape — so attaching an ordinary landscape product shot, which is most product shots, delivers a landscape clip. The vertical premise is cancelled by the very route the table recommends, silently, and nothing errors.

So crop the photo to the target ratio before attaching it. That is a free, local step, and it is what the measured run did: the product photo was cropped to 9:16 (529x941) first, and the clip came back 480x854. Found by reading the two rows against each other rather than by spending anything; the crop is the fix, and putting the crop in the recap is how the user finds out it happened.

The recap for this scenario

Brief:
- Product: the ceramic pour-over kettle — photo at /Users/me/shots/kettle.jpg (given)
- On camera: hands only (you chose)
- Beat: unboxing — parcel, open, lift out, turn, first pour (AI's pick)
- Talk: no dialogue, ambient sound only (AI's pick)
- Vertical 9:16, 10s, 480p draft — with a 720p row below
- Photo prep: your kettle.jpg is landscape, so I will crop it to 9:16 before
  attaching it — free and local. With a frame attached the clip takes the
  photo's shape, so an uncropped landscape photo would produce a landscape clip
- Note: the attached-photo route has one run behind it (8s/480p). It came back
  with the product right and two problems: a ~0.1s flash of the photo then a
  hard cut at the open, because the photo's background was not the scene; and
  a more polished look than the text-only run. The draft is still the
  experiment

Then the full prompt, then the cost table, all in one message — the order and the rule for a change made at the gate are in creative-brief.md's Order, with the approval gate.

What actually reads as UGC

Four positive levers. They are the substance of the clip; the AVOID list is what stops the model reaching for its defaults underneath them.

1. The capture is part of the fiction — the phone is not

No other scenario in this repo has a CAPTURE line, and it is the single most load-bearing sentence here. It carries three things nothing else supplies: where the frame sits, how imperfectly it is composed, and how the lens misbehaves — which together are most of what makes a clip read as phone-shot. Write all three as properties of the shot: a fixed viewpoint at table height, a little low and a few degrees off level, so the near table edge cuts across the bottom of the frame; the focus hunts once when the hands move in; the exposure lifts when the white box fills the frame. The phrasing for the flaws is in the shared file's Consumer capture and Autofocus hunting (UGC texture) rows; use it as written.

⚠️ Never say where the phone is. It will be in the shot. Measured on this skill's one paid run, job 2ecbedec (8s, 480p, 88 cents): the capture line read "the phone is propped against a biscuit tin at the back of the desk, slightly too low and a little off-square" — and the delivered clip has a phone sitting in the box, visible in every frame, with a small glowing screen, which also broke the same prompt's own no legible text clause. The model has no concept of an off-screen camera. A noun given a position in the room is set dressing, and set dressing is rendered. propped against a mug, on a tripod, wedged on a shelf, left recording on a bench are all the same trap, and the gallery's own UGC cases are written exactly that way (static phone propped on bathroom sink), so the phrasing reads as completely natural right up until you watch the clip.

The split that works. A phone named as how the image was made is a format anchor and stays: real phone capture texture, shot on a phone, selfie viewpoint. A phone named as a thing at a place is scene content and has to go. Converting one is mechanical — the phone is propped against a mug at the back of the table, slightly too low becomes a fixed viewpoint from the back of the table, at mug height and a little too low. Identical framing, no object. Then put a phone, a camera, a tripod or a lit screen visible anywhere in the shot in AVOID as a backstop, phrased as objects, which is the form this repo has measured holding (lever 4).

✅ And the fix has now been generated, not just reasoned. Job ede33e6d (2026-09-17, 8s, 480p, 88 cents) ran the converted form — viewpoint wording, no object, plus the AVOID backstop — and six sampled frames contain no phone, camera, tripod or lit screen. The one candidate object, enlarged, is an out-of-focus wooden desk item. So this is a rare pair: the failing construction and its replacement have each been run once, on the same model and tier. That is still one run each, and the conversion is a rule about wording, not a guarantee about every room you describe.

Nothing catches this before delivery, which is why it is worth this much space. That prompt passed --dry-run, passed --print-payload, and passed this skill's own re-read-your-prompt-for-polish-words check. None of them can see a rendered object. Only the extracted frames could.

⚠️ A frame that never re-frames is a tripod. Put one small correction in the timeline — the view tilts up to catch the product, the framing drifts and re-settles — as its own beat with its own stamp, and write it as something the frame does, not as something a hand does to a phone. A camera behaviour written only as an adjective in a style line does not reliably render; the shared file's A camera move needs its waypoint frames, not just a verb is the measured version of that, and it applies to sloppiness exactly as it applies to an orbit.

2. One practical light, and let it be uneven

Name the source and where it is: a window to the left with the curtain half drawn, one warm ceiling bulb, the blue-white of an overhead kitchen light. Then say what it does badly — the near side of the product bright and the far side falling into shadow, a colour cast on a white wall, a blown highlight where the light hits gloss. The shared file's The plastic look is designed out, not forbidden has the general form (name sources, let them fall off); what UGC adds is that the unevenness is not a compromise to be minimised, it is the evidence that nobody set up a light.

3. Framing that is a little wrong

The product entering frame off-centre and being corrected. The item's top cropped for a second. A hand crossing the lens. Written as beats, not as a style word. Two cautions:

• Not so wrong that the product cannot be seen. The clip still has a job. The shared Camera language note on readability outranking the effect is the right tie-breaker.
• Imperfection is an event, so it needs a stamp, for the reason in lever 1.

4. A real room, not a set — with one tension to manage

A UGC set is supposed to be cluttered: the counter has other objects on it, the table has a used cup, there is a bag on a chair. That directly conflicts with the one measured rule this repo has about invented text. The shared file's Unwanted text is designed out of the set, not forbidden in the list found, across four jobs, that a set with no lettered surface produces no invented text and a set full of them produces it regardless of how strongly the AVOID list forbids it — and a real room is full of them. Book spines, packaging, a cereal box, a laptop.

The resolution is not to empty the room, which would cost the scenario its whole point. It is:

• keep the lettered carriers out of the frame or out of focus — the shallow background a phone gives you at close range is doing that work already, and it is the one bit of optical softness that is UGC-native rather than cinematic;
• name the carriers you do not want as objects in the AVOID list, not as "text" — the same file records a bare "no text" failing on a busy set and an object-by-object exclusion holding on a desk;
• accept the product's own packaging as text you want, and lock it with a photo rather than describing it — see below.

People, products and what the API refuses

Two measured ofox-video-core facts decide the route, and together they are less restrictive than they first look:

• A photoreal person in an attached frame is refused at submission on bytedance/seedance-2.5 (input_moderation_failed, nothing billed). It is a rule about the picture you attach. --real-person true lifts it — measured 2026-09-16 — but as an authorisation claim about a likeness the user has the right to use, not as a way past the check, so it is a question for the user and never a flag an agent adds: ../ofox-video-core/references/api-params.md → "--real-person true lifts that refusal on 2.5".
• A photoreal person generated from the prompt text is fine. Several 20–30 second text-to-video jobs built entirely around people have completed in this repo.

So a UGC clip with a person in it is not blocked; it splits:

Half	How
The product must be exactly right (real label, real shape)	attach it as --frame-first-image, with no person and no part of one anywhere in that image — hands, feet and socks included
A person must appear	write them into the timeline as text, entering after the product is established

That split is measured end to end on the ad scenario's job ac927785-92ef-4e28-97b9-ff8172ec5554 (a product-only first frame, a person entering at 4s, the product's appearance unchanged at 19.6s). It is an ad, not a UGC clip, so read it as the route being proven, not the genre.

Hands are the UGC-native answer and they are also the safest one. The same job's lesson about anatomy — escalate how much of a body is in frame, keep the full figure to one distant shot, never make a hand or a face the focal point — matters more here than there, because a phone-close unboxing puts hands in the middle of the frame by definition. Two things that help: keep the hand's job simple (one action per beat, holding or turning rather than gesturing), and keep the product, not the hand, as what the beat is about. A written size limit on a limb is not a defence — one measured job wrote only the fingertips and the first knuckle ever visible twice, once in the shot and once in AVOID, and the render showed most of the finger and the hand to the knuckles anyway. What kept that shot anatomically clean was the macro framing and one simple action, not the limit.

No person survives across jobs. They are generated fresh each time, so a second clip is a different creator. Only the product can be locked.

The prompt template

Vocabulary is not repeated here — timestamp formats and segment lengths, transition phrasing, camera and focus terms, the negative-list rows, asset role sentences are all in ../ofox-video-core/references/prompt-structure.md. What follows is the UGC shape.

If several images should supply appearance without locking the opening frame, use the shared Identity-reference recipe — the authoritative copy. Do not reproduce its payload or limits here; this file owns only the UGC role each image plays.

Slots in <angle brackets>; optional lines in [square brackets].

STYLE: <consumer-capture anchor, from the shared file's Consumer capture row — real phone capture texture, sensor noise, compression artefacts, autofocus hunting, exposure shifts>. Vertical. Nothing is graded, lit or staged.
CAPTURE: <the viewpoint, as a property of the shot and never as a phone standing somewhere: a fixed viewpoint from the back of the table, at mug height and a little too low, so the near table edge cuts across the bottom of the frame | a close handheld viewpoint that drifts and re-settles | an arm's-length frontal viewpoint>. <Its flaws for this clip: the focus hunts once when the hands move in; the exposure lifts when the white box fills the frame.>
SCENE: <a real room, named: a kitchen counter at the end of the day | a desk with the day's things still on it>. <The one light source, and where it is.> <What it does badly: the near side bright, the far side in shadow.>
PRODUCT: <shape, material, colour, and any text on it verbatim in quotes>.
         [@image1 supplies the product exactly — <shape, label, colour>; ignore its background.]
[PERSON: <only what is in frame — two hands, adult, no jewellery | a person in their 20s in a grey hoodie, seen from the chest up>. <No name, no wardrobe manifest: this is not a character that has to survive anything.>]

0–<N>s    <one ordinary action>. <What the frame does about it: it sits a touch low, the focus hunts, the view tilts up.> <the sound this action makes>
<N>–<M>s  <the next action in the chain — no transition word; these are beats in one continuous shot unless the clip genuinely cuts>
…
<…>–<T>s  <the last action just finishes>. <The shot keeps running for a moment after it.>

AUDIO: <room tone of the actual room>, <two or three sounds the actions make: cardboard tearing, tape, a lid, a cup on a counter>. [<Speaker>: "<one line, in the language to be spoken>" — spontaneous, slightly breathless, not scripted.] No music.
AVOID: <the UGC anti-polish row from prompt-structure.md's Negative list — the frequent items, quoted as written>; plus: a phone, a camera, a tripod or a lit screen visible anywhere in the shot; tripod stability, gimbal smoothness, a studio or seamless backdrop, a professional model or presenter, an advertising voiceover, a slogan, a logo card, an end card; <the lettered objects in this room you do not want: book spine, cereal box, magazine, laptop screen, poster>; music, score, soundtrack, instrumental, percussion, humming, singing.

Six things about that shape:

• CAPTURE: describes the shot, never the phone, and the AVOID list names the device as an object anyway. That slot was DEVICE: and read "where the phone is", which put a phone in the delivered clip on run 2ecbedec — the ⚠️ under "The capture is part of the fiction" has that measurement. The corrected wording has since been run and held: job ede33e6d (2026-09-17), six sampled frames, no phone, camera, tripod or lit screen anywhere. So this line now has a measurement on both sides — the construction that fails and the conversion that fixes it — which is the strongest evidence any single line in this template carries. One run each; the conversion rule is still worth re-reading before you write the slot.
• No header FORMAT line and no shot manifest. UGC clips are short and single-shot; the shared file's Short prompts (10 seconds or less) says to skip the manifest, and a manifest is itself a tell.
• The timeline is beats inside one continuous shot, not cuts. Say one continuous shot in the first sentence if you mean it — the shared file records that a timestamped list without that declaration reads as a cut list. A real phone video does sometimes cut (they stopped and restarted); if you want that, say hard cut explicitly, because an unnamed boundary becomes one anyway.
• No climax and no hero frame. The chain just ends. Case 25's ending is the UGC close — but write it as the shot keeps running for a moment, not as case 25's own the camera continues recording for a moment, for the reason in the bullet above: the close is the last place you want to hand the model a device to render.
• No music, ever. Not a taste rule: a prompt asking this model for music came back output_moderation_failed on audio copyright in this repo (unbilled). Ambient and action sounds only; a track goes on in an editor. The shared file's Asking for music can fail output moderation on copyright has the measurement. Case 25 also lists music under STRICTLY AVOID for genre reasons, so both arguments point the same way.
• The AVOID slot is a quote, not a paraphrase. Open prompt-structure.md, find the UGC anti-polish row, and put its wording in. If that file is out of reach, see "If the script isn't found" — this is the one thing in this skill that genuinely degrades without it.

Worked example — a 10-second vertical unboxing, hands only

Text-to-video, no photo. This example has not been generated as written — it is the template filled in, not a clip anyone has paid for. Its CAPTURE line and its last two beats are the corrected form of the construction that put a phone in the shot on job 2ecbedec. That correction is no longer reasoning alone: the same corrected construction was generated on job ede33e6d (2026-09-17, with a photo attached) and no device appeared. This particular filled-in example is still unrun; the shape of its CAPTURE line has a run behind it.

STYLE: real phone capture texture, slight sensor noise and compression artefacts, mild autofocus hunting, exposure shifting when a bright surface enters the frame. Vertical, one continuous shot. Nothing is graded, lit or staged.
CAPTURE: a fixed viewpoint from the back of the table, at mug height and a little too low, so the near table edge cuts across the bottom of the frame and the box sits right of centre. The focus hunts once when the hands move in, and the exposure lifts when the white box fills the frame.
SCENE: a kitchen table at the end of the day, a used cup and a set of keys pushed to one side. The only light is a window to the left with the curtain half drawn, so the near side of everything is bright and the far side falls into shadow; a cool cast on the white wall behind.
PRODUCT: a matte black cylindrical hand grinder with a walnut knob and a folding steel crank, in a plain white cardboard box with no printing on it.
PERSON: two adult hands, no jewellery, sleeves pushed up. Nothing above the wrists is ever in frame.

0-3s    a hand slides the white box into the middle of the table; it arrives off-centre and the far corner is cropped, and the hand nudges it straight. The focus hunts briefly on the tabletop before settling. Sound: cardboard on wood, a chair creaking off-screen.
3-6s    both hands work the tape at the seam, pull it off in one go and drop it beside the box; the flaps come up. Sound: tape tearing, the flaps.
6-8.5s  one hand lifts the grinder out. It comes up slightly out of frame at the top and the view tilts up a few degrees to catch it, then settles. The exposure drops as the dark body fills the frame.
8.5-10s the hands turn it over once, the crank folds out, and they set it down on the table. The shot keeps running for a moment after the hands stop.

AUDIO: kitchen room tone, a fridge hum, cardboard, tape, the crank clicking out, the grinder set down on wood. No dialogue. No music.
AVOID: cinematic color grading, beauty filters, artificial skin smoothing, dramatic slow motion, music, perfect lighting; no beauty filter, grading, cinematic/CG look, fisheye, vignette, or legible text; a phone, a camera, a tripod or a lit screen visible anywhere in the shot; tripod stability, gimbal smoothness, a studio or seamless backdrop, a professional model or presenter, an advertising voiceover, a slogan, a logo card, an end card; book spine, magazine, cereal box, laptop screen, poster, printed packaging other than the plain white box; music, score, soundtrack, instrumental, percussion, humming, singing; warped or extra fingers.

The AVOID list's first two clauses are the shared file's UGC anti-polish row verbatim — that is what "quote it" means. Everything after them is what this scenario adds on top.

⚠️ One clause in that row fights a real product, and you have to notice it. It ends …fisheye, vignette, or legible text — written for a text-only clip where every letter would be invented. If the brief is an unboxing of a real product whose label the user wants readable, a blanket no legible text is now pointed at the thing they are advertising. The example above dodges it by giving the product a plain unprinted box, which is a legitimate choice and not the only one. When the label has to be legible:

• lock it in an attached first frame rather than describing it — text approved on a still and then preserved is the reliable route, measured in this repo, and text rendered from a description is the classic failure;
• drop the legible text clause from the quoted row and replace it with the specific carriers you don't want (the lettered objects already listed in the AVOID slot), so the prohibition removes the background signage without removing the product's own label.

That is the one edit to the shared row this scenario routinely needs. Make it deliberately and say you made it.

If the creator speaks

One line usually beats a script. The shared file's Dialogue and sound → Density — two tiers, not one gives the measured word budgets: a talking-head runs about 3.5 words a second and a drama far below that, so a 10-second clip of continuous commentary is a talking-head budget and a single closing line is nowhere near it. Write the line in the language it should be spoken in — that is what decides the voice's language.

The delivery note is where UGC lives: spontaneous, slightly breathless, not scripted (case 25), a false start, a word repeated. An advertising cadence in a handheld clip is the polish leak in audio form.

--generate-audio defaults to true on the server, which is what you want for ambient sound. Set it false only if the user is laying their own audio over the clip.

Vertical by default

--aspect-ratio 9:16 on pure text-to-video. Two things to know:

• With a photo attached the flag stops being yours. ofox-video-core forces adaptive on bytedance/seedance-2.5 and defaults to it on other models that offer it, so the output follows the image. Crop the photo to 9:16 before generating — crop, never pad, since padding bakes the bars into the video. Relay the NOTE: the script prints.
• Not every model offers 9:16. Aspect-ratio lists differ per model and they are catalog facts, so read them rather than assuming: ofox-video.sh models. A ratio the model doesn't have is rejected locally, free, before anything is submitted.

Recommended defaults

Parameter	Default	Why
--aspect-ratio	9:16 on text-to-video; not passed when a photo is attached	UGC is vertical by default; with a frame attached the photo's crop decides the shape
--model	the script's own default, unless the user named one, which always wins	the id the request will really carry is the MODEL line a generate --dry-run prints, and that is the id the cost table has to name. See "Choosing a model"
--duration	short — a draft at the model's minimum, a deliverable in the low tens of seconds	a flat action chain has no reason to run long, and every second bills. ofox-video.sh models prints each model's range
--resolution	draft at the model's cheapest tier, then re-render the keeper higher. When the user named neither, put the next tier up as a second row of the same cost table rather than asking	UGC is the one genre where a lower tier costs the least credibility — sensor noise and compression are in the brief. Still read the draft's frames before shipping it
--generate-audio	leave at the server default (true)	room tone and action sounds are half of what makes it read as real
--seed	let the script roll one and keep it	printed as SEED and written to the .json sidecar, which is what lets "that take, rendered properly" be re-submitted at all. It does not reproduce it: measured, an identical request on a fixed seed came back a visibly different clip. Say a re-render is another roll aimed at the same shot before the user pays for it
--frame-first-image	only when a product photo exists, and only if it contains no person	the measured product-lock route; see "People, products and what the API refuses"
--real-person	leave unset	this skill's route does not need it — the product is the attached frame and the person is written in text. true is Ofox's privacy-preserving preprocessing path for authorised real-person reference images, measured lifting seedance-2.5's submission refusal on 2026-09-16: an authorisation route, never a way past the check, only offered when the user holds the right to that likeness and has said so, and never set on their behalf. See api-params.md → "--real-person true lifts that refusal on 2.5"

Choosing a model

The model stays never-ask — the agent doesn't raise it. But never-ask is not "never listen": if the user names a model id or a shorthand, use it.

Model ids, prices, resolutions, durations and aspect ratios are deliberately not written down in this file. They are catalog facts, they change, and this repo has recorded defects that trace to a hardcoded copy of somebody else's value table. Read them live, free, with no API key:

bash ../ofox-video-core/references/ofox-video.sh models             # every model, its tiers and ranges
bash ../ofox-video-core/references/ofox-video.sh providers MODEL    # that model's per-resolution rates

providers with no model argument prints the flagship's matrix, not the catalog — pass the id you actually mean. And the rate models shows is the one at each model's own default resolution, which differs between models, so it ranks rather than quotes. The number you put in front of a user comes from generate --dry-run at the parameters you are about to send.

Two things worth saying to a user who is choosing: a cheaper model is a different look, not just a smaller bill — which cuts both ways here, since this is the scenario least harmed by a rough one. And moderation policy is per-model, so a prompt refused on one model can be accepted on another.

Before you spend: the approval gate

Never submit a paid job until the user has seen a cost table and said yes. The rule, the required columns and where the numbers must come from are written down once for every Ofox skill in this repo: ../ofox-video-core/references/approval-gate.md.

Get the numbers from --dry-run, which validates everything and prints the estimate without sending a request:

bash ../ofox-video-core/references/ofox-video.sh generate --dry-run \
  --prompt "..." --duration 10 --resolution 480p --aspect-ratio 9:16 \
  --out-dir /absolute/path/to/out

Relay the Estimated cost: line it prints — never a number of your own — then wait for a yes, then re-run the identical command with --dry-run swapped for --approved. The estimate a real run prints comes microseconds before the request goes out, too late to relay. Pass the same --out-dir to both: the dry run creates and enters it, so a bad path fails as exit 6 with nothing submitted.

--approved is where that yes gets typed out. Since ofox-video-core 2.0.0 the four billable subcommands — generate, create, batch, chain — refuse to run without it, while --dry-run never needs it, so the quote above is still free and still works with no API key. Be exact about what the flag does: it records a stance, it cannot prove one. Nothing in a shell script can observe the conversation you had, and it can be typed without showing anyone a price. What it changes is that spending without quoting is no longer the default — it has to be written into the command, where a transcript shows it. The rule above is still the rule, and it is still yours to follow.

The brief recap goes in the same message as the prompt and the table, above them.

The one cost anchor this skill has, with the parameters it was measured at attached, because a figure without its parameters is not a measurement: bytedance/seedance-2.5, 8 seconds, 480p, 9:16, text-to-video, audio on — 88 cents billed (job 2ecbedec). It is not a quote for another duration, tier, model or for an image-to-video job, and it must never be scaled by hand into one. The dry run at the parameters you are actually about to send is the only number to show a user; this anchor is there to sanity-check that the dry run is in the right neighbourhood, never to replace it.

Afterwards the actual bill is VIDEO_COST from the finished job. Report it as money, not as the raw ten-decimal string.

Several takes, and the cheap vertical route

A UGC clip is a roll like any other, and this one has an extra reason to roll: the thing being judged — "does it read as real?" — is exactly the thing the prompt can ask for and not get.

For several takes of one prompt, batch prices the whole set up front, waits for the takes concurrently, and tiles a contact sheet:

bash ../ofox-video-core/references/ofox-video.sh batch --dry-run \
  --prompt "..." --takes 4 --duration 4 --resolution 480p \
  --aspect-ratio 9:16 --out-dir /absolute/path/to/out

Quote BATCH_COST_TOTAL, not BATCH_COST_PER_TAKE, and give the takes a row each — if one take in four is usable, that clip cost the whole total. Hand over the CONTACT_SHEET path on its own line, then the take paths beneath it; each TAKE line carries seed=N.

If the user wants several vertical drafts to choose from as the actual deliverable — Shorts or Reels raw material rather than one clip — shorts-reels is that scenario, and it carries the cheap-tier ladder and the draft-then-promote flow. The two compose: write the prompt with this skill's template, run the set through that one.

Pricing a job with no API key

models, providers and generate --dry-run all work with OFOX_API_KEY unset. So when a user hasn't signed up yet, quote the job first and let them decide whether it's worth registering — don't open by sending them to a signup form.

If the script isn't found

bash: ../ofox-video-core/references/ofox-video.sh: No such file or directory

Nothing is broken — this skill delegates all execution to ofox-video-core and reaches it by relative path, and that path just missed. Two different situations wear this message, so run the probe in "Where the core skill lives" before deciding which:

• The probe printed a directory — the core is installed and only the directory name was wrong, which is the normal LobeHub case (ofoxai-skills-ofox-video-core). Re-run against what the probe printed.
• The probe printed nothing — ofox-video-core really is absent, and installing it is the user's call to make, not yours: an install writes outside this working directory, so hand over the command and let them run it rather than running it for them. Which command depends on the installer they already have — skills.sh is npx skills add ofoxai/skills --skill ofox-video-core, which asks for that one skill and answers none of the agent, scope or confirmation questions on the user's behalf; this repo's own wrapper is npx ofox-skills ofox-video-core, the same install with all three answered in advance (every agent, user-level, no prompts); on LobeHub or ClawHub, install ofox-video-core from the same publisher. Ask for the one skill that is missing rather than the whole repo, and give all three routes — pointing a LobeHub user at the skills.sh line alone reads as "abandon your installer", which isn't the advice.

Either way, name the missing skill and where it was expected rather than relaying the raw path error, which names neither.

A broken link to a shared reference has the same two causes. This skill degrades more than most when that happens, and it is worth saying so: the question set, the template, the four levers and the defaults are all written out here, but the anti-polish AVOID list is deliberately not, and that list is half the brief. Without it, write the AVOID slot from the four categories named in "Where the anti-polish vocabulary lives" — grading, beauty filters and skin smoothing, dramatic slow motion, perfect lighting and the CG look — and tell the user the canonical wording was unavailable.

Exit codes worth knowing

Full table in ../ofox-video-core/SKILL.md. The ones that come up:

Code	Meaning	What to do
1	Parameter rejected locally, no network call, nothing billed	Fix the flag and retry freely
2	Environment problem — curl/jq missing, or no OFOX_API_KEY	Ask the user to fix it; check reports the same
3	API rejected it, or the job ended failed/cancelled/expired	Read the mapped message; a rejected create was not billed
4	Timed out waiting — the job is still running and billable	poll JOB_ID, never re-run generate
5	Ambiguous network failure on create	Do not retry blindly; check https://app.ofox.ai first
6	--out-dir unusable	Fix the path; if it happened after a create, poll JOB_ID

How long to tell the user it will take

generate blocks while it polls, up to --max-wait (default 540s). A short low-resolution vertical clip is usually one to three minutes. Say so before starting, so the wait isn't silent.

If your tool call can't stay open that long, use create (submits and returns a job id in seconds) followed by poll. That way a timeout can never strand a job whose id you never saw. For batch, the worst case is takes x max-wait — lower --max-wait for drafts, or create and poll the takes yourself.

Where the file lands

Always pass --out-dir, and make it an absolute path. Without it the script writes to the current working directory — which, given that the examples here run from this skill's own directory, would drop the user's video inside an installed skill. Relay the absolute VIDEO_PATH the script prints, on its own line.

Pass --name too — name the file after the clip rather than leaving the script to guess from the prompt's opening words, which here describe a phone's sensor noise. The clip lands as <name>-<short job id>.mp4 with a .json sidecar holding the full job id, the prompt, the seed and the real cost.

Generating

bash ../ofox-video-core/references/ofox-video.sh generate --approved \
  --prompt "<the UGC prompt built above>" \
  --name "<short clip name, e.g. hand grinder unboxing>" \
  --duration 10 \
  --resolution 480p \
  --aspect-ratio 9:16 \
  --out-dir /absolute/path/to/out

Drop --aspect-ratio when a product photo is attached — see "Vertical by default".

--approved is not decoration: without it the script refuses, submits nothing, and prints the quote-first steps instead. Add it only once the cost table has actually gone in front of the user and come back with a yes — the flag cannot check that for you.

This one call validates the parameters, submits the job, polls to completion, downloads the mp4, and prints STATUS, JOB_ID, VIDEO_PATH, VIDEO_SECONDS, SEED and VIDEO_COST. Report the actual values from that output — never the estimate, and never a path or cost you didn't see the script print. Do not re-implement any of the request, poll or download logic here.

Common failure modes and fixes

Symptom	Cause	Fix
The clip is beautiful and reads as an ad	Polish leaked in — style words carried from another scenario, a climax beat, or a camera that behaves	Re-read the prompt for cinematic, studio, professional, film grain, rim light, shallow depth of field, slow motion; flatten the beat chain; put the phone's misbehaviour in the timeline as stamped beats rather than in a style line. See "The inversion". New prompt, new cost table
A phone, camera or tripod is visible in the shot, sometimes with a lit screen	The prompt gave the filming device a position in the room, so the model treated it as set dressing. Measured on job 2ecbedec, from the phrasing the phone is propped against a biscuit tin at the back of the desk	Rewrite the CAPTURE line as a viewpoint (a fixed viewpoint at desk height, a little low and off-square), convert any beat that has a hand touching the phone into something the frame does, and add a phone, a camera, a tripod or a lit screen visible anywhere in the shot to AVOID as objects. New prompt, new cost table — nothing in the tooling can catch this before delivery
The camera is smooth despite "handheld"	A camera texture written once as an adjective. The shared file's measured rule is that a move written only as a verb tends not to happen; the same applies to instability	Write the re-frame as a beat with its own timestamp, as something the frame does — "the view tilts up a few degrees to catch it, then settles" — never as a hand adjusting a phone
The lighting is flattering and even	The prompt named a light but not what it does badly	Name the source, its position, and the shadow or cast it leaves. Positive sentences, per the shared file's The plastic look is designed out, not forbidden
Invented signage or garbled lettering on the shelf behind	A real room is full of lettered surfaces, and the negative list is only partly obeyed on such a set (measured, four jobs)	Move the carriers out of frame or out of focus and forbid them as objects — book spine, box, magazine, laptop — not as "text". See lever 4
Exit 3, input_moderation_failed on an image job	The attached photo contains a person — a hand or a foot is enough	Attach a photo of the product alone and write the person in text. Nothing was billed
Exit 3, output_moderation_failed mentioning audio copyright	The prompt asked for music	Rewrite AUDIO as room tone plus recorded sounds, keep the music words in AVOID, and re-run — a new request, safe immediately
Hands look wrong	Hands are the highest-risk anatomy and this genre puts them centre-frame	One simple action per beat, the product as the beat's subject rather than the hand, and no written "never more than X of the hand" clause — that was measured being ignored. Re-roll, or move to a wider framing
The product's label is wrong or drifts	Text rendered from a description is the classic failure	Attach a real product photo as the first frame; text approved on a still and then preserved is a different and reliable task
The clip has a music bed nobody asked for	--generate-audio true plus a prompt that didn't exclude music	Keep the music words in AVOID and name the ambient sounds you do want
Exit 1 on --aspect-ratio 9:16	The model doesn't offer that ratio	ofox-video.sh models lists each model's ratios. Nothing was submitted, so this is free to fix
The output isn't vertical even though 9:16 was passed	A frame is attached, so the ratio is adaptive and follows the image	Crop the photo to 9:16 and re-run. Crop, never pad
Exit 4, timed out waiting	Still running upstream, not failed	poll JOB_ID with the id printed before the timeout; never re-run generate
Exit 5, ambiguous network failure on create	No HTTP response at all — can't tell whether a job exists	Don't guess or retry; tell the user to check https://app.ofox.ai

When NOT to use

• A polished brand or product ad — seedance-ad-creative. Its UGC-variant table is the ancestor of this skill; anything with a hero frame, a grade or a slogan belongs there.
• Plain catalog or listing footage — seedance-product-video.
• A scene carried by people talking to each other — seedance-short-drama. A creator talking to the camera about a product stays here; a conversation between characters does not.
• Several cheap vertical drafts as the deliverable — shorts-reels, which owns that ladder. Write the prompt here, run the set there.
• The user can just film it. A real unboxing on a real phone costs nothing, shows the actual product and cannot fail an authenticity check. Say so before quoting a job: this skill earns its keep when the product isn't in hand, when the volume is too high to film, or when the shot is needed before the stock arrives.