Morphix

View saved version →

---
name: cinematic-shot
description: Create a short cinematic clip from a written idea or a reference frame, using film camera and lighting language. Trigger when the user asks for a cinematic shot, a film-look clip, a movie scene, b-roll, an establishing shot, a moody or atmospheric video, or describes a scene they want shot. Do not trigger for animating a personal photograph, for product marketing videos, for editing existing footage, or when the user wants a still image rather than motion.
---

# Cinematic Shot

Produce one short clip built with deliberate film grammar rather than a generic description.

## Morphix tool contract

These are the tools available and the order they run in. Follow it exactly.

**Getting the user's image in.** Chat attachments cannot be passed to generation. When a reference is needed, call show_upload_widget, which shows a drop box and returns an asset id once the user uploads. Never call upload_asset directly. If the user uploaded something earlier in the conversation, call get_assets and reuse that asset id rather than asking again.

**Before spending credits.** Call explore_models to choose a model for this job, and get_account to confirm the user has credits. Both are read-only and free. Choosing a model blind, or starting a generation the user cannot pay for, are the two most common wasted turns.

**Choosing a model.** Never hardcode a model name. Read the capabilities returned by explore_models and pick for the job at hand: reference fidelity when the result is anchored to a supplied image, text handling when legible type matters, motion quality for video, voice range for audio.

**Generating.** generate_image and generate_video accept asset ids to edit or reference existing images, including image-to-video. generate_audio takes the script plus a voice chosen through explore_models. Each spends credits and renders a live view that polls itself, so never poll manually and never tell the user you will check back.

**Retry spends again.** A regeneration costs credits a second time. The quality gate in this skill allows one automatic retry. Beyond that, say what is wrong and let the user decide.

**Finishing.** upscale_image and upscale_video raise an existing asset. Offer these instead of regenerating at a larger size.

**Recovering work.** get_assets lists the library newest first. get_generations fetches a generation by id. Use them to pick up earlier work instead of remaking it.

## The difference this skill makes

A generic video prompt returns motion. A cinematic shot returns a decision: a specific lens, a specific light source, a specific camera behaviour, and a reason for all three. Always make those decisions explicitly, and state them back to the user in one line so they can redirect.

## Preconditions

Either a written scene description or a reference frame. If the user supplies a photograph of a real person, route to photo-to-video instead, which protects likeness.

## Intake

Ask once, bundled: the scene; the mood in one word; duration; and aspect ratio. Offer five, eight, and ten seconds. If the user gives only a subject with no mood, choose one, say which you chose, and proceed.

## Build the shot from four decisions

Make all four before writing a word of the prompt.

**1. Shot size.** Extreme wide for scale and isolation. Wide for context. Medium for relationship. Close for intent. Extreme close for texture and tension. Do not default to medium.

**2. Lens behaviour.** Wide angle exaggerates depth and distorts edges, good for space and unease. Normal reads neutral and documentary. Long lens compresses planes and separates subject from ground, good for intimacy and surveillance. Name the implied focal length in the prompt.

**3. Light.** Name the source, direction, and quality. Hard low sun from behind for rim and haze. Soft north window from the side for skin and melancholy. Single practical in frame for noir. Overcast top light for flat realism. Moonlight and fill for night. Light is what separates cinematic from rendered.

**4. Camera behaviour.** Locked for stillness and dread. Slow push for growing focus. Slow pull for reveal and loneliness. Handheld drift for presence and unease. Slow lateral track for observation. Crane or tilt for scale. One move per shot, never two.

## Grade and atmosphere

Name the grade in concrete terms: cold shadows with warm highlights, bleached and low contrast, heavy blacks with a single colour source, natural and unpushed. Avoid the phrase cinematic as a style word in the prompt, because it collapses into a generic look. Describe what makes it cinematic instead.

Atmosphere carries motion cheaply and safely: dust in a beam, rain on glass, fog rolling, steam, embers, curtains, foliage. Add at least one when the frame would otherwise be static.

## Write the prompt

One continuous description: subject and action, setting, shot size, implied lens, light source and direction, camera behaviour and speed, atmosphere, grade, and duration.

Avoid stacked style words, avoid naming directors or films, avoid trademarked looks.

## Execute

Call generate_video with the reference frame when one exists. Use sixteen-by-nine for film feel, nine-by-sixteen for vertical social, and two-point-three-nine to one only if the user asks for anamorphic framing.

## Quality gate

Reject if the clip drifts into a different scene, if the light source changes direction mid shot, if the camera performs a second unrequested move, if faces or hands deform, or if text appears in frame.

Regenerate once with a slower move or a simpler frame.

## Deliver

Return the clip and one line naming the four decisions made: shot size, lens, light, camera. That line is what lets the user ask for a different take rather than a different prompt.

## Stop conditions

| Condition | Response |
|---|---|
| Neither description nor reference supplied | Stop and ask what the scene is |
| User supplies a photo of a real person to animate | Route to photo-to-video |
| Request names a specific film, director, or trademarked look | Offer the underlying technique instead |
| Quality gate fails twice | Deliver the most coherent take and state what broke |

Referenced files: 1

View saved version →

---
name: product-ad-variations
description: Produce an ordered set of creative variations of the same product, each changing one controlled variable, so the user can test, compare, or run them across placements. Trigger when the user supplies a product image and asks for multiple versions, variations, angles, options, A/B test creatives, or a set of ads for the same product. Do not trigger for a single image, for video variations, for applying label artwork to a container, or when the user has not supplied a photograph of the real product.
---

# Product Ad Variations

Produce N ordered variations of one product creative, built so the result of a test can actually be read.

## Morphix tool contract

These are the tools available and the order they run in. Follow it exactly.

**Getting the user's image in.** Chat attachments cannot be passed to generation. When a reference is needed, call show_upload_widget, which shows a drop box and returns an asset id once the user uploads. Never call upload_asset directly. If the user uploaded something earlier in the conversation, call get_assets and reuse that asset id rather than asking again.

**Before spending credits.** Call explore_models to choose a model for this job, and get_account to confirm the user has credits. Both are read-only and free. Choosing a model blind, or starting a generation the user cannot pay for, are the two most common wasted turns.

**Choosing a model.** Never hardcode a model name. Read the capabilities returned by explore_models and pick for the job at hand: reference fidelity when the result is anchored to a supplied image, text handling when legible type matters, motion quality for video, voice range for audio.

**Generating.** generate_image and generate_video accept asset ids to edit or reference existing images, including image-to-video. generate_audio takes the script plus a voice chosen through explore_models. Each spends credits and renders a live view that polls itself, so never poll manually and never tell the user you will check back.

**Retry spends again.** A regeneration costs credits a second time. The quality gate in this skill allows one automatic retry. Beyond that, say what is wrong and let the user decide.

**Finishing.** upscale_image and upscale_video raise an existing asset. Offer these instead of regenerating at a larger size.

**Recovering work.** get_assets lists the library newest first. get_generations fetches a generation by id. Use them to pick up earlier work instead of remaking it.

## The governing principle

A variation set is only useful if the user can attribute a difference in outcome to a specific change. That requires each version to change one axis while everything else is held constant. A set where several things change at once is an assortment, not a test. Never produce an assortment.

## Product Lock

The product in the supplied reference is immutable data across every version.

Preserve its shape, proportion, colour, material, finish, label geometry, typography, logo, and every printed character exactly as they appear. Never redesign, restyle, correct, translate, or recreate any part of the product. If label text is unreadable in the reference, frame it out of focus or out of frame rather than reconstructing it.

## Preconditions

A photograph of the real product is required. Every version spends credits, so resolve the axis and the constants before the first generation.

## Intake

Ask once, bundled: the product photo; how many versions, N; which axis to test; and where the creatives will run. If the user does not know which axis, recommend scene for top of funnel and angle for listings and retargeting, then proceed.

## Permitted axes

Choose exactly one as the axis of the set.

| Axis | What changes between versions | Use when testing |
|---|---|---|
| Scene | the environment around the product | positioning and audience |
| Angle | camera position | product legibility in feed |
| Light and mood | lighting scheme and palette | brand tone |
| Human presence | no person, hands only, partial use | social proof |
| Format | aspect ratio and crop | placement adaptation |

## Fix the constants first

Before generating, write down everything that does not change: the product, camera distance, base palette, aspect ratio, and the lighting treatment unless light is the axis. Every version inherits these verbatim. Writing the constants before writing the variations is mandatory, otherwise the set drifts and the test becomes unreadable.

## Write N prompts in order

Each prompt repeats the constants word for word and alters only the value of the chosen axis.

Order is stable from planning through delivery. Version 1 is always the most conservative and closest to conventional use of the product. Later versions move progressively further from it. This gives the user a readable baseline.

Cap N at six in one round. Beyond six the user cannot compare and the credit cost stops being justified. If more are requested, deliver six and offer a second round after the first is read.

## Execute

Call generate_image once per version, in order, each with the same product reference.

## Quality gate

Reject and regenerate a single version once if the product diverges from the reference, if text is invented, if a duplicate product appears, or if the version accidentally changed an axis other than the chosen one. Frozen constants are the point of the set.

## Deliver

Return the images labelled Version 1, Version 2, and so on, each with one line naming the axis value it carries. Close with one sentence naming the axis tested and what a win for each version would indicate.

Never promise performance, never estimate click-through or conversion, never declare a winner before the test runs.

## Stop conditions

| Condition | Response |
|---|---|
| No product photograph supplied | Stop and ask for one |
| User wants several things to change at once with no axis | Explain in one sentence that the set would be unreadable, then ask for the axis |
| User asks to change the product design | Stop and ask: different product, or different scene |
| More than six versions requested | Deliver six, offer a second round |

Referenced files: 1

View saved version →

---
name: product-photography
description: Photograph a product the user supplies, in a studio or lifestyle setting, for a store listing, website, catalog, or paid ad. Trigger when the user provides a product image and asks for a product photo, product shot, hero shot, studio shot, lifestyle shot, packshot, catalog image, or listing image. Do not trigger for applying flat label artwork onto a container, which belongs to packaging-mockup. Do not trigger for video, for multi-version test sets, or when the user has not supplied a photograph of the real product.
---

# Product Photography

Produce one photograph of a real product supplied by the user.

## Morphix tool contract

These are the tools available and the order they run in. Follow it exactly.

**Getting the user's image in.** Chat attachments cannot be passed to generation. When a reference is needed, call show_upload_widget, which shows a drop box and returns an asset id once the user uploads. Never call upload_asset directly. If the user uploaded something earlier in the conversation, call get_assets and reuse that asset id rather than asking again.

**Before spending credits.** Call explore_models to choose a model for this job, and get_account to confirm the user has credits. Both are read-only and free. Choosing a model blind, or starting a generation the user cannot pay for, are the two most common wasted turns.

**Choosing a model.** Never hardcode a model name. Read the capabilities returned by explore_models and pick for the job at hand: reference fidelity when the result is anchored to a supplied image, text handling when legible type matters, motion quality for video, voice range for audio.

**Generating.** generate_image and generate_video accept asset ids to edit or reference existing images, including image-to-video. generate_audio takes the script plus a voice chosen through explore_models. Each spends credits and renders a live view that polls itself, so never poll manually and never tell the user you will check back.

**Retry spends again.** A regeneration costs credits a second time. The quality gate in this skill allows one automatic retry. Beyond that, say what is wrong and let the user decide.

**Finishing.** upscale_image and upscale_video raise an existing asset. Offer these instead of regenerating at a larger size.

**Recovering work.** get_assets lists the library newest first. get_generations fetches a generation by id. Use them to pick up earlier work instead of remaking it.

## Product Lock

The product in the supplied reference is immutable data.

Preserve its shape, proportion, colour, material, finish, label geometry, typography, logo, and every printed character exactly as they appear. Angle, lighting, surface, background, relative scale, and framing are free. Never redesign, restyle, correct, translate, beautify, or recreate any part of the product. If the request requires changing the product itself, stop and ask whether the user wants a different product or a different scene.

If label text in the reference is blurred, cropped, or unreadable, do not reconstruct it. Frame so that face falls out of focus, turns to profile, or sits partially out of frame. Invented packaging text is the most expensive defect in this category.

## Preconditions

A photograph of the real product is required. Never generate a product from a description, never substitute stock, never complete a poor reference with invention. Without a photo, stop and ask for one.

Every generation spends the user's credits. Resolve ambiguity before the first call.

## Intake

Ask once, bundled, only for what is missing: the product photo; where the image will be used (marketplace listing, website, paid ad, catalog, social); clean studio or lifestyle scene; and output shape. If the destination is a marketplace listing, assume a neutral background with the product centred, and say that you assumed it.

Never ask about models, resolution, providers, or internal mechanics.

## Classify the product before writing the prompt

Category determines treatment. This step is not optional.

- Transparent or liquid (bottle, flask, glass): reflection control is the core problem. Use a gradient backdrop, raking side light, and a vertical edge highlight. Never pure white, which erases the silhouette.
- Cosmetic and skincare: clean surface, soft short shadow, tonal background, no competing props.
- Electronics: edge definition sells. Controlled hard light, specular edge reflection, neutral cool or dark background.
- Textile and apparel: texture and drape. Wide diffuse light, flat lay or on a form, never hard light.
- Food and beverage: freshness cues. Condensation, a crumb, steam, window light. Never perfect symmetry.
- Rigid objects and tools: scale. Include a size cue or use a three-quarter angle.

## Choose the angle by function

Three-quarter at roughly thirty degrees is the hero default. Straight-on front for marketplace listings. Overhead for flat lay and food. Macro for texture and finish. Avoid dramatic low angles unless requested, as they distort product proportion.

## Write the prompt

One continuous description containing: the product named for what it is, the surface, the background, the lighting scheme, the angle and camera distance, the implied lens, the expected shadow and reflection, and the product preservation instruction. No keyword lists. No generic style words such as premium, high quality, or professional.

Always constrain against: added text, watermarks, third-party logos, duplicate products, hands unless requested, plastic skin if a person appears, exaggerated bokeh, and teal-and-orange grading.

## Execute

Call generate_image once, passing the product photograph as the reference image and the validated prompt.

Derive aspect ratio from destination: square for marketplace listings, four-by-five for social feed, sixteen-by-nine for banners and website heroes, nine-by-sixteen for stories.

## Quality gate

Before delivering, reject and regenerate once if the product diverges from the reference in shape, colour, or label; if any text is unreadable or invented; if a duplicate product appears in frame; if a hand shows the wrong number of fingers; or if a third-party brand mark is visible.

## Deliver

Return the image and one line describing what was made. Offer upscale_image to 4K when the destination is print, catalog, or a website hero. Never expose model names, job identifiers, prompts, or internal steps.

## Stop conditions

| Condition | Response |
|---|---|
| No product photograph supplied | Stop and ask for one |
| Product obstructed, cropped, or unrecognisable in reference | Stop and ask for a clearer photo |
| User asks to change the product design | Stop and ask: different product, or different scene |
| Quality gate fails twice | Deliver the closest passing result and state what could not be preserved |

Referenced files: 1

View saved version →

---
name: product-video
description: Turn a supplied product photograph into a short video for an ad, a store page, or social. Trigger when the user has a product image and asks to animate it, make a product video, a video ad, a product clip, or motion for their product. Do not trigger for narrated creator-style videos where a person speaks, for editing existing footage, for still images, or when the user has not supplied a photograph of the real product.
---

# Product Video

Produce one short video of a real product supplied by the user.

## Morphix tool contract

These are the tools available and the order they run in. Follow it exactly.

**Getting the user's image in.** Chat attachments cannot be passed to generation. When a reference is needed, call show_upload_widget, which shows a drop box and returns an asset id once the user uploads. Never call upload_asset directly. If the user uploaded something earlier in the conversation, call get_assets and reuse that asset id rather than asking again.

**Before spending credits.** Call explore_models to choose a model for this job, and get_account to confirm the user has credits. Both are read-only and free. Choosing a model blind, or starting a generation the user cannot pay for, are the two most common wasted turns.

**Choosing a model.** Never hardcode a model name. Read the capabilities returned by explore_models and pick for the job at hand: reference fidelity when the result is anchored to a supplied image, text handling when legible type matters, motion quality for video, voice range for audio.

**Generating.** generate_image and generate_video accept asset ids to edit or reference existing images, including image-to-video. generate_audio takes the script plus a voice chosen through explore_models. Each spends credits and renders a live view that polls itself, so never poll manually and never tell the user you will check back.

**Retry spends again.** A regeneration costs credits a second time. The quality gate in this skill allows one automatic retry. Beyond that, say what is wrong and let the user decide.

**Finishing.** upscale_image and upscale_video raise an existing asset. Offer these instead of regenerating at a larger size.

**Recovering work.** get_assets lists the library newest first. get_generations fetches a generation by id. Use them to pick up earlier work instead of remaking it.

## The failure mode this skill exists to prevent

Video models deform geometry and labels when motion is fast, when the camera crosses the face of the product, or when duration outruns the reference anchor. Every rule below exists because of that.

## Product Lock

The product in the supplied reference is immutable data for the entire shot.

Preserve shape, proportion, colour, material, finish, label geometry, typography, logo, and every printed character. The product stays rigid. Camera, light, and environment move; the product does not change.

If label text is unreadable in the reference, keep that face away from camera rather than reconstructing it.

## Preconditions

A photograph of the real product is required. The product should occupy at least a fifth of the frame in the reference; below that, ask for a tighter photo before spending credits.

## Intake

Ask once, bundled: the product photo; duration; where it will run; and whether the user wants sound. Offer five, eight, and ten seconds.

Above ten seconds, say plainly that product fidelity degrades and propose two short shots instead of one long one.

## Permitted camera moves

Only these, because they preserve product legibility.

| Move | Effect | Best duration |
|---|---|---|
| Slow push-in | focus and desire | 5 to 8s |
| Partial orbit, up to forty-five degrees | reveals form and material | 8 to 10s |
| Tilt reveal, low to high | scale and presence | 5 to 8s |
| Pull-back from macro to wide | texture then context | 8 to 10s |
| Locked camera with environmental motion | product still, light or background moves | any |

Forbidden: whip pan, full three-hundred-and-sixty degree orbit, morph or transformation of the product, hard cut inside the shot, abrupt zoom, and any move that passes behind the product. If the user requests one, explain in one sentence that it deforms the label and offer the permitted equivalent.

## Put the motion in the environment

When the product carries dense label text, the safest way to get a living shot is to keep the product locked and move everything else: airborne particles, smoke, liquid, a falling leaf, a shifting light, a travelling shadow. Recommend this by default for text-heavy packaging.

## Sound

If the user wants sound, generate ambience or music only, using generate_audio.

Do not write spoken narration in this skill. Voiceover with a scripted presenter is a different format and out of scope. If narration is requested, say so in one sentence and offer the ambience version.

## Write the prompt

Name the product, the exact camera move and its speed, what moves in the environment, the lighting, the duration, and the instruction that the product stays rigid with geometry and label unchanged throughout the shot.

## Execute

Call generate_video with the product photograph as reference. Use nine-by-sixteen for vertical social, square for feed, sixteen-by-nine for website and long-form platforms.

## Quality gate

Inspect the first frame, a middle frame, and the last frame. The product must be the same object in all three.

Reject if the shape changes, the label shifts position relative to the form, any text changes, colour drifts, or a second product appears. Regenerate once with a slower move or a shorter duration.

## Deliver

Return the video and one line describing the move and duration. Never expose model names, job identifiers, prompts, or internal steps.

## Stop conditions

| Condition | Response |
|---|---|
| No product photograph supplied | Stop and ask for one |
| Product occupies less than a fifth of the reference frame | Stop and ask for a tighter photo |
| User requests a forbidden move and refuses the alternative | Stop and explain the deformation risk |
| User requests spoken narration | State that this is a different format, offer the ambience version |
| Quality gate fails twice | Deliver the most stable result and state what drifted |

Referenced files: 1

View saved version →

---
name: restore-and-upscale
description: Repair and enlarge an existing image, recovering real detail rather than smoothing it away. Trigger when the user asks to upscale, enhance, sharpen, restore, fix, de-blur, or increase the resolution of an image they supply, including old, scanned, damaged, or low resolution photographs. Do not trigger for generating new images, for changing the content of an image, for video, or for stylistic edits.
---

# Restore and Upscale

Make an existing image larger and better without turning it into a different image.

## Morphix tool contract

These are the tools available and the order they run in. Follow it exactly.

**Getting the user's image in.** Chat attachments cannot be passed to generation. When a reference is needed, call show_upload_widget, which shows a drop box and returns an asset id once the user uploads. Never call upload_asset directly. If the user uploaded something earlier in the conversation, call get_assets and reuse that asset id rather than asking again.

**Before spending credits.** Call explore_models to choose a model for this job, and get_account to confirm the user has credits. Both are read-only and free. Choosing a model blind, or starting a generation the user cannot pay for, are the two most common wasted turns.

**Choosing a model.** Never hardcode a model name. Read the capabilities returned by explore_models and pick for the job at hand: reference fidelity when the result is anchored to a supplied image, text handling when legible type matters, motion quality for video, voice range for audio.

**Generating.** generate_image and generate_video accept asset ids to edit or reference existing images, including image-to-video. generate_audio takes the script plus a voice chosen through explore_models. Each spends credits and renders a live view that polls itself, so never poll manually and never tell the user you will check back.

**Retry spends again.** A regeneration costs credits a second time. The quality gate in this skill allows one automatic retry. Beyond that, say what is wrong and let the user decide.

**Finishing.** upscale_image and upscale_video raise an existing asset. Offer these instead of regenerating at a larger size.

**Recovering work.** get_assets lists the library newest first. get_generations fetches a generation by id. Use them to pick up earlier work instead of remaking it.

## Content Lock

Restoration repairs an image. It does not reinterpret it.

Preserve every element: faces and their identity, text and its exact characters, logos, colours, grain structure appropriate to the era, and composition. Do not add detail that was never photographed, do not invent a face behind a blur, do not complete a cropped edge, do not modernise clothing or setting, and do not colourise unless explicitly asked.

The most common failure is a restored photograph where the person is no longer the same person. Treat that as a hard failure, not a stylistic variation.

## Preconditions

An image file. Ask what the result is for, because the destination sets how far to push.

## Intake

Ask once, bundled: the image; what it is for, such as print, web, archive, or a larger creative piece; and whether it needs repair as well as enlargement. If the image shows visible damage such as creases, tears, stains, fading, or scan artefacts, name what you can see and confirm what they want removed.

## Diagnose before acting

Different defects need different treatment and the wrong order makes things worse.

- **Low resolution but clean**: upscale directly. Nothing to repair first.
- **Soft or out of focus**: upscaling will enlarge the softness. Say so. Sharpening a genuinely out-of-focus image invents detail and changes faces.
- **Compression artefacts**: blocking and ringing around edges. Address before enlarging, or the artefacts scale with the image.
- **Film grain or scan noise**: grain is often part of the photograph's character. Ask before removing it. Removing grain from a film photograph is a stylistic choice, not a repair.
- **Physical damage**: creases, tears, missing corners, stains. Repair is reconstruction of background, never of a face. If damage crosses a face, say explicitly which part will be reconstructed and let the user decide.
- **Fading and colour cast**: correct toward neutral, but preserve the era's palette. A nineteen seventies photograph should still look like one.

## Order of operations

Repair defects first, then enlarge. Enlarging first multiplies the defect.

For a heavily damaged photograph, do it in visible steps and show the intermediate result, so the user can stop at the point they prefer rather than receiving one over-processed final.

## Execute

Call upscale_image for enlargement and detail recovery. When repair requires reconstruction of damaged areas, call generate_image with the source as reference and an instruction limited strictly to the damaged region, naming what must remain untouched.

State the output size in the delivery line.

## Quality gate

Compare against the source at matching scale. Reject if a face reads as a different person, if text characters changed, if a logo altered, if skin became plastic, if edges show halo or oversharpening, or if invented detail appeared in an area that was simply out of focus.

Regenerate once with a lower degree of intervention. Less is the fix.

## Deliver

Return the image, the final dimensions, and one line naming what was repaired and what was deliberately left alone. If grain or character was preserved on purpose, say so, otherwise the user may read it as a failure.

## Stop conditions

| Condition | Response |
|---|---|
| No image supplied | Stop and ask for one |
| Damage crosses a face | Name what would be reconstructed, ask before proceeding |
| Image is out of focus rather than low resolution | Explain the difference, offer enlargement without invented detail |
| User asks to change content rather than repair it | Route to a generation skill |
| Face identity fails twice | Deliver the lower intervention result and say why |

Referenced files: 1

View saved version →

---
name: thumbnail
description: Design a thumbnail or cover image that stays readable at small size, for YouTube, podcasts, courses, blogs, or playlists. Trigger when the user asks for a thumbnail, a video cover, a podcast cover, a channel art piece, or an image to sit next to a title in a feed. Do not trigger for general poster or social graphic design, for product photography, for video, or for images where small-size legibility is not the point.
---

# Thumbnail

Design one image whose entire job is to win a click at roughly two hundred pixels wide.

## Morphix tool contract

These are the tools available and the order they run in. Follow it exactly.

**Getting the user's image in.** Chat attachments cannot be passed to generation. When a reference is needed, call show_upload_widget, which shows a drop box and returns an asset id once the user uploads. Never call upload_asset directly. If the user uploaded something earlier in the conversation, call get_assets and reuse that asset id rather than asking again.

**Before spending credits.** Call explore_models to choose a model for this job, and get_account to confirm the user has credits. Both are read-only and free. Choosing a model blind, or starting a generation the user cannot pay for, are the two most common wasted turns.

**Choosing a model.** Never hardcode a model name. Read the capabilities returned by explore_models and pick for the job at hand: reference fidelity when the result is anchored to a supplied image, text handling when legible type matters, motion quality for video, voice range for audio.

**Generating.** generate_image and generate_video accept asset ids to edit or reference existing images, including image-to-video. generate_audio takes the script plus a voice chosen through explore_models. Each spends credits and renders a live view that polls itself, so never poll manually and never tell the user you will check back.

**Retry spends again.** A regeneration costs credits a second time. The quality gate in this skill allows one automatic retry. Beyond that, say what is wrong and let the user decide.

**Finishing.** upscale_image and upscale_video raise an existing asset. Offer these instead of regenerating at a larger size.

**Recovering work.** get_assets lists the library newest first. get_generations fetches a generation by id. Use them to pick up earlier work instead of remaking it.

## The constraint that defines this skill

A thumbnail is not a small poster. It is viewed at a fraction of its size, next to competing thumbnails, often on a phone, frequently in dark mode. Every decision follows from that.

Three elements maximum. One focal point. Readable in one glance at thumbnail scale.

## Preconditions

Ask for the video or episode title and the topic. Without the title you cannot decide what the image must not repeat.

A source photo is optional. If the user supplies a face shot, use it and preserve the likeness.

## Intake

Ask once, bundled: the title; the platform; whether they have a face or product shot to include; and the channel's existing look if there is one. Never ask about resolution.

## Composition rules

**Subject scale.** The focal element occupies at least a third of the frame. A subject that reads at full size disappears at thumbnail size. When in doubt, go bigger.

**Face rules.** If a face is included: crop at chest or tighter, eyes in the upper third, one clear expression, lit so the face separates from the background. A face reading a single strong emotion outperforms a neutral one. Never manufacture an expression on a supplied photo of a real person; recompose and relight instead.

**Contrast and separation.** The subject must separate from the background by luminance, not only by colour. Test the idea in greyscale mentally: if subject and background merge, add a rim light, a dark vignette, or a flat colour field behind the subject.

**Colour.** Two dominant colours plus one accent. High saturation reads at small size. Avoid muddy mid-tones and avoid palettes that vanish against both light and dark interface backgrounds.

**Negative space.** Leave a deliberate clear zone where platform elements land: the lower right corner carries a duration badge on video platforms, and the lower edge can be overlaid by progress bars.

## Text in the thumbnail

Prefer no text. The title already sits next to the image, and repeating it wastes the frame.

If the user insists on text: three to four words maximum, one line if possible, heavy weight, very large, placed in the clear zone, never crossing the face. Text must add something the title does not say.

Because generated text is unreliable, deliver the image with the text zone left clear whenever the user can add type themselves, and say that is what you did and why.

## Write the prompt

Name the subject and its scale in frame, the expression if a face is present, the background treatment, the lighting that creates separation, the two-colour palette and accent, and the clear zone. State explicitly that the composition must read at small size.

## Execute

Call generate_image, sixteen-by-nine for video platforms, square for podcast and playlist covers.

## Quality gate

Judge the result as if it were two hundred pixels wide. Reject if the subject does not read instantly, if subject and background merge in luminance, if more than three elements compete, if any generated text is misspelled or garbled, or if a supplied face has been altered.

Regenerate once with a tighter crop and a larger subject.

## Deliver

Return the image and one line naming the focal point and the clear zone. Offer upscale_image if the user needs a larger master.

## Stop conditions

| Condition | Response |
|---|---|
| No title or topic given | Stop and ask for the title |
| User wants a long text block in the image | Explain the small-size constraint, offer three to four words |
| Supplied face is too low resolution to hold at crop | Ask for a better photo or offer a non-face composition |
| Generated text garbles twice | Deliver with the text zone clear and say why |

Referenced files: 1

View saved version →

---
name: ugc-video
description: Produce a creator-style video ad for a product, shot the way a real customer would film it rather than like a commercial. Trigger when the user asks for a UGC ad, a creator video, an influencer-style ad, a testimonial-style video, an unboxing, a review video, or a TikTok or Reels ad for their product. Do not trigger for polished product hero videos with no person, for animating a personal photograph, for written ad copy alone, or when the user has not supplied a product reference.
---

# UGC Video

Produce a short vertical video that reads as filmed by a customer on a phone, not produced by an agency.

## Morphix tool contract

These are the tools available and the order they run in. Follow it exactly.

**Getting the user's image in.** Chat attachments cannot be passed to generation. When a reference is needed, call show_upload_widget, which shows a drop box and returns an asset id once the user uploads. Never call upload_asset directly. If the user uploaded something earlier in the conversation, call get_assets and reuse that asset id rather than asking again.

**Before spending credits.** Call explore_models to choose a model for this job, and get_account to confirm the user has credits. Both are read-only and free. Choosing a model blind, or starting a generation the user cannot pay for, are the two most common wasted turns.

**Choosing a model.** Never hardcode a model name. Read the capabilities returned by explore_models and pick for the job at hand: reference fidelity when the result is anchored to a supplied image, text handling when legible type matters, motion quality for video, voice range for audio.

**Generating.** generate_image and generate_video accept asset ids to edit or reference existing images, including image-to-video. generate_audio takes the script plus a voice chosen through explore_models. Each spends credits and renders a live view that polls itself, so never poll manually and never tell the user you will check back.

**Retry spends again.** A regeneration costs credits a second time. The quality gate in this skill allows one automatic retry. Beyond that, say what is wrong and let the user decide.

**Finishing.** upscale_image and upscale_video raise an existing asset. Offer these instead of regenerating at a larger size.

**Recovering work.** get_assets lists the library newest first. get_generations fetches a generation by id. Use them to pick up earlier work instead of remaking it.

## What makes this different from a product video

A product video sells through craft: controlled light, clean surfaces, deliberate camera. A UGC video sells through the absence of craft. Handheld frame, ordinary room, uneven light, a real hand, a slightly off centre composition. If it looks produced, it fails as UGC.

Never apply studio lighting, never centre the product perfectly, never use a locked tripod frame, never grade it cinematically.

## Product Lock

The product from the reference is immutable: shape, proportion, colour, material, label geometry, typography, logo, printed text. It may be held, tilted, partially occluded by a hand, or slightly out of focus, which is normal in UGC. It may never be redesigned.

If label text is unreadable in the reference, let the hand or the angle hide that face. Never reconstruct it.

## People

Any person is auxiliary. Hands, a partial body, a shoulder, an over-the-shoulder view, or a face are all acceptable, but the product is the subject.

Never depict a real, identifiable person unless the user supplies their own photo and says it is them. Never imply a specific real person endorses the product. Generated people are fictional and must not resemble a public figure.

Never state or imply that the video shows a real customer, a real review, or a real result. If the user asks for a testimonial, produce it as clearly fictional demonstration and say so in the delivery line.

## Claims

Write no claim the user has not given you. No statistics, no before and after results, no medical, financial, or performance outcomes, no superlatives about the product. If the user supplies approved claims, use only their exact wording.

## Intake

Ask once, bundled: the product reference; the duration; the format, meaning whether they want spoken voiceover, on-screen captions, or neither; and any approved claims. Offer eight, fifteen and thirty seconds.

## Structure

Write the beats before generating anything.

- **Hook**, first two seconds. A problem stated, an unexpected visual, or the product already in hand. Never a greeting, never a logo, never a title card.
- **Context**, the ordinary setting that makes it believable.
- **Demonstration**, the product being used, held, opened, applied, or worked with.
- **Close**, the result or the reason to act.

For durations above fifteen seconds, plan two or three short shots rather than one long take. Long single takes drift and break the product.

## Look

Vertical nine-by-sixteen. Handheld micro-shake. Mixed domestic light, such as a window plus a lamp, not a softbox. Real background clutter, kept mild. Shallow phone-style depth rather than cinema bokeh. Slight overexposure or a mild colour cast is an asset here, not a defect.

## Audio and captions

Check explore_models for a video model that generates native speech. If one exists, use it and write the spoken lines in the prompt.

If none is available, say so plainly, deliver the video without speech, and produce the voiceover separately with generate_audio, returning both plus the script. There is no tool to combine them, so tell the user they will need to lay the track over the clip in their editor. Never imply the delivered file contains narration it does not.

Captions are not generated into the video. Text baked by an image or video model garbles. Deliver the caption text alongside so the user adds it in their editor.

## Execute

Call generate_video with the product asset id as reference, nine-by-sixteen.

## Quality gate

Reject if the product changed shape, colour, or label; if the result looks studio-lit or tripod-shot; if a hand shows the wrong number of fingers; if a person dominates over the product; if generated text appears in frame; or if a second product appears.

Regenerate once, moving further toward ordinary and handheld.

## Deliver

Return the video, the script if one was written, and the caption text if requested. State plainly that any person shown is fictional and that the video is a concept demonstration, not a real customer testimonial.

## Stop conditions

| Condition | Response |
|---|---|
| No product reference | Call show_upload_widget and wait |
| User asks for a real testimonial or real results | Offer a clearly labelled fictional demonstration instead |
| User asks for a named real person or lookalike | Decline in one sentence |
| User supplies claims you cannot verify | Write around them, say what was left out |
| Quality gate fails twice | Deliver the most believable take and say what drifted |

Referenced files: 1