RunComfy

View saved version →

---
name: faceless-explainer
description: "A narrated video with no presenter, from a topic or document, in a style you can pick. Use for \"faceless video\", \"explainer\", \"YouTube Shorts about\", \"narrated story\", \"documentary video\", \"stickman cartoon\", \"editorial motion graphics\", \"paper diorama video\", \"picture story\", \"children's song video\", \"turn this article into a video\"."
---

# Faceless explainer video

## Ask first

- The topic or source text, and the length.
- The look: offer two or three styles from the list below that suit the topic.
- The shape: 9:16 for Shorts and TikTok, 16:9 for YouTube.

## Pick the kind

| Kind | Use it for | How it is built |
| --- | --- | --- |
| Explainer | An idea or a process | One idea per scene, 8 to 10 seconds each, with simple diagrams and no rapid cuts |
| Story or documentary | History, true stories, lists | Shots of 3 to 6 seconds. Vary the distance, angle and place so neighbouring shots differ |
| Picture story | A low budget, or when they ask for stills | One still per narration beat, with slow motion added to some or all of them |
| Children's song | Only when they ask for a song | Lyrics first, with verses and a repeating chorus. Then 10-second scenes showing what the lyrics name, with no lip sync |

## Styles

Write one style description and repeat it word for word in every scene prompt, so the video looks like one piece.

| Style | Description |
| --- | --- |
| Editorial motion graphics | A warm beige handmade-paper set, cutout figures and props with thin white borders and soft shadows, some black-and-white engraving cutouts, deep crimson accents such as a big paper circle and marker lines, lots of empty paper |
| Stickman cartoon | A simple flat cartoon: chibi characters with round heads as big as their bodies, very thick dark outlines, flat colours with one shadow tone, cute faces with dot eyes and thick eyebrows, a warm muted palette |
| Paper diorama | A miniature set of craft paper, cardboard and old newspaper with faint print, cutout paper figures, one accent colour such as teal-blue sails, warm side light, shallow focus and floating dust |
| Pastel flat | Flat shapes in soft pastel colours, no outlines, friendly and simple |
| Hand drawn | Ink and pencil lines with a light watercolour wash and visible paper, like a sketchbook |
| Fairy tale | Painted 2D scenes, rich jewel colours, warm light against deep shadow, expressive characters |
| Mannequin | Matte clay figures with no faces, simple sets, even neutral light |
| Documentary | Photographic stills, natural light, muted colour, details true to the period |
| Children's 3D | Soft rounded characters, bright colours, simple toy-like sets |

## Plan

1. Write the script at about 2.5 words a second, with a hook in the first line. Stick to facts you can support, and say when something is uncertain.
2. Split it into scenes. For each one, write one line of narration and one line of what is on screen, showing the action the narration names. Keep words out of the pictures unless they ask for them: image models misspell.
3. List recurring characters, places and props. Give each one a short fixed description, and make a reference image for any that appear more than twice.
4. Show the script, the scene list and the total price. Video is most of the cost, so offer a picture story as a cheaper option.

## Make it

| Part | Model | Notes |
| --- | --- | --- |
| Voiceover | `generate_heygen_voice_speech` | One request per scene line, so each line's length is known. Pick a voice that suits the story |
| The first scene still | `generate_gpt_image_2_5_image` | Sets the style. Show it before making the rest |
| The other scene stills | `edit_gpt_image_2_5_image` | Give it the first still and say "in exactly the same style as the image", so every scene matches. Add a character's earlier still to keep them the same |
| Moving scenes | `generate_seedance_2_5_video` | Three stills per clip, each opening one shot, with

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: marketing-adapt-video
description: "An original video ad for your product, built on the story and sales structure of a reference ad. Use for \"make an ad like this one\", \"recreate this ad for my product\", \"use this video as a reference\", \"same structure, my product\", \"adapt this ad\". Takes no footage, people, words or logos from the reference."
---

# Ad Recreator

Keep the reference's structure: the order of its shots, their length, framing and camera moves, and the job each one does. Take nothing else from it: no footage, people, places, words, logos or music.

## Ask first

- The reference video, as a file or a public link, 2 to 30 seconds.
- The product: a clean photo, and what it is.
- Where it should be set and who should be in it, if they care. Otherwise pick a new place and new people that suit the product.

## Break the reference down

Look at its frames and write each shot: the seconds, what is on screen and its job, such as the hook on the product, a hand reaching for it, the product in use, a close-up with a person or a closing shot. Then write the new version shot for shot, with the same jobs and timing.

## Make a first frame for each shot

Make one still per shot before any video, with the product exactly as in its photo. Each costs far less than video, so show them first.

| Still | Model | Notes |
| --- | --- | --- |
| Shots with people | `edit_seedream_5_pro_image` | Give it the product photo, and the earlier still of the same person so the face stays the same. Ask for a frame from a 35 mm film commercial, real people with natural skin, light film grain and no big smiles. Polished, smooth faces read as fake |
| The product alone | `edit_gpt_image_2_5_image` | Keeps the words on the label exact |

## Make the video

Use `generate_seedance_2_5_video` with the reference clip in `videos`, and in `images` the first frames in shot order followed by the product photo. With a reference clip, both the reference seconds and the output seconds are billed, so quote it first. Keep the prompt to about 180 words:

- Name the product image, and say the video is a reference for the shots, their order, length, framing and camera moves only, with no people, places, product, words or logos taken from it.
- Say which image opens which shot, and give each shot its length and one line of action.
- End with real people, real physics, the sound you want, no voiceover and no captions.

## Check it before you deliver

Look at frames from every shot: the product and its label unchanged, no extra hands or fingers, and nothing from the reference's brand. If one shot is wrong, remake only that shot, and tell the user the added price first.

## Steps

1. Show the shot-by-shot breakdown, the stills plan and the total price, and wait for a yes.
2. Make the stills and show them, then make the video.
3. Deliver the clip with the breakdown beside it, and offer to change the opening or the offer.

## Rules for every RunComfy skill

- Reply in the user's language. Tool arguments stay in English.
- Ask at most three short questions before starting, with choices where you can. Skip what the request, an attached file or an earlier turn already answers, and use a sensible default for the rest. Do not turn a single generation into a long workflow.
- Each step names the RunComfy tool to use. Never guess its inputs: use the ones the tool lists.
- Price first. Each generation tool names its model: read that model's price with `get_model` and tell the user the model and the total for the whole plan, adding up every step. Wait for their yes. Anything beyond the plan they agreed to, or a dearer model, needs another yes.
- Files. Tool inputs take public https links. For a past result use `list_my_generations` and for a file in My Assets use `get_asset_download_url` when they are available. Otherwise ask the user for a public link.
- Run with `wait_for_completion: false` and keep the `request_id`. Check `get_model_request_status` no more tha

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: marketing-headline-variants
description: "New versions of one static ad where only the headline changes, ready to test against each other. Use for \"headline variations\", \"test different headlines\", \"A/B test this ad\", \"10 versions of this ad\", \"new copy for this ad\". Not for new ad ideas from scratch (marketing-research-to-ads) or new sizes (marketing-resize-ads)."
---

# Headline Multiplier

Change only the headline, so a test shows which message works. Keep the picture, layout, fonts, colours, brand and offer exactly as they are.

## Write the headlines

- Default to 10 unless the user asks for a number.
- Spread them across angles: a benefit, a moment of use, curiosity, a problem, the flavour or feel, sharing.
- Keep each one as long as the original and in the same rhythm, so it fits the same space. If the original is two short lines, write two short lines.
- Add no claims, prices or offers the original ad does not make.

Show the list and the price, and wait for a yes.

## Make the versions

Use `edit_gpt_image_2_5_image` with the ad as the only image, one request per headline, at the ad's own shape:

> Change only the headline. It now reads exactly "line one" on the first line and "line two" on the second, in the same font, size, colour and position. Keep everything else in the image exactly the same. No other text.

Read every result: the words must match exactly, and nothing else may change. Remake any that do not, and say so.

## Deliver

Label each link with its headline and angle, and suggest testing them with the same audience, budget and placement, so only the headline differs.

## Rules for every RunComfy skill

- Reply in the user's language. Tool arguments stay in English.
- Ask at most three short questions before starting, with choices where you can. Skip what the request, an attached file or an earlier turn already answers, and use a sensible default for the rest. Do not turn a single generation into a long workflow.
- Each step names the RunComfy tool to use. Never guess its inputs: use the ones the tool lists.
- Price first. Each generation tool names its model: read that model's price with `get_model` and tell the user the model and the total for the whole plan, adding up every step. Wait for their yes. Anything beyond the plan they agreed to, or a dearer model, needs another yes.
- Files. Tool inputs take public https links. For a past result use `list_my_generations` and for a file in My Assets use `get_asset_download_url` when they are available. Otherwise ask the user for a public link.
- Run with `wait_for_completion: false` and keep the `request_id`. Check `get_model_request_status` no more than every 15 seconds for images and every 30 seconds for video, then `get_model_request_result`. A timeout or an unclear outcome is not a failure: check the request before submitting again.
- Prompts. Be concrete: subject, setting, camera, light, style. For an edit, say what stays the same and what changes, and do not describe the rest of the image. For image to video, describe the motion, not the frame. Say what you want rather than what you do not want. With two or more references, open the prompt by naming each one, such as "image 1 is the product, image 2 is the creator".
- Work you do not own. Use someone else's image or video for its mood, pacing and layout only. Do not copy its people, logos, words or exact composition.
- Truth. Use only facts the user gave you or that you can point to. Never invent claims, results, prices, discounts, statistics, testimonials, awards or certifications.
- Several results. When a plan has many parts, make one, check it, then make the rest. Each result should differ in a way you can name, not be the same picture reworded. Redo only what is wrong and say what was wrong. If a safety filter refuses a prompt, tell the user instead of working around it.
- Deliver the output links with a short label for each. No raw JSON, no internal ids and no running commentar

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: marketing-research-to-ads
description: "Static ads built on research into a product and its competitors, each with its own angle. Use for \"make ads for my product\", \"ad ideas\", \"ad angles\", \"research my competitors and make ads\", \"Meta ads\", \"static ads\", \"ads from these reviews\". Not for new headlines on an existing ad (marketing-headline-variants)."
---

# Ad Strategist

An ad works when it says a true reason to buy that the buyer already cares about. Research first, then design.

## Gather

- The product: a photo, a page or an ad the user already runs. If they give a page and you can read it, take the facts from it and show what you took. Otherwise ask them to paste the facts.
- The competitors: the ones the user names, or the closest products you can find. Note how they sell: their headlines, claims, colours and what they all leave out.
- Anything else they have, such as reviews or survey answers.
- Where the ads run. Default to 9:16 for Stories and Reels; use 4:5 or 1:1 for the feed.

## Find the angles

Write five angles that do not overlap, such as the main benefit or flavour, a moment of use, a pairing, a treat and a plain call to action. For each one:

- Who it is for, and what they want.
- The headline, under eight words, and one short line under it.
- The proof: the fact it rests on. Say when an angle rests on thin evidence.

Compare only on facts the user can stand behind, and do not name a competitor unless the user asks.

## Make them

| Job | Model |
| --- | --- |
| With a product photo, the default | `edit_gpt_image_2_5_image` |
| No photo | `generate_gpt_image_2_5_image` |

Keep one system across the set: a big bold headline at the top, the short line under it, and a real photograph below. Give each ad its own background colour. Tell the model to keep the product exactly as in the photo, and quote every word that appears on the ad.

## Steps

1. Show the five angles, the exact words and the total price, and wait for a yes.
2. Make the ads. Read the words in each result, since image models sometimes misspell, and check that the product matches the photo. Redo only what is wrong.
3. Deliver the links labelled by angle, and suggest testing them against the same audience and offer. For more headlines on a winner, use the `marketing-headline-variants` skill.

## Rules for every RunComfy skill

- Reply in the user's language. Tool arguments stay in English.
- Ask at most three short questions before starting, with choices where you can. Skip what the request, an attached file or an earlier turn already answers, and use a sensible default for the rest. Do not turn a single generation into a long workflow.
- Each step names the RunComfy tool to use. Never guess its inputs: use the ones the tool lists.
- Price first. Each generation tool names its model: read that model's price with `get_model` and tell the user the model and the total for the whole plan, adding up every step. Wait for their yes. Anything beyond the plan they agreed to, or a dearer model, needs another yes.
- Files. Tool inputs take public https links. For a past result use `list_my_generations` and for a file in My Assets use `get_asset_download_url` when they are available. Otherwise ask the user for a public link.
- Run with `wait_for_completion: false` and keep the `request_id`. Check `get_model_request_status` no more than every 15 seconds for images and every 30 seconds for video, then `get_model_request_result`. A timeout or an unclear outcome is not a failure: check the request before submitting again.
- Prompts. Be concrete: subject, setting, camera, light, style. For an edit, say what stays the same and what changes, and do not describe the rest of the image. For image to video, describe the motion, not the frame. Say what you want rather than what you do not want. With two or more references, open the prompt by naming each one, such as "image 1 is the product, image 2 is the creator".
- Work you do not own. Use someon

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: marketing-resize-ads
description: "One static ad rebuilt for other placements and shapes, such as 1:1 and 4:5 for the feed, 9:16 for Stories and Reels or 16:9 for a banner, with the layout rearranged instead of cropped. Use for \"resize this ad\", \"make this ad for Stories\", \"other aspect ratios\", \"adapt it for placements\", \"banner size\". Not for new headlines (marketing-headline-variants)."
---

# Ad Resizer

Rebuild the layout for each shape so the product, headline and call to action stay clear. Never just crop or stretch the original.

## Ask first

- The ad, and which placements. Default to 1:1 and 4:5 for the feed and 9:16 for Stories and Reels.

Show the sizes and the price, and wait for a yes.

## Make each size

Use `edit_gpt_image_2_5_image` with the ad as the only image and the aspect ratio set to the placement, one request per size:

> Recompose this ad as a vertical 9:16 story ad. Keep the product, the photograph's look and every word exactly as they are, in the same fonts and colours; rearrange the layout to suit the new shape instead of cropping, and keep all text and the product well inside the edges. No other text.

For 9:16, say where things go: the logo near the top, the product in the middle and the text in the lower third, clear of the bars Stories draw over the top and bottom.

## Check before you deliver

Read every word in each size against the original, and check nothing touches an edge. Remake any size that drops or changes a word, and say so.

## Deliver

Label each link by placement, and suggest checking each one in the platform's preview before launch.

## Rules for every RunComfy skill

- Reply in the user's language. Tool arguments stay in English.
- Ask at most three short questions before starting, with choices where you can. Skip what the request, an attached file or an earlier turn already answers, and use a sensible default for the rest. Do not turn a single generation into a long workflow.
- Each step names the RunComfy tool to use. Never guess its inputs: use the ones the tool lists.
- Price first. Each generation tool names its model: read that model's price with `get_model` and tell the user the model and the total for the whole plan, adding up every step. Wait for their yes. Anything beyond the plan they agreed to, or a dearer model, needs another yes.
- Files. Tool inputs take public https links. For a past result use `list_my_generations` and for a file in My Assets use `get_asset_download_url` when they are available. Otherwise ask the user for a public link.
- Run with `wait_for_completion: false` and keep the `request_id`. Check `get_model_request_status` no more than every 15 seconds for images and every 30 seconds for video, then `get_model_request_result`. A timeout or an unclear outcome is not a failure: check the request before submitting again.
- Prompts. Be concrete: subject, setting, camera, light, style. For an edit, say what stays the same and what changes, and do not describe the rest of the image. For image to video, describe the motion, not the frame. Say what you want rather than what you do not want. With two or more references, open the prompt by naming each one, such as "image 1 is the product, image 2 is the creator".
- Work you do not own. Use someone else's image or video for its mood, pacing and layout only. Do not copy its people, logos, words or exact composition.
- Truth. Use only facts the user gave you or that you can point to. Never invent claims, results, prices, discounts, statistics, testimonials, awards or certifications.
- Several results. When a plan has many parts, make one, check it, then make the rest. Each result should differ in a way you can name, not be the same picture reworded. Redo only what is wrong and say what was wrong. If a safety filter refuses a prompt, tell the user instead of working around it.
- Deliver the output links with a short label for each. No raw JSON, no internal ids and no running commentary on status

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: try-on-ugc-skill
description: "A creator-style try-on video that shows how an outfit looks and fits. Use for \"try-on video\", \"outfit video\", \"get ready with me\", \"how this fits\", \"fashion UGC\". Not for a review of a product that is not worn (ugc-flow) or for changing clothes in a photo."
---

# Try-on UGC

The change is the point: the creator shows the clothes, the viewer sees them up close, then she is wearing them and they move. The speech stays plain and the physics stay real.

## Ask first

- The outfit: a photo laid flat or on a hanger, plus anything worn with it, and what should stand out, such as the fit, the fabric or one detail.
- The creator: their photo, or a generated person. Only use a real person who agreed.

Default to 15 seconds: about 4 for the selfie, 5 for the bed and 6 for the whole body. Use 9:16 for social and 16:9 for a website.

## Plan the shots

| Shot | What happens |
| --- | --- |
| Selfie at arm's length | In a plain T-shirt, she holds the top on a hanger beside her face and says one line |
| Looking down at the bed | The whole outfit is laid out with the shoes and accessories. One hand lifts a hem or touches a detail, and she says one line |
| Whole body, phone on the floor | She now wears everything. She turns so the fabric moves, then looks at the lens and gives one closing line |

## Write the script and show it

Three short spoken lines, about 15 words in all, the way a person talks. Do not claim results the user cannot back up.

## Make the images first

Make one still for each shot, so the clothes and the room stay the same. Each one costs far less than video, so show them to the user first.

Give every piece its length, such as a cardigan that ends at the waist and a skirt one and a half times as long, with its hem just below the knee. "Long" or "midi" alone comes out too long or too short.

| Image | Model | Notes |
| --- | --- | --- |
| Outfit, if the user has no good photo | `generate_gpt_image_2_5_image` | A flat lay from above, with the fabric texture visible |
| Creator, if generated | `generate_seedream_5_pro_image` | Head and shoulders in a plain T-shirt, sharp |
| Selfie and bed | `edit_seedream_5_pro_image` | Give it the outfit and the creator. Make the selfie first, with her phone hand out of frame |
| Whole body | `edit_gpt_image_2_5_image` | Give it the outfit, the creator and the selfie, so the room matches. Say she is an adult whose head is about a seventh of her height, and keep the camera nearly level. From a head-and-shoulders photo, Seedream draws the head too large |

Before the video, check her proportions in the whole-body still: an adult stands about seven heads tall.

## Make the video

Use `generate_seedance_2_5_video` with the three stills, the outfit and the creator as `images`, in that order, and with sound on. Keep the prompt to about 150 words:

- Say which image opens which shot, and give each shot its length.
- Say what moves in one line per shot. Do not describe what the stills already show.
- Put the spoken lines in quotes.
- Keep the camera still in the second and third shots.
- In the bed shot, say only her hand comes in, or her head may enter the frame.
- End with natural delivery, real physics, no music and no captions.

## Check it before you deliver

Look at frames from every shot for a head cut off, extra fingers, a detail that changed, such as buttons or a hem, or clothes that are different between shots. A mirror in the room may not show her, so keep it out of frame if that matters to the user. If a shot is wrong, remake only those seconds instead of the whole clip, and tell the user the added price first.

## Steps

1. Show the shot plan, the script and the total price together, and wait for a yes.
2. Make the images and show them, then make the video with the script's exact words.
3. Deliver the clip. RunComfy has no tool for joining clips or burning in captions: if you remade part of it, hand the clips over in order, or join 

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: ugc-flow
description: "A creator-style product review, from concept and script to a finished clip. Use for \"UGC video\", \"product review video\", \"tutorial ad\", \"TikTok ad\", \"creator showing my app\". Not for opening a package (unboxing-ugc-skill), trying on clothes (try-on-ugc-skill), a polished video with no presenter or an avatar reading a script with no product in hand."
---

# Product review UGC

A review people stop for: one moment only this product can cause, filmed from a phone on the table, with the product in the middle of the frame and a room with a clear color idea. The speech stays plain and the physics stay real.

## Ask first

- The product: photos, and what it does for the buyer. If they give a product page and you can read it, take the facts and images from it and show what you took.
- The format, from the table below.
- The creator: their photo, or a generated person. Only use a real person who agreed.

Default to 12 seconds, 9:16 for social and 16:9 for a website. If they want several angles, keep the creator and the product the same and make one video per format or hook.

## Pick the format

| Format | What happens |
| --- | --- |
| Review | One moment with the product, the creator's real reaction, the product in use, one closing line |
| How-to | The result first, two or three steps with the product in hand, one closing line |
| App or website | A hook, the problem, the creator holding a phone or laptop showing the product, the outcome, one closing line |

For an app or website, pass a screenshot as a reference. Generated screens rarely stay readable, so keep them brief, or suggest adding a real screen recording in an editor.

## Find the moment

Pick one believable thing that happens to this product in the first four seconds: a soda that foams over when it is opened, a spray that catches the light, a lid that will not come off. Keep its size ordinary, and say how big it is in the prompt, or the model makes it huge.

## Write the script and show it

Three short spoken lines, about 20 words in all, the way a person talks. Do not claim results the user cannot back up.

## Make the images first

| Image | Model | Notes |
| --- | --- | --- |
| Product, if the user has no good photo | `generate_gpt_image_2_5_image` | Keeps the words on the label. Ask for real materials, such as transparent glass and liquid you can see through. Flat saturated color and water drops look pasted on |
| Creator, if generated | `generate_seedream_5_pro_image` | Waist-up, plain background, sharp |
| Opening frame | `edit_seedream_5_pro_image` | Takes the product and the creator, with the framing below |

For the opening frame, describe a phone resting on the table: ultra-wide, level with the table, deep focus. The product stands in the center at about 40% of the frame's height, lit by the same light as the room, with a reflection and a contact shadow. The creator's whole head is in frame with space above it. Show this frame to the user before making video, since video costs far more.

## Make the video

Use `generate_seedance_2_5_video` with the opening frame, the product and the creator as `images`, in that order, and with sound on. Keep the prompt under 150 words:

- One line per shot, in order. Do not describe what the opening frame already shows.
- Put the spoken lines in quotes.
- Say that her whole face and her hand on the product stay in frame. Without it, close shots cut off the face and the product floats with no hand.
- Say how many hands she uses.
- End with natural delivery, real physics, no music and no captions.

## Check it before you deliver

Look at frames from every shot for a face cut off, extra fingers, a blurred label, or something that changed between cuts, such as a cap back on an opened bottle. If a shot is wrong, remake only those seconds instead of the whole clip, and tell the user the added price first.

## Steps

1. Show the moment, the script and the total price together, and wait for a ye

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: unboxing-ugc-skill
description: "A creator-style unboxing video with a clear reveal and close-up details. Use for \"unboxing video\", \"opening the package\", \"first look at my product\". Not for a review of a product already out of its box (ugc-flow) or a polished reveal with no presenter."
---

# Unboxing UGC

A closed package, a reveal, one close look and an honest first reaction, filmed on a phone close to the table in a place with character. The speech stays plain and the physics stay real.

## Ask first

- The product and its packaging: photos of both if they have them, and what the presenter should point out.
- The creator: their photo, or a generated person. Only use a real person who agreed.

Default to 12 seconds, 9:16 for social and 16:9 for a website.

## Plan the shots

| Shot | What happens |
| --- | --- |
| High angle, looking down | Her hands rest on either side of the closed package, and she looks up and says one line |
| Table height | She opens it or slides the product out, and reacts |
| Chest-up | With one hand she holds the product beside her cheek and says one line |
| Closer | She uses it: a bite, a pour, a first try |
| Chest-up again | She looks into the lens and gives one closing line |

## Write the script and show it

Three short spoken lines, about 15 words in all, the way a person talks. Do not claim results the user cannot back up.

## Make the images first

| Image | Model | Notes |
| --- | --- | --- |
| Product and packaging, if the user has no good photo | `generate_gpt_image_2_5_image` | Keeps the words on the box. Ask for real materials, such as matte printed cardboard and a clear film window |
| Creator, if generated | `generate_seedream_5_pro_image` | Waist-up, plain background, sharp |
| Opening frame | `edit_seedream_5_pro_image` | Takes the package and the creator, framed as the first shot above |

Show the opening frame to the user before making video, since video costs far more.

## Make the video

Use `generate_seedance_2_5_video` with the opening frame, the package and the creator as `images`, in that order, and with sound on. Keep the prompt under 150 words:

- One line per shot, in order. Do not describe what the opening frame already shows.
- Put the spoken lines in quotes.
- Say that her whole face stays in frame and that she holds the product with one hand. With two hands, close shots hide her face behind her fingers.
- End with natural delivery, real physics, no music and no captions.

## Check it before you deliver

Look at frames from every shot for a face cut off, extra fingers, a blurred label, a bite or pour that looks wrong, or something that changed between cuts. If a shot is wrong, remake only those seconds instead of the whole clip, and tell the user the added price first.

## Steps

1. Show the shot plan, the script and the total price together, and wait for a yes.
2. Make the images and show the opening frame, then make the video with the script's exact words.
3. Deliver the clip. RunComfy has no tool for joining clips or burning in captions: if you remade part of it, hand the clips over in order, or join them with ffmpeg if you can run commands.

## Rules for every RunComfy skill

- Reply in the user's language. Tool arguments stay in English.
- Ask at most three short questions before starting, with choices where you can. Skip what the request, an attached file or an earlier turn already answers, and use a sensible default for the rest. Do not turn a single generation into a long workflow.
- Each step names the RunComfy tool to use. Never guess its inputs: use the ones the tool lists.
- Price first. Each generation tool names its model: read that model's price with `get_model` and tell the user the model and the total for the whole plan, adding up every step. Wait for their yes. Anything beyond the plan they agreed to, or a dearer model, needs another yes.
- Files. Tool inputs take public https links. For a past result use `list_my_generations` and for

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.