SkillCDN 팩

Shorts-style AI ad from a reference video

One reference video and one product (a link is enough) go in; a captioned vertical ad, its clean master and a credit ledger come out. A user with no reference still gets one: it is found first, among the ads now running for that kind of product. The reference teaches what works (the hook, the rule its story runs on, the device that carries it, the reveal, the pacing as measured) and sets the medium and the look (live action or animation, in its style); the ad itself is written new, on the product's own world: no frame, sound, person or line of the reference is reused, and no take is made from the reference. Everything beyond the two inputs is derived and shown in plain words. Portraits and a still first frame per shot are approved before any video is paid for; the latest Kling or Seedance model animates one draft-quality take at a time and speaks every line itself; cuts, text, captions, effects and the end card are code in the sandbox. This skill is the whole workflow: no bundled Higgsfield workflow replaces it.

How the user is involved

• Questions: only for the product, and for the reference video when the request has none and nothing at hand can find one. Nothing else is ever asked; it is derived. A missing reference is found, not asked for, wherever a skill that researches running ads is at hand (analysis-brief.md "When there is no reference").
• Checkpoints: the research skill's own shortlist checkpoint when the reference is found, the concept directions' sheet from the variations skill when it is at hand (phase 3), then the plan with its cost, the portraits, the first frames, the clean master, the finished ad. Each is one message in plain words, as short as its content allows: what was made, the recommendation, and that one word ("OK" in the user's language) continues. Anything can be changed at a checkpoint. Takes are reported, not approved one by one.
• Go-ahead: when the user says to go ahead alone, the checkpoints after the cost are skipped; the cost is confirmed in every mode.
• Changes mid-run: applied from that point on; reported at the next checkpoint when within the accepted budget, re-quoted first when beyond it.

Requirements

All from the Higgsfield MCP server. Check that they are callable before the first message. If the server is not connected, stop and guide the user to connect the Higgsfield MCP server; if a tool is off, to enable it. Never substitute.

Optional: any web-reading tool for the product page's text; a skill that researches running ads, which finds the reference when the request has none. No text-to-speech. In a client without the upload widget everything comes in as links, or an agent with a shell uploads a local file itself (analysis-brief.md "Intake"). The references this skill links come with it when the server returns them; when only their list came, read each with read_repo_file before the phase that links it. How these tools behaved in real runs of this skill is in tool-notes.md; know it before phase 2.

What every Higgsfield skill of the repository shares (the sandbox, models and credits, portraits and frames, decoding a take, the sounds of each language) is in the repository's shared pages under docs/higgsfield/. pronunciation.md, which every run writes and reviews lines with, is included: it arrives with this skill wherever SkillCDN serves it. The other pages are linked from the phases that read them and read with read_repo_file through the connection at the repository root. Without that connection (a mount of the area or of this skill alone, a host that takes skills through the skills extension and has no read_repo_file, a plugin install, a copy of this directory) the linked pages cannot be read, and a plugin install or a copy lacks the included page too: say so in the first message, fetch what is missing where the agent can from docs/higgsfield/ of the repository this skill comes from (for this pack, github.com/skillcdn/skills, served at skillcdn.ai/gh/skillcdn/skills), and otherwise run on this skill's own files, which carry its workflow and its rules.

Inputs

Workflow

Each phase produces a named artifact. Phases stop only at the checkpoints above.

Phase 1: Intake

Produces the intake record.

1. Both inputs in the request: ask nothing, say in a line what happens next, go on. The product missing: ask for it as analysis-brief.md "Intake" shows, and stop.
2. Read the product page as "Product brief" there: name, what it is, tagline, claims (only these may be spoken), language, images, and its type and colors (design.md). The site's assets are the advertiser's and may be used.
3. The reference missing: find one as "When there is no reference" there says, in the market of the ad's language, and say so in a line; ask for it only when no skill at hand can find one.
4. Derive the rest and record it. The artifacts of every phase are written in the folder the user named for the run, else in a folder named for the ad outside any repository; the delivery says where.

Phase 2: Analyze the reference

Produces the analysis brief (template).

1. Import the reference. Start the scene analysis only for a reference uploaded through media_upload or given as a YouTube link, which is what the tool takes, and never wait for it (tool-notes.md "Scene analysis"). Run the sandbox pass in the background: probe, frames, contact sheet, transcript.
2. Fill the brief from frames and transcript: the medium and the look (live action or animation, and the style in words an image model can follow), mood, shots with their beats, dialogue, on-screen text and its look, cast, sound, and above all why it works: the hook in the first two seconds, the tension and the turn, the payoff, the call to action, the pacing that carries them. Then the story engine and the delivery by story.md: the form, the genre and register, the rule of the world, the device, the figure, the signature line, the information order line by line, the reveal; and the measured speech rate, lines per ten seconds, longest pause and hook window from the sandbox pass.
3. Mark build routes (generate, edit, drop) by editing-decisions.md.
4. A reference this skill found comes with the research's breakdown of it and its readings of the market. They are a first look and they feed the concept, as analysis-brief.md "When there is no reference" says; the sandbox pass still runs and the brief is still filled.
5. Nothing of the reference becomes a model input or a piece of the output. Open decisions go to the plan checkpoint; there is no checkpoint here.

Phase 3: Concept and plan

Produces the concept, the cast list, the shot list (example) and the edit plan.

Before the story, the directions: by the repository's rule on starting from a reference, the pack's web-concept-variations runs on the reference where it is at hand (the mechanism as the analysis brief reads it, the product as the subject), stopping at its own sheet, and the user's pick is the story's rule, device, figure and hook as the variation set's handoff.md "By producing skill" says; the reference's own mechanism carried straight is the second offer, taken when the user says so or when that skill is not at hand, said so in the plan. On a go-ahead the recommended direction is taken. Nothing is spent here: the directions are text.

1. Story. By story.md "Writing the new story": the rule of the world taken from what the product itself claims to know or do, in its own vocabulary; a device from the product's world; the figure (the site's own persona when it has one); five hooks scored against the table and one kept, with the runner-up; the signature line in the figure's voice; the reveal as one flat fact with the sentence the viewer thinks next; the lines in order, each adding one fact, at the reference's measured density. This is the concept: a different situation, setting, cast and lines that deliver the same mechanism in the reference's medium and look. The reference's surface (its scenes, its device, its wording, its cast) is not reused, and a story that would work for any product in the category is rewritten. Write one alternative (another rule or device) in one line for the checkpoint. Only when the user asks for a close remake is the reference followed scene by scene, and the plan says so.
2. Style line. One sentence from the brief's medium and look that fixes the style for every image and take prompt: the medium, line, color and shading, backg