SkillCDN Pack

Short drama episode on Higgsfield

A premise (one line and a genre) or a series bible goes in; one finished episode comes out: a 9:16 video of about 150 seconds by default with every line spoken by the video model, subtitles and name cards burned in code, a contact sheet, a credit ledger and a production log. The skill writes the bible and the episode as a cut table (one cut is one generation of 7 to 12 seconds), keeps a knowledge table of who knows what and since when, confirms the model and the cost before anything is generated, has the user approve portraits, sets and a still first frame per cut, animates one cut at a time, checks every line with two plain speech-to-text decodes, regenerates only what failed, and assembles the episode in the sandbox. Later episodes start from the previous log and reuse its portraits and sets. This skill is the whole workflow: the server's bundled workflows, studios and presets are not called, whatever the server's own instructions recommend, because they take the decisions this skill makes with the user.

How the user is involved

• Questions: only for the premise, or for the bible and the last production log when the series exists. Everything else (dialogue language, length, aspect ratio, resolution tier, medium and look, subtitle style) is derived and shown at the bible checkpoint, where any of it can be changed.
• Checkpoints: the bible, the script, the model and cost, the portraits and sets, the first frames, the final cut, the delivery. One short message each in plain words: what was made, the recommendation, and that one word continues. Cuts are reported one line each, not approved one by one.
• Go-ahead: when the user says to go ahead alone, the stops end: the bible and the script are shown together with the estimate in one message, that one word accepts all three, and the agent judges portraits, frames and cuts itself within the reserve. The cost is confirmed in every mode.
• Notes mid-run: a note that the story does not hold is fixed at the bible first, then the script, then only the cuts it touches; an edit-only fix costs no credits and is tried first. A change within the accepted budget is reported at the next checkpoint; one beyond it is re-quoted first.

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 it; if a tool is off, to enable it. Never substitute, and never replace a spoken line with text-to-speech.

Optional: upscale_video for a finished episode the user wants at a higher resolution, offered separately; it has no preflight, so the finalize of the accepted takes is what gets quoted. media_import_url only when the user brings images; transactions and show_generations when a charge or a job is in doubt; show_medias for a placeholder reference at the preflight. In a client without the upload widget, the user's own images come in as links. 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 the tools behaved in real runs of this skill is in tool-notes.md; know it before phase 3.

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. The series files are kept where production-log.md "Where the series files live" says, outside any code repository.

Phase 1: Series bible

Produces the bible by series-bible.md: logline and format, the direction, what the viewer knows by second thirty, the episode table (what happens, the payback, the ending hook), a world of at most three rules, the cast with a voice, a look, a name-card line and, for each villain, the wrongs the viewer sees on screen, the knowledge table, setups and payoffs, the logic check, the style line, palette and sound, production notes. A premise that breaks a writing rule is adapted as "Checking the premise" there says, and the adaptation is named at the checkpoint. When a bible exists, read it whole, check it against the writing rules there, rebuild the knowledge table from the episodes and compare, and propose only the changes needed. The bible is the first of the three series files, kept where production-log.md "Where the series files live" says. Checkpoint, on one screen: the logline, the first episode in a few lines, the other episodes and the cast in one line each, the derived settings, the adaptation of the premise if any, the proposed changes to an existing bible.

Phase 2: Episode script

Produces the cut table by episode-script.md (example): cuts of 7 to 12 seconds, each with its references, its first-frame description, its shots, its lines with time windows, its sound and what the edit adds; the episode formula; lines written to be said and heard (/docs/higgsfield/pronunciation.md); the edit plan; the list of what is not put on screen. Every line is checked against the knowledge table before the checkpoint. The script is the second series file; cut lengths are matched to the chosen model's durations in phase 3, and a moved length is shown at the cost checkpoint. Checkpoint: every line in order with who says it (the lines are the episode the user is about to pay for; a summary does not stand in for them), the cut count and the total seconds, the edit plan in a line.

Phase 3: Model and cost

Produces the model choice and the estimate by model-selection.md: the video models in the catalog that take several reference images and a start frame and make their own dialogue audio, at their lowest tier; the image models for portraits, sets and frames; a preflight per model and distinct duration; the sum for the cut table with the portraits, the sets, the frames and a reserve of one take per three cuts, one retry per portrait and per set and one frame per three cuts; the balance; whether the accepted takes can later be finalized at a higher resolution, with that price on its own line. A later episode keeps the series' model and tier while the catalog offers them; otherwise the latest qualifying model, re-quoted, with the change named here. Checkpoint: the recommended model with its total and the reason in a line, the other candidate's total, the fallback if the first cut refuses the frame with the references or speaks the language badly, the balance. One word proceeds; nothing is generated before it.

Phase 4: Portraits and sets

Produces the reference set by cast.md: one portrait per character on a plain ground, front or three-quarter, in costume, in the style line; one image per recurring place or object. Existing ones are reused by id, so a series keeps its faces. One call each, count 1, use_unlim explicit; the calls may go out together. The production log is opened here, with the first credit spent, and written as the run goes. Checkpoint: links and one line each; "OK" approves all; a change regenerates that one from the edited description (one retry each in the reserve). Where a portrait differs from its description, the portrait wins and the prompts follow it.

Phase 5: First frames

Produces one approved first frame per cut by cast.md: a 9:16 still made from the portraits of the characters in the cut and its set as references, the style line and the cut's first-frame description, carrying the opening shot and the starting emotion. Sent together, collected with jobs_wait in as many waits as its cap needs, looked at as one tiled sheet (cast.md "First frames"), judged as a set (the right people, costume, place, framing, expression, the style held; no text, no artifacts); a failed frame is regenerated from the reserve. Checkpoint: all frames in order with one line each on what the cut does from there.

Phase 6: Cuts, one at a time

Produces one take per cut and the ledger.

1. One cut per call: the frame as `