shortform-format-library
GitHub构建短视频格式库,评估内容素材并匹配录制大纲与编辑简报。支持格式发现、页面刷新、内容排名及引流物料准备,确保源证据与创作意图分离。
Trigger Scenarios
Install
npx skills add ericosiu/ai-marketing-skills --skill shortform-format-library -g -y
SKILL.md
Frontmatter
{
"name": "shortform-format-library",
"description": "Build and maintain a visual short-form format library, evaluate supplied content spikes, and match source-grounded ideas to recording outlines and editing briefs. Use for format discovery, refreshing a swipe page, ranking a content table, or preparing a format-library lead magnet."
}
Shortform Format Library
Help a creator move from an idea to a usable delivery structure. Keep source evidence, editorial interpretation, and the creator's own results distinct.
When running inside AI Marketing Skills, use the repository's optional version check and telemetry conventions in repository integration. Standalone use must work without that repository.
Choose the requested work
- Read the existing library, audience, creator roster, source policy, and supplied ideas or analytics before changing anything. Preserve the user's current page and saved outline IDs. Resolve the exact library and publishing destination from user context; do not substitute another account.
- For content spikes or discovery, read research and learning. Analyze supplied evidence before selecting candidates. If an attachment is absent, say so and continue independent work; never invent its contents.
- For ideas and format selection, use the default content table below, including a separate Watch reference column. Keep the idea source separate from the video demonstrating its format. A hook is proposed copy unless verified as an exact quotation. Keep new recordings separate from existing cutdowns and refresh angles. Score only when requested or required by the user's established workflow; label editorial scores as judgments.
- For a visual page, read page and data contract. Let users enter an idea or select a supplied one, compare two relevant format references, and turn the chosen structure into an editable three-prompt outline. Match by the available proof, story, recording effort, and audience, not trendiness alone.
- For editing or lead delivery, read production and delivery. Produce a renderer-neutral brief first; map only verified capabilities to a selected editor. Reuse
manychat-ig-lead-magnetwhen available for comment-to-access delivery. Its absence does not prevent drafting copy and a handoff. - Verify the changed artifact and report what changed, source coverage, evidence limits, and any missing input. Building a page does not refresh its sources or activate a background job.
Default content table
Use this column order unless the user requests a different layout:
| Priority | Idea / proposed hook | Idea source | Suggested format | Watch reference | Three talking prompts | Visual payoff / claim limit | CTA |
|---|
- Idea source: link to the owned video, artifact, or evidence supporting the idea; include a timestamp when available. Label an original premise as such.
- Watch reference: link directly to the specific reference video using the creator and format name as the label. A profile, search page, or unrelated idea-source link is not a substitute.
- Pick one default format and reference so the creator and editor know what to make. An optional alternative must be clearly labeled.
- Keep exactly three concise talking prompts in each row. Carry the same chosen reference URL into the editor handoff and identify what to borrow: delivery, layout, pacing, or proof order.
- If a suitable reference has not been verified, write Reference needed and state what is missing. Do not invent a URL or imply full-video review from a cover alone.
Preserve the workflow boundaries
- Extract structures rather than copying another creator's distinctive wording, story, or footage. Credit and link reference posts.
- Inspect the full relevant sequence before describing full-video pacing or format. A cover supports composition only. Label caption, spoken, and on-screen hooks separately.
- Every content idea needs a source or a clearly requested original premise, a filmable payoff, and a claim boundary. Do not convert an AI prototype, preference, forecast, or proposed test into a measured result.
- A recent reference is not automatically an outlier. A creator-relative outlier is not proof that a format caused performance. Target-account results are a separate evidence layer.
- Preserve raw observations and old format definitions. Archive or lower priority without breaking saved drafts. Opening the page, rebuilding, or a failed fetch must not advance source check dates.
- Refresh schedules, publishing, paid rendering, and live lead delivery are separate operations. Follow the user's actual authorization for each; a research request alone does not enable them. Existing explicit authorization remains valid within its scope.
- Keep private drafts, client information, internal metrics, raw research media, and local paths out of public skill packages and public pages. Public skills contain reusable instructions and fictional examples.
Useful skill handoffs
Use installed skills when their capability fits: creator-pattern-copier for source analysis, shortform-idea-grill for deeper idea intake, natural-yapping-shortform for conversational delivery, and shortform-production or the configured renderer for production. Preserve this task's requested scope rather than importing an unrelated full pipeline. The references below provide a standalone workflow when these skills are absent.
Completion evidence
For research: the source inventory, comparisons, and limitations. For a table: distinct idea-source and reference-video links, one default format, three specific prompts, and honest proof limits. For a page: loaded images, working filters/source links, saved draft reload, export content, and phone readability. For an edit: the actual review render and source alignment. For publishing or merging: the target system's readback, not a queued operation.
Version History
- 09694f0 Current 2026-09-09 07:29


