视频/提示词反推生视频Skill by CreatOK

概述

Video to Seedance Skill

Turn reference material into a reusable, installed Seedance 2 prompt-writing Skill. Three input types are accepted — each maps to a dedicated analysis path:

Extract the creative system behind the examples instead of reproducing a single source. The result is a new Skill that works without the original video or prompt.

CreatOK-First Principle (applies to video/URL inputs)

The CreatOK CLI is the primary analysis engine for TikTok and public video URLs. Before any video design work begins, ensure CreatOK is available and authenticated, then use creatok analyze as the required first analysis layer: original script, storyboard / scene breakdown, source metrics, transcript, hook / proof / CTA logic, and reusable structure. Treat the full CreatOK output as the structured foundation of analysis-notes.md.

For URL inputs, a successful creatok analyze IS the complete analysis layer. When CreatOK returns ok: true with usable output, do NOT download the video, do NOT extract frames, and do NOT run the local analyze_video.py pipeline — proceed straight to step 3 with the CreatOK evidence. Its transcript, normalized vision/storyboard scenes, and response analysis are treated as the frame-level evidence for timing, continuity, product identity, and capture/editing grammar. The URL→CreatOK path ends when the analysis returns; there is no second local pass.

The local download/frame/audio pipeline is the fallback only, used in exactly two cases:

• the source is a local media file (MP4/MOV/WebM), or
• CreatOK is not installed, not authenticated, cannot reach the service, rejects the URL, or returns an error.

When CreatOK fails on a URL, record the failure briefly and then continue locally; never silently skip the CreatOK layer without logging why.

Video prompt text inputs skip the CreatOK and download layers entirely — there is no media to fetch. They are analyzed structurally as prompt text (see Workflow step 2p). Optionally, when the prompt text is accompanied by a video URL or file of the result, run CreatOK/local analysis as cross-validation, but never require it.

Mandatory Entry Check

This check is mandatory only for video/URL inputs. If the user provides video prompt text only (step 2p), skip the CreatOK check and go straight to prompt-text analysis; note in analysis-notes.md that no CreatOK layer applies.

At the start of every run with a video/URL source, verify CreatOK before anything else:

creatok doctor

• If doctor reports ok: true: CreatOK is ready. Proceed with the CreatOK-only workflow below.
• If creatok is not installed: stop and tell the user to install the CreatOK CLI first (see [references/creatok-integration.md](references/creatok-integration.md) for install guidance), then re-run creatok doctor. Do not fall back to local analysis as a silent default.
• If the CLI exists but reports a missing CREATOK_API_KEY: ask the user to configure the API key, or check for CREATOK_API_KEY_PREVIEW; only after a key is set may the workflow continue. Stable CLI uses CREATOK_API_KEY; preview CLI uses CREATOK_API_KEY_PREVIEW.
• If doctor reports skill drift, run:

  creatok skills install --dir "$SKILL_DIR"

($SKILL_DIR is the directory containing this SKILL.md; on Windows use the concrete path such as C:Users82734.workbuddyskillsvideo-to-seedance-skill.)

• If an update is available (creatok doctor reports it, or creatok version shows an update field), suggest creatok update to the user but do not block the workflow on it.

Read [references/creatok-integration.md](references/creatok-integration.md) before using CreatOK.

Required outcome

Create a new personal Skill whose future inputs are images, product information, scripts, or short requirements. The generated Skill must not require the original source video after creation.

Do not return only a video description, shot-by-shot replication prompt, generic template, or uninstalled SKILL.md draft.

Workflow

1. Identify the video family and ask the user

The core value of this Skill is producing a reusable derived Skill (e.g., "a Skill that always generates try-on videos", "a Skill that always generates product reveal ads", "a Skill that always generates UGC talking-head ads"). Therefore, first identify which creative family the source video belongs to, then ask the user targeted questions to complete the context around that family's derivation goal — do not derive from a single video in isolation.

1a. Identify the family first, then infer. On input, immediately classify the video family using the recognition signals in [references/video-family-guide.md](references/video-family-guide.md) (content, narrative engine, capture/editing grammar, speech role, platform/conversion context). For URL inputs, base this initial classification on the URL/creator context and (when already available) CreatOK output; for local files, use a quick storyboard skim. If evidence is weak, present candidate families to the user for confirmation.

1b. Ask the user to complete context (key step). After identifying the family, ask targeted questions. Timing:

• Round 1 (before analysis, required): confirm derivation goal and key boundaries
• Round 2 (after analysis, before design, as needed): clarify when multiple interpretations or conflicting evidence appear

Round 1 required questions (use AskUserQuestion with options + free text; max 4 questions per round; ask the highest-impact first):

1. Derivation goal (派生目标): What will this derived Skill mainly produce in the future? (always generate try-on videos / product reveal ads / UGC talking-head ads / tutorials / story shorts / vlogs / music-led montages / other)
2. Future inputs (未来输入): What will the user supply to this Skill later? (product image / presenter image / outfit image / script / selling points / market / language)
3. Target model: Seedance 2 (default) or another video model? (affects prompt format)
4. Source count & supplements (源数量): Is the current source enough? Can you add 2-4 more same-family videos (different products/people/scenes — cross-source evidence is more robust)?

Family-specific follow-up questions (only the relevant ones, max 2-3):

• Speech-led UGC / talking-head / review / testimonial: fixed presenter? preferred spoken language? voiced by default or silent/post-production?
• Try-on / outfit: fixed model/body? fixed mirror/scene? on-image text needed?
• Product film / cinematic reveal / craft process: fixed product category? no-speech default acceptable?
• Tutorial / demo: must show complete steps? voice-over language?
• Story short / drama: fixed characters? dialogue language? target length?
• TikTok Shop / shoppable UGC / live-style pitch: CTA slot style? platform UI allowed?
• Music-led montage: music direction? beat structure?

Record answers in analysis-notes.md under a "User context" section. Stop asking when the user says it is enough; record refusal and continue if the user declines an item — never block the workflow on questions.

If the user provides no name, create a concise lowercase hyphenated name based on the video family and outcome. Avoid names tied to the source creator, source brand, or one product.

When the user asks whether more reference sources help, recommend 3-5 items from the same creative family but with different products, people, scenes, or executions. One source is enough to create a useful Skill, but it must be labeled single-source and interpreted conservatively. Two sources are directional, not definitive.

Derivation goal × family → output shape (map user intent while asking):

2. Analyze URL sources (CreatOK-only)

For URL inputs, this step IS the analysis — there is no download step. A successful creatok analyze completes the evidence layer; go straight to step 3.

For TikTok single video URLs, CreatOK analysis is mandatory and sufficient. Read [references/creatok-integration.md](references/creatok-integration.md) before using CreatOK. Run:

creatok analyze --url "https://www.tiktok.com/@creator/video/123" --out "/tmp/video-to-skill/creatok/<source-name>"

On Windows, use a concrete output directory:

creatok analyze -url "https://www.tiktok.com/@creator/video/123" -out "C:Users82734WorkBuddyvideo-to-skillcreatok<source-name>"

Apply the CreatOK analysis expectations to the result before designing anything: inspect the JSON envelope and persisted result.json, extract the original script, storyboard / scene breakdown, available metrics, transcript segments, hook, proof, value, CTA, conversion logic, and reusable structure. Record this complete analysis in analysis-notes.md under a clearly labeled CreatOK section — it is the primary structured evidence. Use CreatOK's normalized vision/storyboard scenes as the frame-level evidence for timing, continuity, and scene structure; use its transcript segments as the speech evidence.

Do not download the video after a successful analysis. Skip download_reference_video.py and analyze_video.py entirely for URL inputs whose CreatOK analysis succeeded. The only exceptions (fallback cases) are:

• the source is a local media file (skip download; run the local pipeline directly), or
• CreatOK is unavailable / unauthenticated / rejects the URL / returns an error — record the failure and then continue locally (see fallback commands below).

If the user provides a creator/profile/channel URL instead of direct video URLs, treat it as a source-discovery URL. Use scripts/discover_reference_videos.py to collect and rank recent public videos before selecting which to analyze:

python3 "$SKILL_DIR/scripts/discover_reference_videos.py" "https://example.com/@creator" 
  --limit 5 
  --output "/tmp/video-to-skill/discovered/videos.json"

When supported metadata is available, prefer high-performing videos: rank primarily by view/play count, then use likes, comments, reposts, or recency as tie-breakers. Select a sample of 3-5 videos, then run complete creatok analyze per selected URL — this is required, not optional. Keep profile metadata as selection context only; use CreatOK output as the source-video evidence. If CreatOK fails for any selected URL, record the failure and use local media analysis for that video, then note the mixed evidence in analysis-notes.md.

When using a profile URL, record only lightweight selection context:

• Which videos were selected and why: view/play count, engagement, recency, or user-provided priority.
• Whether each selected URL produced a successful CreatOK analysis.
• Whether any high-priority item failed CreatOK analysis.
• Any account-page clue that should be treated as uncertain until verified by CreatOK or local frames.

Fallback — only when CreatOK fails or the source is a local file. If a source is a URL and CreatOK could not analyze it, first acquire the media with scripts/download_reference_video.py. Resolve SKILL_DIR to the directory containing this SKILL.md, then run:

python3 "$SKILL_DIR/scripts/download_reference_video.py" "https://example.com/video" 
  --output-dir "/tmp/video-to-skill/downloaded/<source-name>"

For TikTok, YouTube, Instagram, X/Twitter, Bilibili, Xiaohongshu, or other yt-dlp-supported public media, the downloader requires yt-dlp in the active Python environment or on PATH, and ffprobe from ffmpeg. If the page requires login but the user has a browser session, retry only when appropriate:

python3 "$SKILL_DIR/scripts/download_reference_video.py" "https://example.com/video" 
  --output-dir "/tmp/video-to-skill/downloaded/<source-name>" 
  --cookies-from-browser chrome

Only proceed when the downloader returns "ok": true, a local media_path, and a non-zero probed duration. Use that media_path as the input video for analysis.

Do not claim an account profile page was downloaded unless a playable media file was actually produced. Prefer direct individual video URLs over creator profile URLs. Do not bypass private accounts, paywalls, DRM, deleted/private media, or access controls. Downloading a reference is not proof that the user has reuse rights.

Before local fallback analysis, verify media usability:

• Confirm a video stream exists, duration is non-zero, frames can be extracted, and extracted frames are not blank or black.
• Confirm whether audio exists and whether audio.wav was produced.
• Record codec/container issues such as HEVC/H.265 files that may appear black in some players while still containing valid frames.
• If local playback is unreliable but frame extraction works, create or request an H.264 proxy when useful. Do not call a video unusable only because one player cannot display it.

2p. Analyze video prompts (text input)

Use this path when the user provides video prompt text — one or more prompts written to generate a video — instead of (or in addition to) a video file/URL. No CreatOK call and no download is required for this input type. The prompt text itself is the source of truth.

Collect the prompt(s):

• One prompt is enough to create a useful Skill, but it must be labeled single-source and interpreted conservatively.
• Two prompts are directional, not definitive.
• 3-5 same-family prompts (same video family, different products/people/scenes) are the preferred evidence set.

Analyze each prompt structurally. A Seedance-style prompt typically contains these sections — read them against the intended output in [references/generated-skill-blueprint.md](references/generated-skill-blueprint.md):

• Reference Materials: which assets ([Image 1], [Video 1], etc.) the prompt assumes; future inputs this Skill must request
• Global Specs: duration, aspect ratio, capture style, continuity, subtitle/speech/music policy
• Video Content / Timeline: shot list, beat order, transitions, camera moves, subject actions, proof/payoff placement
• Audio: dialogue, voice-over, SFX, ambience, music treatment
• Consistency and Constraints: reference locks, negative constraints, style guardrails

For each prompt, extract into analysis-notes.md (clearly labeled as prompt-text evidence):

• Prompt structure skeleton: which sections are present and in what order
• Creative DNA: hook mechanism, narrative beats, shot grammar, performance direction, product interaction, visual treatment, speech-to-visual proof, audio rhythm, payoff/CTA
• Fixed conventions: wording patterns, section framing, constraint style that recur
• Variables: product, person, scene, language, duration, brand details
• Prompt-writing rules worth teaching: how the prompt orders information, how it locks references, how it phrases negative constraints

Apply the same method separation as step 5: distinguish invariant prompt-writing rules (belong in the derived Skill) from defaults, variables, and incidental phrasings. Do not copy the exact source prompt verbatim into the derived Skill; abstract its writing method and fillable slots.

For mixed inputs (prompt + video of the same result), cross-validate: the prompt explains intent, the video/URL analysis (CreatOK-only for URLs; local pipeline for files) verifies what the prompt actually produced. When they conflict, prefer the video as ground truth for visual timing/continuity, and treat the prompt as the authoritative writing style to preserve.

Do not proceed to Skill design until analysis-notes.md for each prompt contains: source role, source count and confidence, prompt structure skeleton, Creative DNA extraction, fixed conventions vs variables, prompt-writing rules, and invariant/default/variable/incidental separation.

3. Classify video type and narrative engine

Applies to both input types. For videos, classify from playable-video evidence (for URL inputs, this evidence comes from the CreatOK analysis completed in step 2 — its transcript, vision scenes, and response analysis; for local files, from frames/audio). For prompt-text inputs, classify from what the prompt explicitly directs (described shots, audio instructions, capture style); never invent classifications the prompt does not support.

Before detailed extraction, classify each source and the source set. Do not design the derived Skill until the dominant narrative engine is explicit. Family recognition: see [references/video-family-guide.md](references/video-family-guide.md) (recognition signals, analysis priorities, derived output shape, subtitle/speech/music policy, common traps, follow-up questions).

Classify the video family:

• Speech-led UGC: talking-head, street encounter, off-camera interview, try-on, review, testimonial, recommendation, demo with presenter explanation.
• Visual-led demonstration: actions and product behavior carry the meaning without necessary speech.
• Cinematic reveal: mood, process, and payoff carry the meaning; audio/subtitles are mostly post-production layers.
• Music-led montage: beat, movement, dance, or edit rhythm carries the structure.
• Hybrid: identify the primary and secondary engines.

Then classify the narrative engine:

• Primary speech: dialogue or voice-over explains the hook, claims, objections, proof, and CTA.
• Primary visual action: the viewer can understand the arc without words.
• Primary music/editing: beat drops, lyrics, or rhythm determine the structure.

For UGC, review, try-on, tutorial, talking-head, street interview, testimonial, or presenter-led ads, assume speech may be primary until proven otherwise. Classify the source from playable-video evidence first (for URLs: CreatOK transcript + vision scenes; for local files: audible speech, visible mouth movement, off-camera questions, gestures, claim/proof timing, and transcription when available). Account-level impressions can suggest what to check, but they are not enough to declare the derived Skill's dominant format.

The classification must determine analysis priorities and the derived Skill's defaults. Do not carry over the general "no speech/no subtitles/no BGM" default blindly. That default is correct for cinematic craft/process reveals, but often wrong for speech-led UGC, reviews, try-ons, testimonials, tutorials, and talking-head ads.

Also classify the capture/editing grammar:

• One continuous take: one handheld or stabilized shot, with walk-up, push-in, pan, tilt, sidestep, follow, or reframing instead of visible hard cuts.
• Hidden-cut continuous style: feels continuous but likely has subtle stitched cuts.
• Jump-cut UGC: presenter or action advances through visible quick cuts.
• Montage/edit-driven: meaning comes from many short shots, rhythm, or music synchronization.
• Cinematic multi-shot: planned coverage with distinct angles, inserts, and edits.

If the source is one continuous take or hidden-cut continuous style, the derived Skill must preserve continuous camera logic. Do not convert it into a 6-8 shot montage.

Classify the platform/conversion context from playable-video evidence when relevant (for URLs: CreatOK transcript/vision; for local files: frames/audio):

• TikTok Shop/shoppable UGC: trust-building proof followed by a natural shop/link/product CTA.
• Organic social post: entertainment, personality, or discovery with no explicit purchase action.
• Brand ad: controlled product messaging and more explicit CTA.
• Affiliate/recommendation: creator trust and personal recommendation drive conversion.
• Review/testimonial: credibility comes from experience, objection handling, or comparison.
• Tutorial/demo: conversion comes from teaching and visible utility.
• Live-style pitch: presenter uses live-shopping style urgency, direct address, rapid objections, or deal explanation even if the video is pre-recorded.

Use only evidence from the playable video, downloaded metadata, or explicit user-provided context: spoken CTA, presenter wording, visible shopping/product sticker as an actual UI element, demonstrated shopping action, title/description metadata, or user-provided campaign goal. For URL sources, take this evidence from the CreatOK analysis (transcript, vision scenes, metadata). Do not OCR subtitles or on-screen text for this classification. Do not infer TikTok Shop or affiliate intent merely because the URL is TikTok, the creator sells products, or the account grid looks commercial. If evidence is weak, mark platform/conversion context as uncertain and design a flexible CTA slot instead of a hard platform rule.

Platform context affects CTA, UI, subtitle, and claim rules. Do not generate fake platform UI, product stickers, checkout screens, prices, discounts, stock, shipping, or guarantees unless the user provides or requests them.

4. Analyze every source video

This step applies to video files and URLs. For prompt-text inputs, the equivalent analysis is step 2p — do not run video extraction on text.

For URL sources, the CreatOK analysis completed in step 2 IS the analysis layer. Do not run the local pipeline below for URLs whose CreatOK analysis succeeded. For local media files, run the local pipeline below (and additionally run creatok analyze on the originating URL whenever the local file was downloaded from a TikTok URL — CreatOK remains the primary layer even then, with local frames used to verify timing/continuity).

Resolve SKILL_DIR to the directory containing this SKILL.md, then run for each video that needs local inspection (local files, or URLs where CreatOK failed):

python3 "$SKILL_DIR/scripts/analyze_video.py" "/absolute/path/to/video.mp4" 
  --output "/tmp/video-to-skill/<video-name>" 
  --max-frames 40

Inspect storyboard.jpg, metadata.json, scenes.json, analysis-notes.md, audio.wav, and relevant individual frames. The script creates analysis-notes.md as a required working template; fill it before designing the derived Skill.

CreatOK analysis is the primary structured evidence for TikTok/URL sources. Inspect its JSON envelope and persisted result.json for every source it was run on, and treat these fields as required structured evidence:

• transcript segments and timestamps
• video metadata and public metrics when present
• normalized vision/storyboard result
• remote response text and suggestions
• original script and storyboard / scene breakdown in the style required by creatok-analyze-video
• hook, value, proof, CTA, conversion logic, and reusable-template analysis

Map the complete CreatOK analysis into analysis-notes.md, labeled as the primary CreatOK evidence section. The CreatOK section must include the original script summary, storyboard / scene breakdown, available public metrics, transcript, hook/proof/CTA logic, and source-derived reusable template before any derived Skill design begins. For URL sources analyzed successfully by CreatOK, its vision/storyboard scenes are the frame-level evidence for visual timing, continuity, object identity, and on-screen action; its transcript is the speech evidence. Local frame/audio inspection only overrides CreatOK when a local file was actually inspected (local-file sources or documented CreatOK-failure fallback); in that case, trust local playable media for visual timing, continuity, object identity, and actual on-screen action, and use CreatOK transcript only when it aligns with visible mouth movement/audio context or when local transcription is unavailable.

If the video is speech-led or could be speech-led, audio/speech analysis is mandatory. For URL sources, use the CreatOK transcript as the speech evidence. For local files, use an available transcription capability for audio.wav when possible. If speech is not reliably recoverable, do not skip the layer: analyze visible captions, mouth movement, presenter gestures, timing, and source metadata; mark the speech content as unrecoverable and infer only the speech function, not exact words.

When running the local pipeline (local files or CreatOK-fallback sources), inspect more carefully than the contact sheet when the video has fast edits, product steps, meaningful gestures, or a reveal. Open individual frames around the hook, every detected cut, every product interaction, and the final reveal. If scene detection is weak, sample evenly and compare adjacent frames manually. For URL sources analyzed successfully by CreatOK, its vision/storyboard scene list provides the same cut-level structure; rely on it directly instead of a local frame pass.

Build an internal evidence table:

After the evidence table, write a short core logic pass in analysis-notes.md before designing the derived Skill:

Emotional hook:
What tension does the first 1-2 seconds create?

Story engine:
What cause-and-effect makes the viewer keep watching?

Proof chain:
What visible actions prove the claim, value, transformation, or craft effort?

Emotional reversal:
How does the subject's feeling or viewer judgment change by the end?

Reusable formula:
What is the portable method independent of the source product, person, language, music, or platform?

Narrative engine:
Is the source driven primarily by speech, visual action, music/editing, or a hybrid? Why?

Capture/editing grammar:
Is the source one continuous take, hidden continuous style, jump-cut UGC, montage/edit-driven, or cinematic multi-shot? What evidence proves this?

Dialogue logic:
If speech exists or is implied, what is the chain of speaker roles, questions, answers, follow-ups, proof, and CTA?

Platform/conversion context:
Based only on playable-video evidence (for URLs: CreatOK transcript/vision; for local files: frames/audio), downloaded metadata, or explicit user context, is this TikTok Shop/shoppable UGC, organic social, brand ad, affiliate recommendation, review/testimonial, tutorial/demo, live-style pitch, or uncertain? How does that affect CTA and claim rules?

Speech-to-visual proof:
If speech is important, what spoken claim is made and what visual evidence immediately proves it?

Do not treat facial performance as decoration. If the opening face looks like crying, holding back tears, frustration, disbelief, pride, embarrassment, shock, relief, or vindication, record that exact emotional function and decide whether it is an invariant, default, or optional variant. Preserve the function, not the source person's identity.

Do not analyze subtitles, captions, OCR, or existing on-screen text unless the user explicitly asks. Seedance should generally not generate subtitles/captions; subtitles are expected to be added in post-production.

For speech/audio analysis, decide whether speech is primary, secondary, decorative, absent, or unrecoverable. In speech-led UGC, review, try-on, tutorial, testimonial, and talking-head videos, preserve the spoken function as an invariant and make generated speech a default in the derived Skill unless the user asks for silent/post-production mode.

For music/audio analysis, record reusable editing information such as energy, beat placement, reveal timing, or whether process SFX matter. Do not make generated BGM default only because the source had music. Do not make generated subtitles default; subtitles are normally added in post-production.

Do not proceed to Skill design until analysis-notes.md contains:

• Source metadata and frame count
• Source role, source count context, and confidence level
• URL/profile selection context only when applicable, clearly separated from CreatOK/video evidence
• CreatOK analysis status (mandatory for TikTok/URL sources): command result, artifacts directory, API-key/auth/reachability issue if it failed, and which transcript/metadata/vision/metrics fields were used
• Complete CreatOK video analysis (mandatory for TikTok/URL sources): original script, storyboard / scene breakdown, available metrics, hook, value, proof, CTA, conversion logic, and reusable source structure
• Media usability: for URL sources, the CreatOK status (no local download performed on success); for local files, playable status, codec issues, frame extraction, audio extraction, and any proxy/transcode need
• Video type and narrative-engine classification
• Capture/editing grammar classification
• Core logic pass
• Evidence table
• Beat map with timestamps
• Dialogue logic map when speech or off-camera voice drives the sequence
• Speech-to-visual proof map when speech is important
• Subtitle handling: do not analyze source subtitles/captions/OCR; default no generated subtitles/captions; leave subtitles for post-production unless the user explicitly requests otherwise
• Audio/music handling
• Evidence-based platform/conversion handling, including TikTok Shop/shoppable CTA rules only when supported
• Invariant/default/variable/incidental/uncertain separation
• Derived Skill requirements

4b. Synthesize across multiple sources

If the user provides more than one source (multiple videos, multiple prompts, or a mix), complete the per-source analysis first (analysis-notes.md for each), then write cross-video-synthesis.md in the working folder before designing the derived Skill. For prompt sources, name the synthesis cross-prompt-synthesis.md and note that evidence is prompt-text based.

The synthesis must include:

• Source set: count, formats (video / prompt), durations or lengths, and why the examples belong to the same family
• Shared core logic: the emotional, narrative, proof, conversion, or entertainment mechanism repeated across sources
• Invariants: patterns that recur in most sources or are necessary for the family to work
• Defaults: useful common choices that can be overridden
• Variants: different supported hooks, pacing choices, scene types, product categories, presenter behaviors, and reveal styles
• Single-source details: memorable details that appeared only once and should not become hard rules
• Timing ranges: typical proportion per beat rather than one source's exact seconds
• Narrative-engine synthesis: whether speech, visual action, music/editing, or a hybrid drives the family
• Capture/editing grammar synthesis: whether the family is continuous-take, hidden-continuous, jump-cut, montage, or cinematic multi-shot
• Dialogue logic synthesis: repeated speaker roles, question/answer/follow-up pattern, claim order, proof order, and CTA handoff
• Speech-to-visual proof pattern when relevant: repeated claim types and the visual proof that follows each claim
• Prompt-writing synthesis (when prompt sources are present): recurring section order, wording conventions, reference-lock style, negative-constraint phrasing, and fillable slots
• CreatOK contribution: which repeated transcript, metadata, metrics, or vision patterns were confirmed by the remote analysis, and which conclusions still came from local playable-media inspection (local inspection only applies to local-file sources or documented CreatOK-failure fallbacks)
• Complete CreatOK analysis contribution: repeated script functions, storyboard beats, metrics signals, hook/proof/CTA logic, and conversion patterns confirmed before Skill design
• Subtitle handling: do not analyze source subtitles/captions/OCR; default no generated subtitles/captions; leave subtitles for post-production unless explicitly requested
• Audio/music handling: repeated function, beat/reveal timing, and whether future prompts should exclude generated audio
• Evidence-based platform/conversion synthesis: TikTok Shop, organic, brand ad, affiliate, review, tutorial, live-style pitch, or uncertain context; CTA and claim constraints
• Reference-input implications: what future users must provide, what is optional, and what can be inferred
• Confidence: high, medium, or low, with a short reason

5. Separate method from incidental content

Applies to both video and prompt-text sources. Classify each observation:

• Invariant: defines the video family and belongs in the new Skill
• Default: useful starting choice that users may override
• Variable: product, person, outfit, scene, language, claim, or duration
• Incidental: source-specific detail that should not be preserved
• Uncertain: insufficiently supported; omit or label as optional

Use evidence thresholds:

• With one source video, preserve only clearly functional patterns and add single-source caution to the derived Skill's evidence reference. Do not promote every visible detail into a rule.
• With two source videos, treat repeated patterns as directional defaults unless they are clearly necessary to the mechanism.
• With 3-5 source videos, promote repeated or functionally necessary patterns to invariants, preserve meaningful variants, and exclude details that do not recur or explain the effect.
• If the source videos conflict, prefer a flexible rule that names the alternatives instead of forcing one style.

Abstract:

• Exact source dialogue → dialogue function and fillable script slot
• Exact source emotion → emotional function and performance direction
• Specific product → product authority and interaction rule
• Specific actor → presenter authority and performance profile
• Specific location → scene function and visual criteria
• Specific music → tempo, energy, beat and edit behavior
• Exact shot timing → reusable proportional or bounded timing
• Source subtitles/captions → exclude from generated prompts by default; leave subtitle work for post-production
• Source BGM/audio → optional post-production guidance, not required generated audio

Never preserve watermarks, platform UI, source creator identity, unrelated brands, copyrighted lyrics, or unsupported product claims.

For CreatOK-derived transcript or script text, preserve only the reusable line function unless the user explicitly asks to quote or adapt the source script. Do not bake source wording, source captions, source creator identity, or platform UI into the derived Skill. For URL sources, the derived Skill must not depend on CreatOK, the original URL, or remote artifacts after creation — the method is distilled from the analysis, not from the source URL.

For prompt-text sources, preserve the reusable prompt-writing method (section order, phrasing conventions, constraint style, fillable slots) — never copy the full source prompt verbatim into the derived Skill. Keep at most a short abstracted example.

6. Design the derived Skill

Read [references/generated-skill-blueprint.md](references/generated-skill-blueprint.md).

The derived Skill must contain:

• Precise trigger description
• Expected input/reference map (including the prompt-writing input slots the derived Skill will fill)
• Defaults and clarification policy
• Extracted creative DNA
• Source count, evidence confidence, source type (video / prompt / mixed), and either single-source caution or cross-source synthesis
• Video type and narrative-engine classification
• Core logic formula: emotional hook, conflict, proof chain, reversal, and payoff
• Speech-to-visual proof formula when speech or presenter explanation drives the source
• Reusable narrative and shot grammar
• Family-shaped output structure: try-on/speech-led UGC → default spoken lines + pose/scene variants; cinematic reveal/product film → shot list + complete process coverage; tutorial/demo → full step coverage + voice-over policy; story short → dialogue logic + character continuity; vlog → continuous-take grammar; music-led montage → beat/edit rhythm structure (see [references/video-family-guide.md](references/video-family-guide.md))
• Prompt-writing rules when the source is a prompt: section order, reference-lock style, negative-constraint phrasing, wording conventions, fillable slots
• Explicit Reference Lock section (required whenever future inputs include reference materials — presenter/outfit/product images or reference videos): follow the Reference lock template in references/generated-skill-blueprint.md — lock-sentence three conditions (name the reference using the host's exact token syntax / lock scope covers subject or motion authority only / explicitly carve out exceptions) + per-host lock templates + anti-drift guards. Without this section the derived Skill repeats the "person / outfit / product did not lock" failure
• Subtitle, speech, and music policy
• Complete process/step-coverage policy when the source relies on craft, demonstration, transformation, tutorial, assembly, or proof
• Seedance 2 prompt assembly procedure
• Reference-authority and consistency rules
• Variation rules so outputs are not near-duplicates
• Validation checklist
• Exact default output format

Keep evidence separate from instructions. If source-derived detail is useful later, place it in references/source-patterns.md and link it directly from the derived SKILL.md. Do not bundle the original video unless the user explicitly requests it.

Use the completed analysis-notes.md as the source for references/source-patterns.md. Summarize evidence; do not copy raw frame paths, private media paths, full transcripts, or exact source captions unless explicitly requested.

7. Create and install the derived Skill

Read the complete current skill-creator Skill before writing. Follow its personal-Skill initialization, metadata generation, validation, saving, and verification workflow.

Initialize the new Skill in the personal-skills checkout using init_skill.py. Include only necessary resources:

• Usually references
• Add scripts only for deterministic processing the derived Skill will reuse
• Do not add examples, README files, or raw extracted frames by default

Generate agents/openai.yaml so its default prompt explicitly mentions the new $skill-name.

Validate the actual installed Skill with quick_validate.py. Additional output testing is optional and only needed when the user asks for it.

8. Quality gate

Confirm:

• The new Skill can run without the source video or source prompt
• It captures a reusable method rather than literal content
• For video/URL sources: creatok doctor was run at the start; CreatOK status (ready / failed-with-reason) is recorded
• For every TikTok/URL source, CreatOK analysis was attempted first, and its structured output (script, storyboard, metrics, hook/proof/CTA) is cited as primary evidence in analysis-notes.md
• For URL sources with successful CreatOK analysis, no local download/frame/audio pass was performed — CreatOK was treated as the complete analysis layer; local analysis appears only for local-file sources or documented CreatOK failures
• CreatOK analysis, when used, is cited internally as structured evidence and the derived Skill does not depend on CreatOK, the original URL, or remote artifacts after creation
• Only if CreatOK is unavailable or unauthenticated does the workflow rely on local download/frame/audio analysis — and that fallback reason is documented
• For prompt-text sources: no video extraction or CreatOK call was forced; analysis-notes.md records prompt structure skeleton, Creative DNA, fixed conventions vs variables, and prompt-writing rules; the derived Skill abstracts the writing method and does not copy the source prompt verbatim
• It reflects the source count and evidence confidence
• User context was collected via Step 1b questions (derivation goal / future inputs / target model / source count / family-specific questions); refusals recorded in analysis-notes.md and workflow continued
• It correctly identifies whether the source family is speech-led, visual-led, cinematic, music-led, or hybrid
• The derived output shape matches the family (spoken-lines default for speech-led UGC; step coverage for tutorials/demos; dialogue+continuity for story shorts; continuous-take grammar for vlogs; beat/edit structure for music-led montages)
• For speech-led sources, it makes speech/dialogue a core default unless the user explicitly wants silent/post-production mode
• For non-speech-led sources, it does not add dialogue just because audio exists
• If multiple sources were provided, it distinguishes cross-source invariants from supported variants and one-off source details
• If only one source was provided, it avoids overclaiming and records single-source caution
• It captures the source's core logic, including emotional hook, conflict, proof chain, reversal, and payoff
• analysis-notes.md was completed before writing the derived Skill
• Its Creative DNA explains why the sequence works; it is not merely a list of shots or aesthetic traits
• Fixed rules, defaults, variables, and variants are distinguishable
• Future references have explicit authority roles
• Seedance prompts use visible, executable directions
• Timelines fit the requested duration
• Subtitle, speech, music, and BGM follow the user's requested generation policy
• Derived subtitle/speech/music defaults match the source type: speech-led UGC may default to natural spoken lines; cinematic craft/process reveals may default to no speech/subtitles/BGM and leave subtitles for post-production
• Audio and dialogue are not fabricated
• Craft/demo/process Skills show the complete construction or transformation logic, compressed as needed, rather than letting the final product simply appear
• Product identity and character continuity rules are present when relevant
• Outputs vary while retaining the learned creative DNA
• The trigger description activates on realistic future requests

Final response

After successful creation, tell the user the new Skill's name, extracted method, accepted future inputs, an example invocation, availability, and detail-page link. When the source was prompt text, explicitly note that the Skill was distilled from prompt(s) and what future inputs it accepts.

Do not make the analysis table the primary deliverable. Do not expose internal saving mechanics.