| name | xuskill |
|---|---|
| description | Run a repeatable AI creator workflow from audience-coherent topic selection and blind script review through transcript-first editing, dynamic-person proof videos, platform covers, approval-gated publishing, and evidence-based retro. Use for AI/Codex/Claude Code content, talking-head videos, screen-proof tutorials, multi-platform publish packs, and account iteration. |
XuSkill
Use this skill as an orchestration system. It turns a real creator experiment into a repeatable content pipeline without claiming guaranteed traffic or unattended publishing.
Non-negotiable boundaries
- Prefer real proof over abstract claims. Never invent a screenshot, result, metric, or successful operation.
- Treat publishing as semi-automatic. Verify every field and stop before the final publish action until the user confirms in the current turn.
- Do not delete old renders, upload files, or project assets without a dry-run list and explicit confirmation.
- Keep private paths, tokens, cookies, account data, raw creator assets, and all
*.local.mdfiles out of public packages. - If a required local capability is missing, run
scripts/check-environment.ps1, explain the missing dependency, and use the documented fallback.
Load rules by priority, not all at once
Read references/priority-router.md first. Then load the shared governance rules and only the lane needed for the current task.
Global precedence:
- safety, permission, truth, and approved inputs;
- account/audience coherence and task correctness;
- viewer outcome and retention;
- proof, trust, and creator identity;
- visual/language quality;
- speed and automation.
Never sacrifice a higher level merely to save render time or automate more steps.
Core workflow
- Account and audience state - Read
references/decision-governance.mdand a private file based onreferences/account-state.example.md. A topic can be attractive and still be wrong for the current audience. - Topic gate - Read
references/topic-human-hook.md. Ask why a stranger would stop and whether this topic naturally leads to the next two posts. - Hook and script gates - Read
references/hook-semantic-gate.md, thenreferences/retention-first-script-gate.md, thenreferences/script-preflight.md. The first line must make sense without the title or visuals; proof should appear by about four seconds and the first useful step by about twelve seconds. - Blind draft review when alternatives exist - Remove authorship and source labels, score A/B/C drafts with the same table, reverse the labels once, then issue one overall verdict. A better hook does not automatically mean a better full script.
- Content form and visual mode - Read
references/content-formats.md,references/video-mode-router.md, andreferences/visual-design-tokens.md. Choose Mode A, Mode B, Hybrid, Dynamic-Person Proof, or Creator-Stage HUD Proof before editing. Usereferences/dynamic-person-proof-template.mdfor 4:3 creator-led proof andreferences/creator-stage-hud-proof-mode.mdfor landscape creator-stage footage with negative space. - Source inventory and rough cut - Read
references/rough-cut-standard.md. Build the cut from complete semantic takes, not silence detection alone. Keep audio and video cuts identical. - Transcript timing and scene map - Time captions from the final audio. Create
03_edit/SCENE_MAP.mdfromreferences/scene-map.mdbefore rendering. - Preview and production - Read
references/video-production-sop.mdand, when real software operation is claimed,references/screen-recording-proof.md. Render a representative 30-second preview, apply the approved system to the full video, and run both QA passes. - Covers and metadata - Read
references/cover-title-publish.md. Produce a Douyin 9:16 cover and a Xiaohongshu 4:3 cover. Titles and covers must complement rather than duplicate each other. - Approval-gated publishing - Read
references/platform-publishing.md. Build and verify05_publish/publish_pack.json, use native platform topic selectors, and stop before final publishing for confirmation. - Retro and rule promotion - Read
references/metrics-retro.md. Rank one primary cause, one secondary cause, and one amplifier. Log candidate lessons locally; only promote a reusable rule after evidence review.
Visual-mode routing
- Mode A - Advanced infographic: judgement, concepts, comparisons, and saveable checklists. Use large mobile-readable hierarchy and varied diagrams. Do not turn narration into dense PPT pages.
- Mode B - Screen-proof demo: any claim that a tool operated a browser, computer, file, project, or app. Real recordings are mandatory; crop by information anchor and add purposeful focus changes.
- Hybrid - default tutorial mode: Mode A for the hook and structure, Mode B for operational claims and results.
- Dynamic-person proof template: creator-led experiments, personal judgement, repository/project evidence, and workflow demos. Keep the talking head on the top layer, enlarge the whole creator frame during personal proof, and place readable evidence behind it.
- Creator-stage HUD proof: landscape creator-led experiments where the person is the main visual stage and repository/chat/data proof occupies the surrounding negative space. Use persistent micro-motion, functional sound feedback, semantic proof crops, and separate skin-tone QA.
Do not substitute decorative cards for evidence that should be shown as real proof.
Required production artifacts
01_source/ talking head, screen recordings, screenshots
02_script/ final script and timed transcript
03_edit/SCENE_MAP.md one spoken beat mapped to one visual beat
03_edit/previews/ review renders
04_renders/final/ approved upload master
04_covers/ Douyin and Xiaohongshu covers
05_publish/ publish_pack.json and platform metadata
06_retro/ publish record and retrospective
Run scripts/create-ai-media-project.ps1 to create this structure.
Quality gates
Reject or revise when any gate fails:
- audience coherence, hook semantics, or retention structure has not passed;
- captions or visuals do not match the final audio;
- a tutorial claim has no readable real proof chain;
- real UI is cropped without the page/repository/chat context needed to understand it;
- screen proof stays visually static for more than 5-8 seconds;
- opening lacks a meaningful visual change within five seconds;
- titles, cards, captions, icons, or evidence overlap or are unreadable on a phone;
- the talking-head window comes from a downscaled lossy intermediate;
- the later half reuses unrelated early visuals or drifts from narration;
- the final master is not H.264 High, yuv420p, AAC, MP4 faststart;
- the selected cover is not the approved platform-specific asset;
- platform topics were typed as plain text instead of selected from native suggestions.
For videos longer than four minutes, require a complete timeline map and inspect every late section; a good opening preview does not prove that the full video is correct.
Tooling and portability
Read references/dependencies.md on a new computer. The workflow may use ffmpeg/ffprobe, transcription, OBS, and a code-based renderer such as Remotion or HyperFrames. These are optional by role and are not bundled.
Use environment variables instead of personal paths:
AI_MEDIA_HOME- media workspace root;FFMPEG_PATH- optional ffmpeg executable;FFPROBE_PATH- optional ffprobe executable;CODEX_HOME- Codex configuration root when needed.
Public-release safety
Before packaging or uploading, read references/open-source-safety.md and run:
powershell -NoProfile -ExecutionPolicy Bypass -File scripts/validate-release.ps1 -SkillRoot .
Then create the clean archive:
powershell -NoProfile -ExecutionPolicy Bypass -File scripts/package-release.ps1 -SkillRoot . -Version v0.4.0