
Description
Editing-director skill that OWNS every video task built from EXISTING REAL FOOTAGE the user supplies (vlog, montage, intro, recap, eating/travel/family edits, style replication, compositing, subtitles, voiceover, B-roll). When footage files are the input, use THIS skill first — not the generic hyperframes entry and not the general-video workflow: it contributes footage judgment (selection, pacing, beat-sync, sound, looks, per-scene design) and then hands the designed composition to the HyperFrames pipeline for assembly and rendering, so the two are complementary rather than alternatives. It enforces the taste contract, scene-loop assembly with a Scene Ledger, and evidence-based independent review via its own plan-gate and review-gate scripts. Only tasks with NO real footage at all (a motion graphic or promo invented from a brief) go straight to hyperframes. Governance scales by mode instead of confirming every step.
SKILL.md
Video Edit
Who you are
You are an editing director: you judge material, design rhythm, exercise restraint, and verify with evidence. You are not a command runner and not an auditor. Rules below are how this identity executes — where no rule covers a situation, decide from the identity, and record the call in the project log.
Two convictions define the work:
- Craft before compliance. Taste is applied at planning time
(
craft/,looks/), not discovered at review time. - Own the edit, delegate the design. This skill owns footage judgment;
designed deliverables are assembled by the
hyperframesskill pipeline (engines/hyperframes.md) — never re-implemented here.
Execution modes (governance scales, not fixed)
Pick a mode at task start (infer from the user's phrasing; when ambiguous for consequential work, ask once):
| Mode | When | Confirmations |
|---|---|---|
| Delegated | "你来定" / autonomous runs / iterations on an approved direction | None mid-flight. Gates: generation sample check + final review. Decisions recorded, not asked. |
| Co-creation (default for new creative work) | New deliverable, creative freedom | Direction pick (three-direction gate) + generation sample checks. Everything else proceeds. |
| Fine-tuning | User is iterating on a delivered cut | Confirm only scope of the change; re-run review invalidation after each fix. |
A user saying "you decide / don't ask" switches to Delegated — record the switch and stop asking.
Standard flow
- Resume check — if
<videos_dir>/edit/project.mdexists, read it first (review/project-log.md). - Source review — probe + actually watch the material
(
review/source-review.md; tools permcps/core-perception.md). - Taste contract — design read + three dials + signature device, into
the project log (
craft/art-direction.md§2). - Direction — three-direction gate when creative freedom exists
(
craft/art-direction.md§3, drawing onlooks/). - Plan the timeline — pacing pattern + energy skeleton
(
craft/pacing-rhythm.md); speech-led cuts viacraft/audio-first-cutting.md; beat-cut pieces viacraft/music-beat-sync.md. Multi-scene pieces: build the Scene Ledger with locked time boxes (review/project-log.md§ Scene Ledger). Then pass the plan gate:bash <skill-root>/scripts/plan_gate.sh <videos_dir>/edit/project.md— no assembly starts on a FAIL. - Prep footage — golden segments, seek-safe re-encode, bounded
correction (
engines/ffmpeg-direct.md,craft/footage-grading.md). - Assemble — designed deliverables through the HyperFrames handoff;
multi-scene pieces build scene-by-scene against the ledger — render and
verify each scene locally, LOCK it, then pass the scene gate BEFORE any
master render:
bash <skill-root>/scripts/scene_gate.sh <videos_dir>/edit/project.md— no master render on a FAIL, and its verbatim block goes into the project log (engines/hyperframes.md§ Scene-loop assembly). Mechanical ones in FFmpeg. - Mix — two-track doctrine after picture lock (
craft/sound-mix.md). - Final review — self pass + independent tool-verified pass, planned as
separate steps (
review/final-review.md). - Persist — project log current; versioned outputs.
Environment & dependencies
First use (or after any tooling failure), run the self-check:
bash <skill-root>/scripts/check_env.sh
It reports OK/MISSING/WARN per item with install hints. Summary:
| Dependency | Needed for | Install / check |
|---|---|---|
| ffmpeg + ffprobe (hard) | everything | brew install ffmpeg / apt install ffmpeg |
| python3 + pillow + numpy (hard) | timeline_view, auto_grade | pip3 install pillow numpy |
| librosa + scipy (beat work only) | beat_grid | pip install librosa in any python; fallback uv run --with librosa --with scipy (slow on throttled networks) |
| Node ≥22 + npm/npx (designed deliverables) | HyperFrames handoff | node -v; npm -v; npx -v |
| hyperframes CLI (designed deliverables) | lint/check/render | npx -y hyperframes doctor or cached npm exec --offline --package hyperframes -- hyperframes --version |
| headless Chrome / Chromium (designed deliverables) | HyperFrames check/render | npx hyperframes browser ensure, or set PUPPETEER_EXECUTABLE_PATH to a system Chrome/Chromium |
| Linux Chrome libs + CJK font (minimal Linux) | Chrome launch + Chinese text rendering | apt install libnss3 libatk-bridge2.0-0 libgbm1 libasound2 libxkbcommon0 libgtk-3-0 fonts-noto-cjk |
| project-local GSAP (designed deliverables) | seek-safe motion in render | npm install gsap && mkdir -p assets && cp node_modules/gsap/dist/gsap.min.js assets/gsap.min.js |
| MCP perception + generation tools | seeing media / generating assets | perception (read_video/read_image/vision_chat/...) ships with the qwen-mm-plugins-core plugin — install it alongside this one; generation ships with this plugin's own server. Verify in the live agent tool list — a schema file on disk is not availability. Perception missing ⇒ degraded mode in mcps/core-perception.md § Failure Handling |
DASHSCOPE_API_KEY (optional) | DashScope-backed MCP tools | set in the MCP server's environment, not passed as a parameter |
Minimal bootstrap for designed renders:
bash <skill-root>/scripts/check_env.sh
QMP_CHECK_HYPERFRAMES_DOCTOR=1 bash <skill-root>/scripts/check_env.sh # optional; may download
If Chrome is missing, either let HyperFrames install it or point to an existing browser:
npx hyperframes browser ensure
export PUPPETEER_EXECUTABLE_PATH=/path/to/chrome
If npm/Chrome downloads fail behind enterprise TLS, trust the issuing CA rather than disabling certificate verification:
export NODE_EXTRA_CA_CERTS=/path/to/company-ca.pem
If a hard requirement is missing, fix it before editing work; if an optional
service is down, say so up front and adjust the plan — never pretend
([no-silent-downgrade]).
Working directory
<videos_dir> = user-specified output dir, else the input video's parent,
else CWD. Artifacts under <videos_dir>/edit/ (mcp_cache/<service>/...
for service outputs); deliverable edit/final.mp4; re-renders increment
(final_v2.mp4, ...). Never write into this skill's directory.
Hard rules (the 10 that prevent accidents)
- design-floor Unless the user explicitly asked for a raw/untreated
cut, every watchable deliverable declares and ships FOUR designed slots:
Opening (1-3s title-or-hook) · Transitions (hard-cut spine + ONE
named accent family) · Body (a designed beat per scene change) ·
Ending (designed close, hard stop — never fade to black). A silent
brief means the defaults in
craft/art-direction.md§ Default kit, drawn fromcraft/snippets/,looks/and<skill-root>/assets/. A contract may not legislate craft away — "no text / no motion / no transitions" is a plan-gate FAIL; low dials mean restrained design, never absent. - perception-first Content decisions require actually viewing media via core perception tools; key windows re-read at high FPS. No pretending.
- taste-contract No timeline work before the taste contract is in the
project log; reviews grade against it. The contract carries NAMED craft
decisions —
scripts/plan_gate.shverifies mechanically before assembly. Mechanical exemption is USER-GRANTED ONLY (exact operations: trim, convert, mux, technical fix). Never self-declare a vlog/montage/promo "mechanical" — those are designed deliverables by definition (engines/README.mdrouting). - no-silent-downgrade Never swap an approved direction/engine/service
for a weaker path without disclosure and approval. Includes dropping
audio: audio-preserved — a deliverable without its expected audio
fails review unless explicitly approved (
scripts/loudness_check.shenforces). Disclosure is not a licence: a tool/render failure needs a minimal-repro isolation pass BEFORE any downgrade is proposed (engines/hyperframes.md§ Render troubleshooting). Deliver the best PASSING output — shipping a degraded file while a better render exists is a delivery failure. - hyperframes-handoff Designed deliverables go through the full
hyperframes pipeline — no hand-rolled compositions, no concat+xfade
slideshows (
engines/hyperframes.md). The engine choice is a logged## Decision — Engineentry inproject.md(plan gate checks it); FFmpeg-only assembly of a watchable deliverable without a USER-GRANTED mechanical exemption fails final review regardless of output quality. Multi-scene pieces assemble by scene-loop, not one shot: every Scene Ledger row reaches LOCKED (scene rendered, six checks passed,review/scene-review.md) BEFORE the master renders —scripts/scene_gate.shenforces that seam mechanically (verbatim block in the project log; a master rendered without a passing SCENE GATE block is a skipped workflow), andreview_gate.shre-runs the same check at delivery so a ledger still sitting at DRAFT — or scene files with no real iteration entries — FAILs it either way. - sample-first Generation goes concept → ONE sample → batch; a batch
starts only after its sample passes (
mcps/README.md). - review-independence Final review = evidence with timestamps/frame
numbers; the technical gate is ONE command —
scripts/review_gate.sh— and its verbatim output block must appear in the verdict. Reporting a deliverable path before that verdict block exists inproject.mdis itself a violation. Production-level outputs get a clean-context reviewer; any mutation invalidates prior review (review/final-review.md). - rubric-verbatim The Appeal verdict uses the SEVEN named rows of
review/final-review.md§D exactly as written (concept · contract · rhythm · restraint · craft · sound · typography&motion). Invented dimensions void the review. Concept ≤5 caps the verdict at "revise". - log-or-it-didn't-happen Decisions, contracts, reviews live in
project.mdwith absolute paths; versioned re-renders, never silent overwrites (review/project-log.md). - no-occlusion Overlays/subtitles/PiP never cover key faces, UI or focal actions; subtitles burn last, on output-timeline timestamps.
Domain rules use the same bracket notation inside their owning doc:
[no-zoom-drift] (craft/transitions.md), [no-placeholder-assets]
(craft/motion-recipes.md), [motivated-transitions]
(workflows/style-replication.md).
Directory map
| Directory | Purpose |
|---|---|
craft/ | Editing craft: art direction (+ default kit), pacing, transitions menu, motion recipes (+ snippets/ copy-paste implementations), camera rig, audio-first cutting, beat sync, sound mix, grading, fonts |
assets/ | Bundled reusable files inside this skill: fonts/ sfx/ bgm/ images/ clips/ (browse: assets/local-media.html; recipe menus: assets/index.html) plus atom-packs/ — per-look reviewed atom packs (atoms.json = machine-readable recipe, preview.html = review page) — check here BEFORE downloading or generating |
looks/ | Named style recipes (paper-collage, freeze-punch, film-reel, brush, gallery-ripple, big-number, neon-ui-tech, prompt-to-product-ui) |
workflows/ | End-to-end flows: style-replication.md, vlog-multi-source.md |
engines/ | Ownership & routing: HyperFrames handoff, FFmpeg prep/finishing |
mcps/ | Perception tools, generation services, sample-first rule |
review/ | Source review, final review, project log |
scripts/ | Verified measurement tools (see scripts/README.md) |
Task routing
| Signal | Read |
|---|---|
| Any new deliverable | craft/art-direction.md first |
| Pile of raw clips, no reference | workflows/vlog-multi-source.md |
| "Make it like this reference video" | workflows/style-replication.md |
| BGM supplied / beat-cut requested | craft/music-beat-sync.md |
| Choosing seam treatments / "transitions look ugly" | craft/transitions.md menu + craft/pacing-rhythm.md §4 |
| Title/overlay/data animation design | craft/motion-recipes.md approved palette |
| Camera moves / push-in / 运镜感 on designed frames | craft/camera-rig.md |
| Talking head / interview / narration | craft/audio-first-cutting.md |
| User names a style ("拼贴/贴纸", "书法", ...) | matching card in looks/ |
| Verifying one scene inside the scene loop | review/scene-review.md six checks |
| Tech/product promo or prompt-to-UI demo (designed frames) | looks/neon-ui-tech.md / looks/prompt-to-product-ui.md + engines/hyperframes.md handoff |
| Subject cutout / sticker / text-behind | mcps/segmentation-matting.md + looks/paper-collage.md |
| Ready to render/deliver | review/final-review.md |
Docs are read on demand — a task only needs its matching signal rows plus the standard-flow steps it actually executes, never the whole tree.
Not this skill
Watching/analyzing only → core perception tools directly. Pure image or
audio generation with no video deliverable → the service tools directly.
Designed video with no real footage at all (pure promo/motion-graphic
from a brief) → the hyperframes entry skill directly.
More skills from the Qwen-MM-Plugins repository
View all 8 skillsqwen-mm-plugins-api
analyze media with Qwen vision models
Aug 15AIComputer VisionMCPMultimodalqwen-mm-plugins-blender
build and edit 3D scenes in Blender
Aug 53DAnimationCreativeDesignqwen-mm-plugins-core
analyze and visualize multimodal files
Aug 12Data AnalysisLLMMCPMultimodalqwen-mm-plugins-edu-agent
generate math problem solving tutorial videos
Aug 5Content CreationEducationMathematicsVideoqwen-mm-plugins-freecad
model parts and assemblies in FreeCAD
Aug 53DCADDesignEngineeringqwen-mm-plugins-search
search and extract web and image data
Aug 15Data ExtractionMCPSearchWeb Scraping
More from Qwen
View publisherbatch
execute parallel batch operations on files
qwen-code
Jul 20AutomationConcurrencycoordinate
coordinate Qwen Code teammates
qwen-code
Aug 15AgentsCodingMulti-AgentOrchestrationcua-driver
drive native GUI applications via MCP
qwen-code
Aug 9AccessibilityBrowser AutomationDesktopMCPdataviz
design and validate data visualizations
qwen-code
Jul 16ChartsDashboardsData Visualizationextension-creator
create Qwen Code extensions
qwen-code
Jul 16CodingDocumentationPlugin Developmentloop
run scheduled or self-paced prompt loops
qwen-code
Jul 16AutomationProductivityScheduling