Communitygithub.com

video-production

制作或修改视频:确认项目与基准,创建项目接续文件,组织素材筛选、初精剪、声音、调色、包装、字幕、反馈和交付。按用户系统、软件与实际可用工具执行,支持新片、已有粗剪和局部任务。

video-production とは?

video-production is a Claude Code agent skill that 制作或修改视频:确认项目与基准,创建项目接续文件,组织素材筛选、初精剪、声音、调色、包装、字幕、反馈和交付。按用户系统、软件与实际可用工具执行,支持新片、已有粗剪和局部任务。.

対応✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/doublelikefish/video-editing-skills/tree/HEAD/skills/video-production

お気に入りのAIに質問する

このエージェントスキルを事前に読み込んだ状態で新しいチャットを開きます。

ドキュメント

Video Editing

Use when the user says "edit a video," "continue this project," "apply feedback" or similar production requests. After installation, the Agent reads this Skill, fills in required information, and executes; the user does not need to learn new commands, manually create specification files, or install a full pipeline. Do not activate production when only discussing methods or developing software.

This document is simplified from an existing One Master + Nine Stage Skill system: it retains each stage's inputs, decisions, tool calls, handoffs, and verification, removing dependencies on dedicated repositories, fixed drive names, services, and registration formats. It can guide a complete workflow on its own; when detailed Stage Skills are installed, read only the ones needed for the current task, without requiring all of them.

Overview of the Full Process: 10 Stages, Enter Based on Existing Results

StageOriginal Responsibility / Optional Detailed SkillWhat This Stage Passes to the Next
1. Project Configuration & ContinuationThis Skill video-production (inherits the original master control role)Persisted configuration, current baseline, valid requirements, and available capabilities
2. Creative, Script & Storyboardvideo-preproductionCore narrative, paragraph responsibilities, reference usage, and content to be shot/produced
3. Media Receipt & Screeningmedia-screeningSource files, candidate windows, observation basis, and unchecked scope
4. Rough & Fine Cutvideo-first-cutEdit with clear source ranges, continuous audio/video, and reviewable preview
5. Input Restoration & Color Gradingvideo-colorCorrect input interpretation, camera matching, preserved style and references
6. Motion Graphics & Compositingvideo-motionApplied templates/graphics/split screens and editable source
7. Sound Design & Mixingvideo-soundClear VO/Audio Sync/BGM/SFX relationships and actual listening results
8. Subtitlesvideo-subtitlesText, timecodes, and display corresponding to actual dialogue and current edit
9. Review, Feedback & Revisionsvideo-reviewModifications for each requirement, verification, and version awaiting recheck
10. Output, Delivery & Archivingvideo-deliveryCorrect delivery version, project/dependencies, inspection results, and continuation point

This is a responsibility map, not ten approval steps or ten sequential processes. New projects typically go: 1 → 2/3 → 4 (interleaving 5–8) → 9 → 10; existing rough cuts start from 1 → 4; feedback starts from 1 → 9 → relevant stage → 9; incremental media starts from 1 → 3 → affected stages. Missing shots go back to 3, narrative changes go back to 2. For single cuts, transcodes, or sound fixes, go directly to the corresponding stage, only executing input, output, temporary location, and verification, without creating a full project framework.

The rough cut should already have comprehensible sound, essential basic color, and the camera angles and graphics layouts requested by the user. Stages 5–8 are listed separately for on-demand refinement; this should not be used to push them all after the user's review. Dialogue cues are placed before picture editing; formal subtitle styles can be done later.

1. Project Configuration & Continuation: Persist First, Then Execute

Acquisition/Interaction: First, read the information the user has already provided and the current project record; only ask for missing items. If there is no production goal, first ask what needs to be edited this time and what inputs exist; if storage configuration is missing, ask "Do you have a preferred storage method? Where are the source files and where should the project be saved?" Simultaneously confirm the software to be used, delivery purpose/aspect ratio/duration, any existing rough cut, and content that must be preserved. Do not repeat known information; the Agent decides on ordinary tool choices.

Execution: Distinguish between the user's device and the Agent/tool execution endpoint, check accessible paths, editor versions, read/write interfaces, and viewing/listening capabilities. Installed software does not guarantee controllability; ability to read files does not guarantee ability to modify projects. For existing projects, first restore the actual current state; do not infer from the maximum version or most recent file name. After confirming the location and write permissions, immediately create/update the entry file in the project root using the template below, read it back to verify, and continue the current round; do not just suggest the user create it, and do not end the editing task after writing the file.

File Actions: For new projects, by default create AGENTS.md and a CLAUDE.md for import only; preserve existing project rules section by section, only adding missing items. If the current host already has a suitable entry point, use it and point to shared rules; for other hosts, check their loading method. Create PROJECT.md if no project record exists; if one exists, reference it, and do not maintain a separate "latest" copy. Do not overwrite existing user rules or modify global settings. If unable to write to the user's computer, explain the limitation and deliver the file content and location that can be saved; do not claim the files were created.

Saving and Naming: Follow the user's directory/naming habits; if none exist, suggest a project structure with work/ for projects, production scripts, and used assets, output/ for version previews/deliveries, and temp/ for recreatable temporary files; create directories only when needed. Keep original source files read-only; name new outputs using <ProjectName>_v001_<Purpose>.<extension>, do not overwrite old versions or change camera-original file names. Check the target disk and available space; explicitly specify output, download, frame extraction, and tool temporary directories. If the disk is offline or space is insufficient, do not silently write to the system drive, session, or source media drive. Record the actual database location separately; do not move active libraries.

Output / Next Steps: The project entry point is readable; source and working locations are separated; baseline, requirements, and capability gaps are clear. If scripts/rough cuts exist, skip redundant preparation; if only a specific capability is missing, continue with independently executable parts and record pending inputs.

2. Creative, Script & Storyboard

Acquisition/Interaction: Read the goal, audience, script/references, and existing content. Only ask the user if the core narrative, style, or scope of rework is unclear; do not ask for a re-selection of an adopted approach.

Execution/Capabilities: Clarify what each segment should make the audience know, see, and feel; map the script, references, planned shots, and actual camera candidates together. Shooting order is not necessarily narrative order; the user's existing edit is the baseline. The Agent handles text organization; only search for references if needed, only call generation APIs for storyboard images or generated assets if a connected capability exists; do not call generation APIs for normal structural discussions.

Deliverable / Next Step: Save a concise narrative/storyboard plan, including paragraphs, content responsibilities, reference usage, existing inputs, and gaps; estimated duration is not real VO timecode. Send existing media to stage 3; if actual shots are missing, clarify the shooting requirements; if structure exists, use it directly for stage 4; do not force the creation of a layout/storyboard product.

3. Media Receipt & Screening

Acquisition/Interaction: Read the media directory, existing indexes/thumbnails/transcripts, user-specified shots, and their intended use. If files are missing, first check provided paths and related candidates; do not ask for reshoots just because a single search didn't find them.

Execution/Capabilities: First read existing evidence, then use metadata tools to verify source identity, frame rate, audio tracks, color basis, and sync. Extract frames/create thumbnails/transcribe/analyze continuous segments as needed; observe subject, action phase, direction and speed of movement, focus, stability, usable head/tail, and source audio. Also compare relevant unused and unplanned shots; a single imperfection does not disqualify a whole group. A still frame cannot prove a complete action; ASR cannot prove tone and listening perception; dialogue-driven content forms cues aligned with the actual sound before picture editing.

Deliverable / Next Step: Candidates come with source identity + source time range + actual observation scope + intended role + issues/unchecked items. Candidates are not adoptions; pass to stage 4 for comparison selection. If narrative changes are needed, go back to stage 2. Reuse cache, only supplement missing/changed observations; leave media as files, review by issue batch, do not repeatedly stuff base64 or full thumbnails into the conversation.

4. Rough & Fine Cut

Acquisition/Interaction: Read the native rough cut/current edit, actual dialogue, candidates, manual modifications, and content that must be presented in this round. Only consult the user if the adopted structure or major creative decisions need to change; execute clear local modifications directly.

Execution/Capabilities: First save a recoverable copy and verify the active project, timeline, and complete audio/video baseline; execute via supported API/MCP/CLI. Allow only one writer per application instance; switching projects also counts as writing. Without a native interface, do not reconstruct an "equivalent" rough cut from incomplete endpoints; do not stuff a whole MP4 into a track and call it an editable project.

First select complete sentence groups/performances, then handle repeats, failed takes, breathing spaces, action starts/ends, and rhythm. For each B-roll, clarify its explanation/evidence/emotion/cut-hiding role, determine the source window, entry point, exit point, and the main shot after exit; continuously check surrounding sentence groups and all related layers to avoid dialogue flashbacks or covering the wrong sentence. Use straight cuts, action match, static match, crop/scale, split screen, picture-in-picture, J/L-cuts as needed for expression; do not just list technique names. When speeding up overall or by segment, maintain corresponding picture, audio, subtitles, graphics mapping, and user rhythm.

Deliverable / Next Step: A truly playable rough cut or local modification, retaining source/time mapping and key adoption reasons. Incorporate necessary elements from stages 5–8 into this version; for new complex effects, first verify with representative segments in context before expanding. If inputs are missing, use clear placeholders to indicate scope; do not claim it is a complete rough cut. After verification, go to stage 9.

5. Input Restoration & Color Grading

Acquisition/Interaction: Read real source information, project color management, existing manual grades/LUTs, and approved references; only ask the user if the style choice is missing or source information affects the conversion.

Execution/Capabilities: First confirm the conversion ownership to avoid missing or double conversions; if input is unknown, retain 'unknown'; do not guess Log based on camera model. First handle exposure, white balance, skin tones, and camera matching, then process the style. Apply through supported color interfaces; do not claim "parameters are set" without actual effect.

Deliverable / Next Step: Before/after comparison for the same source/time, applied objects, and unchecked items; check modified segments and adjacent camera angles. After other stages replace materials/re-export, verify that existing color grades are not lost; return to stage 4/9.

6. Motion Graphics & Compositing

Acquisition/Interaction: Locate the project's adopted templates, text/branding, references, source windows, and picture responsibilities. If no specific style is defined and it affects the viewer's impression, use representative frames/short segments for alignment; do not reinvent existing templates.

Execution/Capabilities: Determine the content hierarchy (single frame/split screen/picture-in-picture/labels), entry, readable hold, and exit. Reuse the correct template version, replace content for this project; do not directly flash an old template placeholder. Choose native compositing, existing template renderer, or generation tool based on capability; do not generate if existing assets suffice. Record the purpose, source, and placement of key effects in existing notes/timeline markers; do not force SFX for every effect.

Deliverable / Next Step: Editable source, actual used objects, and composite preview; check safe area, text, occlusion, transparency, dynamic start/end, and sound relationship. Successful rendering does not guarantee correct timeline composition; verify and return to stage 4/9.

7. Sound Design & Mixing

Acquisition/Interaction: Restore actual VO, sync sound, ambience, BGM, SFX, and existing processing. Use already selected music first; if a new emotional direction is needed, provide truly different audible candidates; if the user wants multiple versions, keep the comparison versions.

Execution/Capabilities: First determine the role of each layer and the sound-picture connection, then select sources, arrange, and mix levels. When changing music segments, compare phrases, instrument/drum layers, density, timbre, entry, and outro; choose J/L-cuts, rhythmic transitions, instrument/SFX bridges, or intentional silence based on intent. Fades and short whooshes cannot replace musical arrangement. Only search connected sound libraries/APIs for new sounds; compare real source windows, record source and adoption reason; do not re-search for ordinary level fixes.

Clean up unwanted camera noise/continuous noise; keep vocals natural; use content-based ducking and dynamics, not fixed values. Check the actual channel and routing of effective dialogue: a stereo container does not mean VO is present in both channels. Retain stereo music; do not unnecessarily sum the entire project to dual mono.

Deliverable / Next Step: Sound source/window, entry/tail points, and separated tracks/mix; first listen to problem layers in isolation, then with picture and surrounding context, checking VO clarity, BGM continuity, SFX relative loudness, transitions, and tails. If no audio inspection capability, write "not listened to" and do not select a "best" option; return to stage 4/9.

8. Subtitles

Acquisition/Interaction: Use actual adopted dialogue, proofread transcripts, subtitle templates, and output requirements. Only ask the user about keywords that are unclear or use technical terms; do not ask the user to re-submit already provided scripts.

Execution/Capabilities: Only call ASR if no transcript exists and there is dialogue, then calibrate against the sound; skip if there is no dialogue or subtitles are not needed. Share the mapping of source clock, cut points, speed, and timeline; preserve manual calibration. Use subtitle tools/editors to generate or modify tracks/SRT files, etc.; do not overwrite existing manually adjusted tracks.

Deliverable / Next Step: Text, timecode, and actual display; check start/end of text, across cuts, line breaks, reading time, font, safe area, and graphics occlusion. Verify subtitle tracks, burned-in settings, and exported visible subtitles separately; return to stage 4/9.

9. Review, Feedback & Revisions

Acquisition/Interaction: Read all valid requirements from chat, documents, or Review, excluding deleted/explicitly superseded items; retain the original ID/phrasing, version, and location. The user expresses naturally; the Agent organizes; only ask about items that cannot be located or are contradictory. If the user gives an example to point out a general problem, restore the full scope of application; do not fix only the example.

Execution/Capabilities: First break down each verification condition from the original requirement, then make modifications; do not infer passage from "what I did." Update each item as original version/location → requirement → this round's decision → actual result version/location → technical verification + actual viewing/listening → remaining, handling compound requirements separately. Execute in relevant stages, check the modified point, adjacent transitions, and affected sound/color/subtitles/graphics; reuse unchanged valid evidence.

Review products are optional: new projects can use chat/documents directly; for existing Review chains or when the user requests a Review, you must complete version registration/linking, page media and location verification, then provide a link and return to the original opinion. Registration is an independent delivery action; exporting is not registration. If the service is unavailable, explain the gap and continue feasible parts via an authorized alternative; do not silently drop the link.

Deliverable / Next Step: Return each item as Modified awaiting user recheck / Partially complete / Awaiting input / Awaiting location / No change needed (basis). "Modified awaiting user recheck" requires that this round's internal verification is complete; if required actual viewing/listening has not been done, write "Partially complete: technically verified, awaiting audio/visual check" and retain the Agent's inspection action in the next step; do not pass internal omissions to user validation. Do not mark a whole item as passed if anything is missing or unchecked; record user acceptance separately. If feedback is incomplete, continue related stages; if this round's delivery conditions are met, go to stage 10; subsequent feedback still returns to this stage.

10. Output, Delivery & Archiving

Acquisition/Interaction: Determine the specific output version, project/final video scope, platform specifications, subtitle method, and delivery location. Execute if the user has authorized export; for publishing, follow existing clear authorizations; do not turn every output into a confirmation gate.

Execution/Capabilities: Save a snapshot from the correct project/timeline, set and read back the scope, dimensions, frame rate, codec, color, audio tracks/subtitles, then render the current job. Check the actual file, not just the return code; for full video delivery, continuously review the entire piece; for partial delivery, check the changes and context, clearly state unchecked scope. If the full video is incomplete, provide a working preview; do not report delivery as passed.

Output / Completion: Deliver accessible final video, requested editable project, necessary assets/dependencies, and per-item results; update the current editing state/pending references. Project backup does not mean original source files are included. Only clean up recreatable temporary files created this round that are not in use and have no dependence on project results; retain adopted assets and receipts; historical cleanup is handled separately. Clearly state file completion, content verification, and user acceptance separately; do not auto-publish or fill in acceptance on the user's behalf.

How Tools Are Selected and Called

Each time a new capability is needed, follow the process below; reuse verified and unchanged configurations directly:

  1. Find capabilities by stage need: files/metadata, real media observation, transcription, timeline editing, sound, color, rendering, feedback. First check the current tool list, installed Skills, software public interfaces, and existing project integrations; avoid re-creating tools.
  2. Verify support scope: Confirm separately for the user's system and the execution endpoint, software version, OS support, input/output formats, authentication, and fees; systems and software are not one-to-one bound. Only probe specific capabilities in read-only mode; do not switch active projects for testing.
  3. Use existing execution entry points first: Prefer CLI/SDK/API, available MCP/plugins; first check the current help/schema, then pass clear inputs, source/time ranges, output and temporary locations. For projects with existing Workflow/Pipeline integrations, follow their route/brief and original stage adaptation notes; for other projects, do not require these commands.
  4. Add only what's missing: Check official documentation or verifiable tools; recommend based on the user's system/editor. Only install dependencies that are truly needed for this round and already authorized, using the tool's dependency list/installation method and verify; this Skill has no unified media requirements.txt. If existing interfaces are insufficient, write a minimal script for this project, then evaluate reusability; do not build another framework within the edit.
  5. Failure and alternatives: If an interface fails, first diagnose the cause; do not automatically take over the foreground, switch software, or reconstruct the timeline. If a reliable GUI path exists, explain that it requires screen interaction and operate along authorized paths; if the user's software is inaccessible, provide importable results/clear steps, marking which operations remain incomplete. When the user requests an automated final video, a guide does not count as delivery.
  6. Read back after call: Bind the actual output/project identity; verify parameters and content. For external generation, only return file references, necessary parameters, and a brief conclusion; keep the full payload locally; do not print base64. A path reference itself is not proof of viewing/listening.

Project File Templates: Filled and Written by the Agent

For a new complete project, use the two files below plus a Claude import file. Replace placeholders with the project's facts; for non-critical unknowns, write "to be confirmed/unverified"; if the project root and this write location are unknown, ask first; do not guess. For existing projects, merge missing items, retain original rules; if a status record exists, omit PROJECT.md and replace the name in the entry point with a reference to the real record.

AGENTS.md: Project Specification & Continuation Entry Point

# <Project Name> · Video Production Conventions

## Goals & Configuration
- Project goal/audience/delivery specs: <fact>
- User software & version: <fact>; Agent execution endpoint: <fact>
- Source files (read-only at original location): <location>
- Project/production assets: <location>; Output: <location>; Temp: <location>
- Naming: <follow habit or ProjectName_v001_Purpose>; do not rename originals, do not overwrite historical output.
- Native database/project location: <fact or unverified, do not move without permission>
- Verified tools & capabilities: <tool, operation, limits>; Capabilities to be added: <specific gaps>

## Every Session Start
- Read PROJECT.md (or the equivalent existing record) first; restore the baseline, current working version, valid feedback, and next action; verify against this round's user instructions.
- When producing/editing, read video-production from the host's installed Skills; read the available detailed Skill for the current stage; do not rely on a specific machine's Skill absolute path.
- If the entry point is not auto-loaded, read it explicitly; if a Skill is not discoverable, explain and locate the user-provided installation directory; do not claim it is loaded.

## Production & Verification
- Project configuration → creative/media → rough/fine cut (interleaving color, motion graphics, sound, subtitles) → review/feedback → output/delivery. Skip redundant stages for existing results; for local issues, only modify the affected area and its impact.
- Preserve original files, user rough cuts, and manual modifications; verify the project/timeline and baseline before writing, create a restore point; allow only one writer per application instance.
- Compare real candidates and source windows against the narrative; use dialogue cues to aid understanding; for B-roll, verify the cut placement, entry/exit, and the next shot's content.
- The user's requested sound, color, camera angles, and graphics layout cannot be absent due to "rough cut" status; after changing cut points/speed/materials, verify the associated sound, subtitles, color, and graphics.
- Before calling an existing tool, verify the interface and input/output; read back after the call; if an API fails, do not automatically take over the interface or reconstruct the project.
- Check the actual output and context against each original requirement; separate technical pass, actual viewing/listening, and user acceptance; clearly list unchecked/pending inputs.
- Continue feedback via the carrier specified in PROJECT.md; if an existing Review chain exists, register the correct version and write back to the original opinion; if there is no Review, the chat can close the loop.

## End of Each Session
- Update the project record's artifacts, per-item results, inspection scope, incomplete items, and next actions in the same file; do not create another "latest" copy.
- Handle temporary files based on result dependency and running state; do not arbitrarily delete original files, historical assets, or shared caches.

PROJECT.md: Use When No Existing Record Exists

# <Project Name> · Production Record

## Current Continuation
- User baseline/accepted version: <exact file, project/timeline; do not guess if unknown>
- Current working version/preview: <exact location; keep separate from user acceptance>
- Must preserve: <structure, cut points, manual edits, VO, color grade, subtitles/templates, etc.>
- This round's goal & applicable stages: <scope of this round>
- Feedback carrier: <actual entry point for chat/document/Review>
- Next action/pending input: <actionable step or specific missing items>

## Inputs & Production Decisions
- Inputs/references/existing observations: <path and purpose; reuse existing list>
- Paragraphs & candidates/adoptions: <paragraph role → source window → timeline position → adoption reason; or reference to existing plan>
- Key sound/graphics/color: <current decision, source and scope; or reference to existing record>

## This Round's Requirements & Verification
| ID/Phrasing | Original version/location & preserved items | Actual modification/new version location | Technical verification | Actual viewing/listening & scope | Status/Remaining |
| --- | --- | --- | --- | --- | --- |
| <original opinion ID; use F01 if none> | <fact> | <actual result> | <actual test> | <checked or not checked> | <awaiting recheck/partial/pending input, etc.> |

## Delivery
- Output/project/necessary dependencies: <path>
- Inspection scope and unchecked items: <fact>
- User acceptance: <clear basis or not yet confirmed>
- Temporary file handling: <cleanup scope or reason for retention>

CLAUDE.md: Share the Same Specification

For new files, write a single line @AGENTS.md (without code fences/backticks). For existing files, preserve all original text; only append this line if no equivalent reference exists. If the current host already loads a shared entry point, do not duplicate the rules. Other agents should use their supported entry point and point to the same specification.

How to Communicate with the User

At the start, use a few sentences to state the adopted goal/baseline, save location, this round's stages, and what will be delivered first; only ask about missing items; if things are clear, proceed. At completion, report using result and version → per-item results from original requirements → actual inspection scope → incomplete items/next steps. Information comes from tools and real media; do not use templates to replace execution, and do not make the user perform the Agent's internal verification.

関連スキル