Responsibility and Location
Responsible for understanding the goal, selecting capabilities, arranging the order, bridging outputs, and summarizing the delivery. First, read video-shared: it maintains shared rules; deterministic operations are uniformly executed by the local video-cli; domain Skills maintain business judgment; this entry point handles only cross-domain orchestration.
The local entry point, shared layer, and domain capabilities are stored entirely under the current project's .agents/skills/ directory, referenced as adjacent directories. ChatCut continues to use its official companion Skill.
Selecting Capabilities
Based on the current goal, only read the required Skill, then read the reference materials it requires. The following paths are relative to this entry point directory:
| User Goal | Sibling Capability | Primary Deliverable |
|---|---|---|
| Transcribe video or audio to text, supplement speaker labels | video-to-text | Timestamped transcript, JSON, and subtitles |
| Sort topics, insights, quotes, controversies, and action items from existing text | insight-extractor | Insight summary with source localization |
| Clip short segments from long video by topic, speaker, or time range | video-clipper | Segment list and rough-cut clips |
| Remove re-recordings, repeated narrations, and unfinished sentences; retain complete expressions | speech-video-dedup | Dedup candidates, review preview, or cleaned full video |
| Resegment existing subtitles semantically into short sentences | subtitle-resegment | New short-sentence SRT |
| Evaluate opening appeal, content progression, and engagement triggers | clip-quality-monitor | Copy diagnosis with original text evidence |
| Explicitly use ChatCut or continue its engineering | ChatCut official talking-head-guide and its associated Skills | In-project editing, subtitles, or user-requested packaging |
The ChatCut native Desktop companion entry is typically at ~/.agents/skills/chatcut-talking-head-guide/SKILL.md; prioritize the official version already provided in the current session.
Clarifying Task Scope
- Reuse existing tools, projects, material paths, segments, or confirmed editing methods. Do not change the user's choice simply because a tool is available.
- "Just transcribe", "Just evaluate", "Find duplicates" should only perform the corresponding work, not expand into a full finished product.
- For requests like "Help me edit" with insufficient goals, first confirm whether to retain the full narrative and clean it, or extract a short video from a long video. If materials are missing, first check the user's path and project context, then ask for necessary information.
- If a local capability is missing, first look in the current working directory; do not switch to a global copy with the same name, do not automatically download or reinstall. For ChatCut official dependencies, look from its plugin or Desktop companion sources.
Orchestrating Combined Tasks
Clipping a Short Video from a Long Video
Execute video-to-text → insight-extractor → video-clipper as needed. Skip already satisfied steps if transcripts, summaries, or clear time ranges exist. Segment integrity checks and confirmation are the responsibility of video-clipper.
Rough cutting and silence removal do not replace semantic deduplication. Only continue with speech-video-dedup when the user requests "Also clean each clip"; do not add precision editing without request.
Deduplicating a Full Voiceover or Course
Use speech-video-dedup to check the entire film and cut points. Call video-to-text uniformly for transcription when needed; reuse and backend selection follow the shared transcription contract; deduplication does not maintain a second recognition pipeline. Do not compress a full course into a short video unless requested by the user.
Subtitles and Quality Diagnosis
When subtitles are needed, first determine the final edit, then obtain the corresponding SRT and necessary word-level timestamps from the final timeline, then call subtitle-resegment. For an original transcript, must map and convert according to the actual edit or re-transcribe the final video; do not apply directly.
clip-quality-monitor can assist with segment selection evaluation or provide post-edit copy diagnosis; it only gives suggestions based on text, does not verify actual playback, audio-visual, or platform effects.
Specifying ChatCut
When the user chooses ChatCut, use its official workflow to achieve the goal; do not cut externally first and then import arbitrarily. When text analysis is needed, use sibling insight extraction or quality diagnosis capabilities.
When Desktop or local project is explicitly specified, use Desktop tools; when cloud project is explicitly specified, use the corresponding connection. If both are available and project affiliation is unclear, first check project context; if still uncertain, ask. First determine the speech timeline, then produce the user-requested subtitles, music, and visual packaging.
Output Handoff
- Determine an output directory for this task for all stages to use. Directory rules, transcription reuse, and time baseline unification reference the shared layer; do not rebuild workspaces at each stage.
- Pass source file identity, transcription coverage range, and actual edit mapping across stages to avoid mistaking timestamps from one stage for another.
- Follow each capability's original clip protection, confirmation, and export rules; retain confirmed segments and cut points from this task.
Delivery
Provide the requested openable output and path, explain which capabilities were used and what scope was processed. Report on segment integrity, cut point and visual checks, technical checks, and copy diagnosis according to actual completion level; one passing does not mean the rest are verified.