Turn explicit editing requests into existing CLI operations, always reading back the result in the same project. This is not another editing engine, nor is it a required entry point before using other Skills each time.
First identify the action
| What the user wants to do | This Skill's path |
|---|---|
| View the current project, find a sentence, view existing clips and visuals | Read-only locate, see Operation method |
| Place an animation or video over a sentence, keeping the original spoken audio | Visual overlay; the timeline does not get longer, see Operation method |
| Delete visuals, words/sentences, or clips | Distinguish the object and its effect on audio, see Deletion methods |
| Split into two segments, shorten one segment, swap the order of existing clips | EDL modification, see Operation method |
| Change landscape/portrait ratio, have the visual fill the frame or display completely | Distinguish between the project canvas and media fit, see Operation method |
| Insert new material at any position on the main timeline and shift subsequent content | Check current capabilities first; if there is no general insert command, state the gap clearly and do not pass off an overlay or a shift as an insert |
When the user says "insert", first determine whether it means covering the visual while the spoken audio continues, or adding new duration and shifting what follows; when the user says "delete", first determine whether it means removing a visual or cutting out both audio and time. If the intent is clear, execute it without repeatedly asking; if there are repeated sentences, reused material, or the target is not unique, list the positions for confirmation and do not guess an instance from the text.
When automatic judgment of slips of the tongue, repetition, and which version to keep is needed, hand it to the cut-spoken-audio Skill; subtitle content and layout go to the subtitle Skill; shot/animation creation goes to the corresponding design Skill; final output goes to the export Skill; connection failures and upgrades go to maintenance and troubleshooting. Do not automatically start additional Agents, and do not automatically install missing Skills.
Fixed process for a single operation
- Connect: On first use or when identity changes, read Connect and submitin full. Keep the confirmed Runtime parameter array, explicit origin, server projectId, and capabilities, and do not rediscover the environment for every action.
- Read: Read only the relevant documents; save the document's own revision, target ID, original value, and expected result. When locating words/sentences, read in full with pagination if necessary, and confirm the source identity and instance; do not use a subtitle cue ID as a word ID.
- Plan: Briefly explain which object will be changed and whether it shortens time or affects audio. Do not repeatedly confirm scope the user has already explicitly authorized; new effects (such as additionally removing associated layers) require authorization first.
- Submit: Call according to the actual
workbench commands --jsonfields and mode. For write operations,--confirmedindicates existing authorization, not a switch that grants the Agent permission. Edit the formal project only through Runtime; do not directly modify the project JSON. - Verify: Read the affected resources, verify the ID, differences, and duration, and check the word anchors, subtitle status, and visual timing when necessary. When making multiple patches in succession, only construct the next step after the previous step has been successfully read back; this is not an atomic batch transaction.
- Delivery: Separately explain that the data has been written, whether the visuals/audio were actually reviewed, and any unsupported or unverified items. A successful submission does not mean the preview/final cut has passed, and the page not refreshing does not mean nothing was written; read back first, and do not submit again to "refresh".
Stop writing on failure
- Conflict: reread the current differences; an old approval cannot be automatically replayed by replacing it with a new revision; only make another plan after confirming the new target and impact.
- Timeout, disconnection, or readback failure: the result may already have been written; save the original request and identity. Query precisely by the corresponding resource, asset hash, or original operationId; not finding it does not prove it was not submitted. Do not blindly retry with a different ID.
- Insufficient resources, ambiguous target, or missing capability: report the specific blocker and keep the existing project. Do not automatically upgrade, restart, retranscribe, modify the registry, or switch to other tools.
- Do not provide fictional commands for generic undo/redo, volume mixing, playback control, or inserting new media at any position with ripple.
playbackis reading the playback word stream,operation-recoveris constrained transaction recovery, not arbitrary undo. - Deletion refers only to the specified object in the project; it does not clean up the original video, recordings, imported assets, module files, or cache.
Typical requests
- "Overlay this MP4 on the second sentence, leave the audio unchanged." → Locate the real word ID, import the media, keep the old layer and add a new overlay, and read back.
- "Remove the little black animation that was just added." → Locate the unique layer and remove only that layer; do not cut the voiceover and do not delete the animation source file.
- "Delete the sentence said the third time and join the following part." → Verify the third instance, follow the supported word/segment deletion process, and verify the new duration.
- "Split the first segment into two halves, then put the latter half at the end." → Use segment-local trimming followed by second-based splitting, read back the new ID, then move it according to the magnetic timeline rules.