Workbench Operations
Translate explicit editing requirements into existing CLI operations, always reading back results within the same project. This is not a separate editing engine, nor is it a mandatory entry point before using any other Skill.
First, Identify the Action
| What the User Wants | This Skill's Approach |
|---|---|
| View current project, find a phrase, see existing clips and overlays | Read-only positioning; see Operations |
| Place an animation or video over a phrase, keeping the original audio | Overlay layer; timeline does not extend; see Operations |
| Delete an overlay, a word, or a clip | Differentiate between object and audio impact; see Deletion |
| Split a clip into two, trim a segment, reorder existing clips | EDL modification; see Operations |
| Change canvas orientation (landscape/portrait), set image fit to fill or contain | Clarify canvas vs media fit; see Operations |
| Insert new media anywhere on the main timeline and shift subsequent clips | First check current capabilities; if no generic insert command exists, identify a specific gap—do not pass overlay or move off as insertion |
When the user says "insert," first determine whether it's placing an overlay while keeping the voiceover, or adding new duration and pushing the rest of the timeline forward. When the user says "delete," first determine whether it's removing an overlay or cutting out audio and time. Execute immediately if the intent is clear; do not ask repetitive clarification questions. For duplicate phrases, reused media, or ambiguous targets, list positions for confirmation—do not guess the instance based on text alone.
Tasks like automatically judging speech errors, repetitions, or which take to keep should be delegated to the "Crop Speech" Skill. Subtitle content and styling go to the "Subtitles" Skill. Lens/animation creation goes to the respective design Skill. Final video export goes to the "Export" Skill. Connection issues and upgrades go to "Maintenance and Troubleshooting." Do not automatically start additional Agents or install missing Skills.
Fixed Workflow for Each Operation
- Connect: On first use or when credentials change, fully read Connect and Submit. Retain confirmed Runtime parameter arrays, explicit origin, server projectId, and capabilities. Do not rediscover the environment on every action.
- Read: Read only relevant documents; save the document's own revision, target ID, original value, and expected result. For word positioning, paginate to read fully if necessary, confirming source identity and instance. Do not confuse subtitle cue IDs with word IDs.
- Plan: Briefly describe which object will be changed, whether the timeline length will be altered, and whether audio will be affected. Do not repeatedly confirm actions already explicitly authorized by the user; new impacts (e.g., removing additional related layers) must first obtain authorization.
- Submit: Call using the actual fields and mode of
workbench commands --json. The--confirmedflag on write operations indicates authorization has already been given; it is not a switch granting the Agent permission. Edit the official project only through the Runtime; do not modify the project JSON directly. - Verify: Read the affected resources, check IDs, diffs, and durations; if necessary, verify word anchors, subtitle state, and overlay timing. When chaining multiple patches, only construct the next step after the previous step's readback succeeds; this is not an atomic batch transaction.
- Deliver: Report separately: data has been written, whether the visuals/audio have actually been reviewed, and any unsupported or unverified items. A successful submission does not mean the preview/final product is valid; the page not refreshing does not mean the data was not written. Read back first; do not re-submit just to "refresh."
On Failure, Stop All Writes
- Conflict: Re-read current diffs; old approval cannot be automatically replayed by replacing the revision. Make a new plan only after confirming the new target and impact.
- Timeout, Disconnect, or Readback Failure: The result may have already been written. Save the original request and identity. Query precisely by the relevant resource, asset hash, or original operationId; not finding it does not prove it was not submitted. Do not blind-retry with a different ID.
- Insufficient resources, ambiguous targets, missing capability: Report the specific blocker and preserve the existing project. Do not automatically upgrade, restart, re-transcribe, modify registry, or switch to another tool.
- No fictional commands for generic undo/redo, volume mixing, playback control, or insertion of new media at arbitrary positions with timeline shift.
playbackis for reading the playback word stream;operation-recoveris for limited transaction recovery, not arbitrary undo. - Deletion only refers to the specified object within the project; it does not clean up the original video, audio, imported assets, module files, or cache.
Typical Requests
- "Cover the second sentence with this MP4, keep the audio." → Locate the real word ID, import the media, retain the old layer, add a new overlay, read back.
- "Remove the small black animation I just added." → Locate the unique layer, remove only that layer; do not cut the voiceover, do not delete the animation source file.
- "Delete the third occurrence of this sentence, and join what's after." → Verify the third instance, use the supported word/clip deletion workflow, and verify the new duration.
- "Split the first clip in half, then move the second half to the end." → Use clip partial cut followed by second-based split, read back the new IDs, then move according to magnetic timeline rules.