Frame-aware edits
Scrub to a moment, select the subject, and make the change exactly where it belongs.
Built for exact edits.
Three parts of the workflow, shaped for fast review and repeatable agent-driven edits.
Agent controlled
Describe the outcome and let the backend run the generation, review, and timeline update.
Compare the result
Review original and edited clips before committing the best version to your reel.
CLI-01
fiebatt works in the terminal.
The fiebatt CLI wraps the backend for scripting, automation, and agent-driven editing. Any agent that can run shell commands can inspect projects, trigger localized changes, review variants, and export results without a custom integration.
The same workflow, scriptable: preview footage, generate changes, score variants, accept the best result, and export when the cut is ready. A portable skill file gives Codex and Claude the setup context they need.
found 3 active reelsanalyzing timeline... selected source clip man_walkinggenerating localized edit with qwen prompt planvariant ready: compare original vs edited, then acceptexport complete ./exports/demo-reel.mp4How it works
A tight loop from prompt to export.
01
Open a reel
Import footage or reopen a project. The timeline, source media, and accepted edits stay synced with the backend.
02
Describe the change
Scrub to the moment, select the region if needed, then ask for the edit in plain language through chat or CLI.
03
Review and commit
Compare variants against the original, accept the best result, propagate continuity, and export the final cut.
Why different
Not another prompt-to-video wrapper.
fiebatt treats generation as one step inside an editable reel system, not the whole product.
Current video tools
Timeline-first, manual, and hard to automate.
Prompt tools usually regenerate broad shots instead of making localized timeline edits.
Traditional editors rely on manual masking, keyframing, exporting, and re-importing.
Automation is brittle because the UI, project state, and render pipeline are not agent-native.
Variant review is often disconnected from the original cut and final export path.
Object identity and continuity usually require separate tracking, roto, or compositing workflows.
AI results are difficult to reuse because prompts, variants, scores, and timeline decisions are not stored together.
fiebatt
Prompt-first editing with a real timeline underneath.
Every prompt is tied to project context, playhead time, selected region, and conversation history.
CLIP, SAM2, qwen, scoring, and generation workers cooperate around the same edit window.
Claude, Codex, and shell scripts can drive the same backend loop through the CLI.
Accepting a variant updates the EDL, keeps continuity available, and exports from one surface.
Entity search can find matching appearances so localized edits can propagate beyond a single moment.
Generation briefs, tool calls, scores, variants, and accepted edits stay attached to the project history.
Technical flow chart
What happens after a prompt.
The same pipeline powers the editor UI and terminal workflow: collect intent, bind it to timeline context, plan tool calls, generate variants, then commit the accepted result back into the reel.
Loading the technical diagram…