Editing a take is maybe 80% mechanical: cutting dead air, pulling the "ums", re-timing captions to the shorter timeline. Rough-cut does that part on its own. It deliberately leaves colour, b-roll, transitions and sound design alone — that is the 20% that is actually yours. Take the build brief below and have Claude build you the basic version, or skip the build and get the finished plugin.
Under the hood
01
The useful line to draw is not "can AI edit?" but "which parts of editing have no taste in them?" Trimming silence has no taste. Choosing the b-roll does. Automate the first, never the second.
02
The Python is unremarkable — a few hundred lines of ffmpeg wrangling any competent developer would write the same way. A bug in a script gives you a slightly wrong cut; a bug in the file that decides which script runs when gives you a tool that cuts the wrong things confidently. Spend your review time there.
03
It stops before anything subjective — which lines to cut, which opening is the strongest hook — and asks. A pipeline that asks is worth more than one that guesses confidently.
The basic version
The full build brief for the basic version: 8 stages and 6 gated phases, in build order. It asks Claude to stop at every phase boundary and explain what that phase bought you, so you come out understanding the tool rather than owning one.
The pipeline · 8 stages
deterministicjudgment · stops and asksWord-level transcript with timings. Everything downstream is a query over this one artifact.
Find the dead air. Pure signal processing — no transcript needed, no taste involved.
Locate the ums, uhs and false starts by matching the word list against the transcript.
Claude reads the transcript and proposes which lines are worth losing — then stops and asks.
stop & ask
Merge every upstream signal into one ordered list of segments to keep. The only stage that resolves conflicts.
Claude ranks the candidate openings and says which earns the first three seconds — then stops and asks.
stop & ask
Cut the file to the approved keep list. One ffmpeg pass, no re-encode where it can be avoided.
Re-time the transcript onto the shortened timeline. Captions that drift are worse than none.
mkdir rough-cut && cd $_
claude
# paste the brief as your first messageIf you only remember one thing
The orchestrator is the product.
The Python is unremarkable — a few hundred lines of ffmpeg wrangling and interval arithmetic any competent developer would write the same way. What makes it a product is the file that decides which script runs when, what to do with the output, and when to stop and ask a human.
Spend your review time on SKILL.md and the reference docs — not the scripts.
Would rather not spend the tokens?
The finished plugin is the same pipeline with the two phases this brief leaves out — the buyer-facing layer and the adversarial gauntlet — already done.
Nothing here left your browser — the prompt is generated on this page, with no account and no upload.