One prompt, full edit

Give the agent a single outcome prompt and watch it storyboard, cut, and review its own work — then learn to redirect, undo, or polish by hand.

The fastest way to edit in Pluged AI is to stop steering. This tutorial takes a raw 4–5 minute talking-head recording — a founder update, a lesson, a rant — and turns it into a finished vertical short with one prompt, in about 10 minutes. Along the way you'll learn to read what the agent shows you at each stage, so you know when to trust it and when to step in.

The scenario

You recorded yourself talking for five minutes. Somewhere in there is a great 40-second video. You don't want to scrub, cut, caption, and mix — you want to describe the outcome and check the result.

Step 1: The one prompt

Import your recording (drag it into the editor), open the Agent panel, and type:

"Edit this into an engaging vertical short — about 40 seconds. Open on my strongest line, captions on, music under it. Cut everything that doesn't earn its place."

Press Enter, then keep your hands off the timeline for a few minutes. Here's what happens.

Step 2: What you'll see — stage by stage

The route (and at most two questions)

The agent routes by what's already in your project, not by interviewing you. With footage present and a clear ask, it states its plan in one line and starts. It will only ask a question when the answer changes the cut — for example, if your project is empty ("wait for an import, or build from stock?") or the destination is genuinely ambiguous (a 3-minute explainer cuts very differently from a 30-second reel). Two questions is the ceiling; everything else it decides from sensible defaults and tells you.

The storyboard card

Before its first edit, the agent looks at your footage — visual analysis plus the transcript — and commits to an editorial plan: a one-sentence thesis, a target duration, and 3–6 beats (hook, context, payoff, CTA), each mapped to a specific range of your footage. This appears as a collapsible card above the chat and persists across the whole session.

The card is the contract. Every cut the agent makes serves a beat, and beats flip to done as it completes them. If the plan looks wrong — wrong thesis, wrong hook line — say so now (see "Redirecting" below). More in Storyboard.

Tool rows

Each edit streams into the chat as a tool row: analyzing the footage, keeping the segments that serve the thesis, reframing for the vertical canvas, captioning timed to speech, laying a music bed under the whole cut. Every row is a real, immediate, undoable change on your timeline — there is no separate "preview mode" being rendered somewhere else. The full catalog is in the tools reference.

The finish sequence

This is the part that makes delegation safe. Before the agent is allowed to say "done", it runs a mandatory self-check, in order:

Critique (critique_edit) — a structural review of the cut, including deterministic lint checks: overlapping text, text outside the frame, a music bed that dies partway through, machine-gun cuts.
Verify (verify) — confirms the concrete targets: duration, captions present, correct canvas.
Review (review_edit) — renders your actual timeline and has an AI watch the video like an editor, returning problems, suggested fixes, and a ship/no-ship verdict. (Timelines over 120 seconds skip this step.)

The agent fixes what each check finds and re-runs until they pass. And there's a backstop: if it tries to conclude while its latest check is still failing, a quality gate bounces it back to keep fixing — it cannot talk its way past a failed review. Full detail in How the agent checks its work.

When the agent finally reports done, you're looking at a cut that has been watched, not just assembled.

Step 3: Redirecting mid-run

You don't have to wait for the finish to disagree. Type into the chat at any point:

"Actually make it 30 seconds, and the hook should be the line about firing our biggest client."

The agent updates the storyboard first — you'll see the card change — then re-edits against the new plan. Direction changes go through the plan, so the card always reflects what it's actually building.

Step 4: Undo, or take over

Undo the run. Every agent action is a normal undo step. Cmd/Ctrl+Z steps back through its changes one by one — keep pressing to walk the entire run back to your untouched footage. Nothing the agent does is irreversible.
Polish manually. The agent's output is ordinary timeline elements. Trim a cut that lands a beat too early, retype one caption, drag the music down 2 dB in the audio controls. You can hand it back to the agent afterwards — it reads the current timeline state fresh every turn.
Review-before-applying. If you'd rather approve changes before they land, turn on the review toggle in the agent settings menu and the agent proposes instead of applying.

Make it yours

Same one-prompt shape, different outcomes:

"Turn this into a 60-second clip for LinkedIn — landscape, minimal-clean captions, no music."

"Make a 30-second teaser that ends on a cliffhanger. Don't reveal the answer."

"Edit this product walkthrough into a tight demo — cut my rambling, keep every step on screen long enough to read."

If a prompt produces a flow you'll reuse, you can save the run as a replayable recipe and apply it to next week's footage.

Troubleshooting

The agent asked a question and stopped. Answer it — it only asks when the answer changes the route. One reply and it's off.
The storyboard's thesis is wrong. Redirect before it edits far: "Wrong angle — this video is about X, not Y." It rewrites the card, then re-cuts.
It finished but the pacing drags. Say "tighten the middle — cut anything that repeats the point." Refinements are conversational; you never restart.
The review step was skipped. Cuts longer than 120 seconds skip the AI watch-through; critique and verify still run. Ask for a shorter target or review the long cut yourself.
It stopped mid-run. On the free plan, hitting 100% of your monthly AI budget blocks new requests (in-flight work finishes). Check the usage bar — see Plans and limits.