How do you prompt an AI to give useful feedback on a rough cut?

Upload the rough cut to the invideo agent that already holds your treatment doc, then prompt in four slots: (1) what the film is and its stage, (2) the rules to judge it against, (3) the focus areas — pacing, sound, emotional register, continuity, (4) the output format — timestamped notes with severity. Then push back: ask it to challenge the weakest editorial decision.

Send the cut to the same invideo agent that has been on the project — the one holding your treatment doc, character sheets, and shot breakdown. Feedback quality collapses without that context, because the agent needs your rules to judge against. If you're new here: the invideo agent is an agentic video tool where you can spin up role-based sub-agents (a storyboard agent, a DOP agent, an editor agent) that carry your project context across every step.

Slot 1 — Define the film and the stage. Tell the agent what it's looking at: "This is the rough cut of a 70-second psychological short, assembled to picture, no sound design yet, no final grade. Treat this as a pre-lock editorial review, not a polish pass." Stage matters — notes for a rough assembly are different from notes for a fine cut. This mirrors the briefing model behind tools like Eddie AI (video type, stage, objective, focus).

Slot 2 — Anchor it to your rules. Point the agent at the treatment doc and any frameworks already locked in context: "Judge this against the treatment doc and the five emotional stages we defined. Flag any shot where the stage register is wrong." In one documented horror short, the invideo agent caught that the entity reveal was running at Stage D instead of Stage C — a structural beat the director had missed. Without an anchored rule set, you get generic notes; with one, you get diagnostic notes.

Slot 3 — Name the focus areas explicitly. List what you want examined and what you don't: pacing and shot length, editorial density (cuts per beat), emotional progression, continuity (props, lighting, eyelines), sound design implications, and the ending sequence. Be explicit about exclusions too — "ignore color grade, that's a later pass" — so the agent doesn't waste output. This is the role-criteria-scoring posture that structured review frameworks like TeamBench recommend for any AI content review.

Slot 4 — Force a structured output. Demand timestamped notes with severity, not paragraphs: "Return a table: timecode, issue, severity (blocking / should-fix / nit), and proposed fix. End with a 3-line verdict on whether the cut is ready to lock." Timestamps make the notes actionable inside an NLE; severity stops you from chasing nits before structural problems.

Then push the agent past surface notes. After the first response, send a second prompt: "Now challenge the weakest editorial decision in this cut — the one I'd defend but probably shouldn't." AI assistants default to agreeable; explicitly inviting disagreement is what produces the notes you actually need, and Reddit threads on AI feedback consistently flag this as the unlock. A useful third pass: "What did I cut that the story needed, and what's in the timeline that the story doesn't need?"

One example prompt you can paste in: "You have the treatment doc and shot breakdown in context. Attached is the rough cut of a 90-second short. Review it against the treatment's five-stage structure and the locked visual grammar. Focus on: pacing, cut density, emotional stage register, prop/eyeline continuity, ending sequence. Ignore color and final sound. Return timestamped notes as a table — timecode | issue | severity | proposed fix — and end with a 3-line verdict. Then in a second pass, challenge the weakest editorial decision in the cut."

This step is the one most filmmakers skip and the cheapest one in the whole pipeline — no generations burned, just a context-grounded review. Skipping it is the most common mistake in AI-directed work.

Watch some of these to see what works for you:

it got one thing that I would have never noticed, the entities reveal shot. The moment it first appears clearly was running at the wrong stage register.

— Hridaye, invideo's creative director