Use Second Look safely

Prerequisites

• RoughCut Pro. Turning the toggle on without a subscription opens the Pro paywall and leaves it off.
• Bad-take removal switched on. The Second Look toggle is disabled while it is off, because Second Look re-reads the output of the first AI pass. With no first pass there is nothing to refine.
• An AI provider key, for the same reason.

What it does

Second Look sits under Bad-take removal in Cut controls, with a PRO badge, off by default. Its help text in the app describes it precisely:

Re-reads the kept footage once and removes only high-confidence leftovers (missed retakes, false starts). Borderline cuts become Review suggestions — never silent deletions. One extra AI call (~1–4 min).

The important word is kept. Second Look does not re-read your recording; it reads only the transcript that survived the first pass. Anything the first pass already removed is out of scope, so Second Look cannot restore a cut, and it cannot reconsider a decision that has already been made.

It is described in the app's own source as experimental. Treat it that way: useful, bounded, and worth turning off if it is not earning its place on your material.

Passes

The toggle on a job requests one pass. That is the only value the job screen sets.

The underlying configuration allows up to three passes, and the shipped default is zero — Second Look is off unless something turns it on. Turning the job toggle on writes one pass; turning it off writes an explicit zero that overrides any configuration default you may have set.

The loop also stops on its own. A pass that auto-applies zero spans ends the loop — the run is treated as converged and no further pass is attempted, regardless of how many were allowed. The model is never asked to judge whether it is finished; the stopping rule is arithmetic.

What it is allowed to do without asking

This is the part worth reading carefully, because the limits are what make an extra automated pass safe to run at all. Every proposal is checked against four gates, and failing any of them turns the proposal into a suggestion instead of a removal.

Tangents and filler are always suggestion-only. They are never auto-applied, at any confidence, in any configuration. Those are the two categories where the editor is judging intent rather than repetition, and where an automated second opinion is least trustworthy — over-cutting screen demonstrations as "tangents" is a known failure mode.

Read the gates together and the design becomes clear: on a ten-thousand-word transcript, Second Look may remove at most ten spans, totalling at most a thousand words, all of them repetitions or noise, all of them at very high confidence. Everything beyond that budget stops being an action and becomes a question.

Suggestions, not silent deletions

Anything that fails a gate appears in the review list as a suggestion: the span is still in your cut, marked with its proposed reason, with a one-click control to remove it if you agree.

Second Look suggests removing (<label>) — kept until you decide

Nothing is deleted quietly. A suggestion you ignore stays in the cut. A suggestion you accept is applied when you run Recompute Cuts, exactly like any other decision you make in review.

Spans Second Look did auto-apply are labelled too — their reason is prefixed to mark where they came from, so you can tell first-pass removals from refine-pass removals when you are reviewing. Group them with By category and you can audit the whole refine pass in a minute. See Review by category.

Cost and time

Second Look is one additional AI call, and the app estimates roughly 1–4 minutes for it. That is on top of everything the standard job already does, and it is billed to your own provider key like any other call. Longer transcripts cost more, on both counts.

If a job already takes as long as you are willing to wait, that is a reasonable argument against leaving Second Look on by default.

There is one limit worth knowing: a surviving transcript above the editor's single-call word limit is too large for the refine pass, and the pass does not run. The record says so; the cut is unaffected.

Failure is isolated by design

Second Look is built so that it cannot damage a job that already worked. Its service is written never to throw: any failure is captured, passes that already succeeded stand, and the pipeline keeps the intact first-pass result.

If it fails, the job completes anyway and the header tells you plainly:

Second Look could not run — this cut is the standard single-pass result.

The specific error is appended after that sentence. Your rough cut is finished, exportable, and identical to what you would have got with Second Look switched off. Nothing needs to be re-run to recover; if you want to try again, process the job again.

How to use it well

1. Process the job without Second Look first, so you know what the standard pass produces on your material.
2. Turn Second Look on and process the same recording again.
3. Compare. Look specifically at what it auto-applied and whether you agree.
4. Read every suggestion. They are the proposals the system itself was not confident enough to act on, which makes them the ones most worth your attention.
5. Keep it on only if step 3 was consistently favourable on your kind of recording. It is not a setting to leave on unexamined.

Common mistakes

• Turning it on for a first job. Learn what the standard pass does first. Two variables at once tell you nothing.
• Expecting it to catch tangents. It can only suggest them, never remove them. That is deliberate.
• Expecting it to undo a first-pass cut. It only sees what survived. It cannot restore anything.
• Treating suggestions as deletions. A suggestion left alone stays in your cut.
• Reading the failure banner as a broken job. The banner means the cut is the standard single-pass result — a complete, exportable cut.
• Assuming more passes are better. A pass that accepts nothing ends the loop, so extra passes often do nothing at all.

Troubleshooting