TaherZiad/montage-crew

Cut Mistakes & Repeats (Pipeline Agent 2)

Second step of the automated edit, run after cut-silences. It removes the things a human editor cuts on a second pass: stutters ("the the", "I- I-"), immediately repeated words, false starts (an abandoned phrase that restarts), and retakes (a line re-recorded — keep the clean take, drop the botched one).

Deciding what counts as a mistake is judgment, not a formula — emphatic repetition ("never, never") and rhetorical doubling ("the first piece is, is this…") look identical to a stutter mechanically. So this agent is review-gated by default: it surfaces candidates with full context; the agent (and the user) decide; only approved cuts are rendered.

When to use

• "cut the mistakes", "remove stutters / repeats", "clean up the flubs", "keep the best take", "cut false starts"
• As the second stage of the master edit workflow, on cut-silences' output.

The three-part flow

1. Find candidates (mechanical)

node .claude/skills/cut-mistakes/scripts/find-cut-candidates.mjs \
  <silence-transcript.json> --out-dir video-projects/<slug>/assets

Input is normally Agent 1's <stem>.silence-transcript.json, so cuts land on the already-silenced timeline. Writes:

• <stem>.cut-candidates.json — structured candidates (type, confidence, proposed cut range, removes/keeps text, context, recommendation)
• <stem>.cut-candidates.md — readable proposal

Candidate types: stutter (immediate word repeat), retake (duplicate/near-duplicate segment via Jaccard similarity), false_start (short abandoned phrase that restarts).

2. Review (the gate — REQUIRED)

Read each candidate in context and decide keep-or-cut. The mechanical detector cannot tell intentional emphasis/rhetoric from a real flub, so do not trust recommend blindly — read removes + context. Common false positives: emphatic repetition ("it's never, never hands-off"), copula-then-question ("the piece is, is this…"), listing.

Present the candidates to the user with a recommendation per item; collect approvals. Write the approved cuts to a file:

{ "cuts": [ { "start": 100.93, "end": 101.06, "reason": "I- false start" }, ... ] }

(start/end in seconds on the input transcript/video timeline. You can widen a candidate's range — e.g. for a retake, cut from the botched take's start to the clean restart.)

3. Apply approved cuts

node .claude/skills/cut-mistakes/scripts/apply-cuts.mjs \
  <silence-transcript.json> --cuts approved-cuts.json \
  --video video-projects/<slug>/assets/edited-silenced.mp4 \
  --output video-projects/<slug>/assets/edited-clean.mp4 --apply

Writes the EDL, a re-timed <stem>.mistakes-transcript.json (feeds the motion-graphics agent + validate-beat-sync.mjs), a decisions log, and — with --apply — the cut video (ffmpeg trim+concat, A/V in sync). Drop --apply for a dry run.

Reviewing the result

Use the shared review tool to eyeball the cuts and spot-check boundaries:

node scripts/build-edl-review.mjs <stem>.mistakes-edl.json \
  --original video-projects/<slug>/assets/edited-silenced.mp4 \
  --edited   video-projects/<slug>/assets/edited-clean.mp4 \
  --output   video-projects/<slug>/assets/mistakes-review.html
npx serve . -p 8080 -n

Notes

• A very clean delivery may yield few or zero real cuts — that's a valid outcome; don't cut natural speech to hit a quota.
• Cuts between two spoken words are hard joins. They're usually clean for stutters/false starts; for tighter audio a 20-30ms fade can be added later.
• Hand the mistakes-transcript.json + the clean video to Agent 3 (motion graphics / tiered cards).